GSD-Lite Work Log

1. Current Understanding (Read First)

<current_mode>

execution

</current_mode>

<active_task>

Task: READER-001 - GSD-Lite Worklog Reader (Mobile-First)

Status: POC COMPLETE (READER-001a + READER-001b) - See LOG-048

Key insight: Purpose-built worklog browser with positional parsing (ID, TYPE only), collapsible outline, resizable sidebar overlay, desktop/mobile toggle.

Completed: Parser + HTML renderer with full UX

Remaining: READER-001c (Mermaid), READER-001d (scroll sync), READER-001e (CSS polish), READER-001f (E2E test)

Parked (resume after READER-001):

Task: TASK-EVAL-002 - Constitutional Evaluation Pipeline

Status: SCHEMA REFACTOR COMPLETE - See LOG-046

Next: TASK-EVAL-002d — Implement Layer 2 Vertex AI integration

</active_task>

<parked_tasks>

- TASK-EVAL-001: Build OpenCode session parser — SUPERSEDED by SQLite parser (LOG-045)

- TASK-CONST-002b: Write remaining rubrics (Pillar 1, 3, 4) — needed for full evaluation

- TASK-CI-L1-001: Implement Layer 1 structural checks — integrated into TASK-EVAL-002c

- RQ-3: Evaluate SKILLS.md pattern — defer until current architecture matures

</parked_tasks>

<vision>

GSD-Lite is a pair programming protocol where:

- Engineer (Driver): Owns decisions, curates what gets logged, stakes reputation on outcome

- Agent (Navigator): Proposes solutions, executes tasks, narrates reasoning

- Artifacts (Memory): Persist beyond sessions, become institutional knowledge

Core insight: Agents are brilliant but ephemeral. Engineers are permanent and must own/comprehend every decision.

Fork & resume workflow: Pair program → log rich findings → kill session when tokens rise → resume with fresh agent pointing at curated logs.

Distribution Goal: Unified "Hotel Model" — gsd-lite install (global) or install --local creates identical structure at different roots.

Token Budget: First-turn headroom is 10k tokens. Agent instructions must stay under this (currently 4,913) to leave capacity for actual work. See LOG-020.

</vision>

<decisions>

DECISION-004: Universal Onboarding in PROTOCOL.md

- Rationale: Fresh agents using direct workflow entry (e.g., discuss.md) lacked high-level context (PROJECT.md). Onboarding must be a universal boot sequence, not workflow-dependent.

DECISION-005: Decomission Sticky Note Protocol

- Rationale: Artifacts are now mature enough to be the source of truth. Repeating status in every turn adds noise/tokens without value.

DECISION-013: Unified Installation Model (LOG-013)

- Rationale: Single command structure for both global and local. Same output (opencode config + gsd-lite artifacts), different root. Smart check: skip user artifacts if exist, always overwrite templates. Simplifies mental model vs fragmented install/update/local/force flags.

DECISION-020a: Consolidate artifact documentation to agent instruction (LOG-020)

- Rationale: HTML comments in templates are invisible to grep-first onboarding. Agent instruction is ALWAYS loaded. Token budget allows: 4,913 current + 1,250 docs = ~6,163, under 10k headroom.

DECISION-020b: 10k token limit as CI gate (LOG-020)

- Rationale: Token count is deterministic and measurable. Use as fast gate before expensive LLM behavioral tests. Complements LOOP-001 (Intern Test).

DECISION-028a: Constitution-first approach for CI (LOG-028)

- Rationale: Distill immutable pillars into machine-auditable document. All other artifacts can evolve as long as they don't violate the Constitution. Decouples philosophy (stable) from implementation (evolving).

DECISION-028b: Three-layer CI architecture (LOG-028)

- Rationale: Order checks by cost — L1 structural (free), L2 constitutional (~50k tokens), L3 behavioral (~500k-1M tokens). Catch obvious breaks fast, reserve expensive LLM calls for philosophy/behavior testing.

DECISION-032a: Use OpenCode native JSON for eval data (LOG-032)

- Rationale: OpenCode already persists all session data to disk as JSON. Parse existing files instead of building custom instrumentation.

DECISION-032b: Vertex AI Gen AI Eval as primary judge (LOG-032) — SUPERSEDED BY DECISION-042b

- Rationale: Original decision based on "Bring Your Own Data" model. Superseded because Vertex AI trajectory metrics require reference_trajectory (golden path), but GSD-Lite Constitution defines behavioral patterns, not expected sequences.

DECISION-042a: Session as evaluation unit (LOG-042)

- Rationale: GSD-Lite is stateless by design. Each session is independent. Handoff (S1-H1) and onboarding (C3-H2) are session-scoped behaviors.

DECISION-042b: Promptfoo with llm-rubric as primary evaluation platform (LOG-042) — SUPERSEDED BY DECISION-043a

- Rationale: Original decision favored Promptfoo for YAML-native rubrics. Superseded because Vertex AI rubric-based metrics offer adaptive rubrics + custom guidelines parameter, and user is Google Cloud partner seeking hands-on experience.

DECISION-043a: Vertex AI rubric-based metrics for Layer 2 evaluation (LOG-043)

- Rationale: Vertex's GENERAL_QUALITY with guidelines parameter can evaluate Constitution behaviors. Adaptive rubrics dynamically generate pass/fail tests per prompt. Agent-specific metrics (TOOL_USE_QUALITY) align with GSD-Lite's tool-heavy workflow.

DECISION-043b: Hybrid architecture — Programmatic L1 + Vertex L2 (LOG-043)

- Rationale: Deterministic checks (handoff presence, grep-before-read sequence) are free and fast in Python. Qualitative checks (reasoning quality, challenge tone) benefit from Vertex's adaptive rubric intelligence.

DECISION-043c: Constitution as guidelines parameter (LOG-043)

- Rationale: Inject distilled Constitution (P2-H, S1-H, J4-H*) directly into Vertex's guidelines parameter. No rubric format translation needed.

DECISION-042c: Hybrid orchestration — Option C (LOG-042)

- Rationale: Batch extract to individual files, evaluate each session independently, aggregate into summary. Enables surgical debugging + CI gates + re-runnable evals.

DECISION-042d: Turn-structured output schema (LOG-042)

- Rationale: Current flat schema (concatenated prompts/responses) can't support turn-level evaluation. New schema adds turns[] array with per-turn tools for LLM-as-judge correlation.

DECISION-033a: Fingerprint sessions via fs-mcp tool call paths (LOG-033)

- Rationale: When user spawns OpenCode from home dir but connects to different fs-mcp servers per session, project identity is invisible to OpenCode metadata. Extract absolute paths from tool outputs to fingerprint which project a session touched.

DECISION-045a: Migrate eval_ingest.py to sqlmodel (SQLite) (LOG-045)

- Rationale: OpenCode migrated from JSON files to SQLite. pathlib.glob logic is dead. sqlmodel provides typesafe ORM for the new opencode.db schema.

DECISION-045b: Update eval_transform.py to preserve full context (LOG-045)

- Rationale: Truncating responses to the last paragraph destroyed context for Vertex rubric evaluation. Rubric metrics like GENERAL_QUALITY need the full reasoning chain to be effective.

DECISION-046a: Implement Vertex-native turn-structured schema in eval_ingest.py (LOG-046)

- Rationale: Layer 1 checks failed on flat schema (losing turn context). Vertex MULTI_TURN_GENERAL_QUALITY requires structured turns. Refactored ingest to output request.contents[] directly.

DECISION-046b: Decommission eval_transform.py (LOG-046)

- Rationale: Transform layer existed only to bridge flat schema to Vertex format. Since ingest now outputs Vertex-native, the transform step is redundant complexity.

</decisions>

<blockers>

None - POC complete and working.

</blockers>

<next_action>

Fork paths (choose one):

1. READER-001c → Integrate Mermaid CLI (pre-render diagrams to base64 PNG)

2. READER-001d → Add scroll sync breadcrumb (IntersectionObserver)

3. Pivot to evaluation → Resume TASK-EVAL-002d (Vertex AI integration)

4. Test on mobile → AirDrop worklog.html, review mobile UX thoroughly

</next_action>

2. Key Events Index (Project Foundation)

3. Atomic Session Log (Chronological)

[LOG-001] - [DISCOVERY] - Vision Archaeology: Unearthing GSD-Lite's True Purpose from Artifacts and PRs - Task: BOOTSTRAP-001

Timestamp: 2026-02-03 14:30

Details:

This session performed "vision archaeology" — reconstructing the true purpose of GSD-Lite by examining its evolution through planning artifacts, merged PRs, and a production deployment. What emerged was not a "session handoff framework" but something deeper: a pair programming protocol that ensures engineers own and comprehend every decision, even when agents do the heavy lifting.

The Investigation: Three Sources of Truth

We examined three artifact layers to understand what GSD-Lite had become:

1. Original PROJECT.md (.planning/PROJECT.md)

- Started as "Data Engineering Copilot Patterns" — a documentation project

- Core value: "Maintain ownership of the reasoning process"

- This was the seed, but the plant had grown beyond it

2. Merged PRs (7 PRs from GitHub)

- PR #10: Echo-back onboarding — agents must prove understanding before executing

- PR #8: Summary in headers — logs are grep-scannable

- PR #4: Workflow decomposition — 929-line monolith → 5 focused files

- PR #2: INIT_PROMPT — 120 lines vs 300+ docs for activation

- Pattern: Every PR optimized for agent comprehension and token efficiency

3. Production Deployment (Meltano pipeline, /workspaces/.../estrid-meltano-el)

- 26 rich log entries demonstrating the pattern in action

- LOG-017 "Time Traveler Bug" — journalism narrative with analogy

- LOG-025 "Blank String Philosophy" — decision with "Silent Nod" metaphor

- PR_DATA_339.md — PR description generated from WORK.md logs

The Revelation: Fork & Resume as Core Mechanic

The production deployment revealed a workflow not explicitly documented:

SESSION 1 (tokens: 0 → 60k)
├── Pair program with agent
├── Hit finding → "Log this with journalism narrative"
├── Agent writes rich log entry
├── Tokens rising toward 80k
└── FORK: Kill session, keep artifacts

SESSION 2 (tokens: 0 → fresh)
├── "Read LOG-017, continue from there"
├── Agent onboards from curated artifact
├── Continue pair programming
└── FORK again when needed

The Insight User Emphasized:

> "Agents are ephemeral however brilliant. It is the engineers that has the final stake and should hone their skill learning from the best."

This reframed everything. GSD-Lite isn't about making agents more effective — it's about making engineers more effective by:

1. Capturing agent brilliance in permanent artifacts

2. Forcing engineers to curate what's worth preserving

3. Creating onboarding docs that teach reasoning, not just outcomes

The "Silent Nod" Analogy (from LOG-025):

To explain why this matters, consider this analogy from the production logs:

Imagine you ask a colleague, "Do you want coffee?"

- NULL Response: They stare blankly. You don't know if they heard you. (Unknown)

- Empty String Response: They give a Silent Nod. No words, but the message is clear.

In GSD-Lite terms:

- Chat history is like the colleague's spoken words — ephemeral, expensive to replay

- WORK.md logs are like written meeting notes — permanent, searchable, teachable

An engineer reading LOG-025 doesn't just learn that empty strings mean "local currency" — they learn the reasoning pattern for how to think about data semantics. They can apply that pattern to the next ambiguous field they encounter.

The Logging Standard (Captured in PROJECT.md):

User provided the explicit prompt pattern for rich logging:

> "Please include specific code snippet, reasoning, extended context in a journalism narrative style so that whoever with 0 context can onboard and pick up this key record of decision without ambiguity or friction. Add a cherry on top to include exact/synthesized example to explain the concepts/findings mentioned so that other weaker reasoning agents can grasp topic with little friction and ambiguity."

This is now documented in PROJECT.md under "The Logging Standard" with a table of required elements:

- Narrative framing (hook the reader)

- The symptom (what went wrong)

- The evidence (concrete proof)

- The root cause (why it happened)

- The analogy (ELI5 for onboarding)

- The decision (what we chose)

- Code snippet (executable proof)

Decisions Made:

Files Modified:

- gsd-lite/PROJECT.md — Complete rewrite with new sections: Philosophy, Pair Programming Model, Logging Standard

- gsd-lite/WORK.md — This log entry + Current Understanding updated for checkpoint

Cherry on Top: The "Driver vs Passenger" Mental Model

To help weaker agents (or future engineers) understand GSD-Lite's philosophy:

Imagine learning to drive:

- Passenger mode: You sit in the back, agent drives. You arrive at the destination but couldn't drive yourself.

- Driver mode (GSD-Lite): Agent is in the passenger seat with the map. They suggest "turn left here because of traffic." You make the turn AND understand why.

Six months later:

- Passenger: "Why did we go this route?" "I don't know, the agent chose it."

- Driver: "We avoided Main Street because it floods during rain. I learned that from the agent and now I always check weather before routing."

GSD-Lite ensures you're always in the driver seat — learning the "why" behind every decision, able to defend it, able to apply the pattern next time.

[LOG-002] - [DISCOVERY] - Protocol: Mining User Patterns from Storage (Discarding Sub-Agents) - Task: PROTOCOL-001

Timestamp: 2026-02-03 15:15

Details:

This session established a rigorous protocol for mining user behavior from the storage/ directory, specifically distinguishing between "pure" GSD-Lite sessions (manual fork) and multi-agent GSD sessions (programmatic fork).

The Discovery:

- Initial scan conflated gsd-lite sessions with gsd multi-agent sessions.

- gsd sessions use Task() calls to spawn sub-agents (gsd-planner, gsd-executor).

- gsd-lite sessions use manual session handoff ("Fork & Resume").

- Mixing these data points corrupts the insight into the user's manual workflow.

The Protocol for Future Agents (How to Mine Storage):

1. Discovery:

- List storage/project to find project contexts.

- Read project JSONs to identify relevant worktrees.

2. Filtering (CRITICAL):

- Read storage/session/{proj_id}/*.json.

- MUST filter for agent: "gsd-lite".

- MUST discard agent: "gsd-planner", gsd-executor, etc.

- Sort filtered sessions by date to find recent relevant work.

3. Reconstruction:

- Read storage/message/{session_id}.

- Read storage/part/{message_id} to get text content.

- Note: First message often contains the SYSTEM PROMPT injection in its text part. Look for the second message or distinct user parts for actual user input.

4. Insight Extraction:

- Look for "Fork" moments: Session Start → Read Context → Continue.

- Look for "Thinking Partner" moments: Agent challenging user assumptions vs just executing.

- Ignore "Spawn" moments: Task() tool calls are programmatic implementation details, not user behavior.

Why this matters:

Analyzing the wrong agent type leads to "optimizing the machine" instead of "optimizing the human". GSD-Lite is about the human workflow.

[LOG-003] - [DISCOVERY] - The "Driver" Persona: Pattern Recognition from Storage Mining - Task: PROTOCOL-001

Timestamp: 2026-02-04 10:30

Details:

Mining the storage logs for gsd-lite sessions revealed a distinct "Driver/Navigator" usage pattern that defines this user's productivity flow. The user doesn't just "chat" — they execute a precise rigorous loop of context setting, validation, and Socratic inquiry.

The "Fork" Pattern (Token Optimization Strategy):

The most productive sessions start with a specific signature:

1. The Trigger: "Let's load protocol and continue..."

2. The Context: Explicit pointer to WORK.md or a specific file path.

3. The Intent: "I'm at this part... my question is..."

This is the manual fork. The user curated the artifacts in the previous session (the "Pit Stop") so the new agent (the "Fresh Driver") can start immediately at high speed without carrying the baggage (token cost) of the previous 50 turns.

The "Driver" Persona (User Behavior):

The user acts as the Lead Engineer (Driver) using the Agent as a Senior Technical Investigator (Navigator).

* Context-Heavy Prompts: Instead of "fix this," the user says "Look at tmp/reference/.../streams.py, I see X, but ticket says Y."

* Validation Queries: "Is it valueTextType or itemNumber?" — specific A/B testing of hypotheses.

* Mental Model Checks: "Does this mean we need to create a new plugin child?" — verifying understanding before coding.

The "Thinking Partner" Requirement:

The user explicitly prompted an agent to be a "Patient Coding Teacher" with "Socratic Methodology." They don't want silent fixes; they want to derive the answer.

Cherry on Top: The "Pit Stop" Analogy

To explain why this "Fork & Resume" pattern is so effective:

Imagine Formula 1 racing:

* The Car: The Project State (Code + Artifacts).

* The Driver: The Agent (Ephemeral).

* The Race Engineer: The User (Permanent).

In a long race (complex feature), tires wear out (context window fills up, reasoning degrades).

* Bad Strategy: Keep driving on bald tires. The driver gets confused, hallucinates, crashes.

* GSD-Lite Strategy: Pit Stop.

1. Box Box Box: User calls "checkpoint" or manual fork.

2. Tire Change: Old context window is dumped.

3. Fresh Tires: New agent starts, reads WORK.md (Telemetry), and exits the pit lane at full speed.

The car (project) is the same, but the driver (agent) is fresh and focused. The "Fork" is just a Pit Stop.

Code Snippet: The "Fork" Signature

Future agents can identify these high-value start points by looking for this pattern in the first user message:

{
  "role": "user",
  "text": "let's load protocol and continue our work... I'm at [specific file path]..."
}

Status: Discovery complete. Validated that questioning enhancement must support this "Driver" workflow (Technical Investigation & Mental Model Validation), not just high-level "Vision Extraction".

[LOG-004] - [DECISION] - Storage Protocol v2: Direct Agent Filtering - Task: PROTOCOL-001

Timestamp: 2026-02-04 10:15

Context:

Refining the storage mining protocol from LOG-002 based on successful navigation.

The Finding:

- storage/session/{id}.json does NOT contain the agent field.

- storage/message/{msg_id}.json DOES contain agent: "gsd-lite".

- finish: "tool-calls" in message JSON indicates no text output (silent execution).

- finish: "stop" indicates text output or completion.

Protocol v2 (The Efficient Path):

1. List Sessions: ls storage/session/{proj_id}/ (sort by time if possible, or ID).

2. Sample First Message: Read storage/message/{first_msg_id}.json.

3. Check Agent: Filter for agent: "gsd-lite".

4. Scan for Interaction:

- Look for messages with role: "assistant" AND finish: "stop".

- Messages with finish: "tool-calls" are often silent execution loops.

- User patterns live in role: "user" message parts.

Why:

This saves token budget by avoiding reading full session contents for "ghost" sessions (sub-agents) or "silent" sessions (pure tool loops).

[LOG-005] - [DISCOVERY] - User Pattern: The Socratic Engineer - Task: PROTOCOL-001

Context:

Analyzed deep session ses_4011d8550... where user pasted a "Patient Coding Teacher" persona.

The "Fingerprint" (How this user works):

1. Persona: "Patient Coding Teacher" / "Collaborative Mentor".

2. Core Mechanic: Socratic Inquiry. Don't just answer; ask questions to help derive the answer.

3. Decision Making: "The Menu Technique" — present numbered lists of options (Concept vs Code, Strategy A vs B).

4. Pacing: "Mandatory Confirmation Loops" — pause and verify before moving on.

5. Philosophy: "Concept Over Code" — establish mental models first.

Implication for GSD-Lite:

The "Questioning" enhancement isn't just about requirements gathering (OG GSD style). It must be about Coaching.

- Old Model: "What feature do you want?" -> "Here are the specs."

- New Model: "Why this feature?" -> "Have you considered X?" -> "Let's map the mental model."

Action: : New questioning.md reference will focus on "Thinking Partner" techniques, not just project scoping.

[LOG-006] - [VISION] - The Questioning Enhancement: Learning Accelerator Through Socratic Pair Programming - Task: QUESTION-001

Timestamp: 2026-02-04

Context:

User initiated enhancement request: "The OG GSD framework has something great about its prompt engineering — the agent really asks the right questions." After analyzing OG GSD's .claude/get-shit-done/references/questioning.md and .claude/commands/gsd/discuss-phase.md, we interviewed the user to extract their specific vision for GSD-Lite.

The Core Discovery: GSD-Lite Is Not a Task Manager

Through Socratic dialogue, we derived that GSD-Lite's true purpose isn't task management or project planning — it's a learning accelerator. The artifacts (WORK.md, INBOX.md) aren't just logs; they're crystallized understanding that:

1. The user derived through dialogue (not received passively)

2. Are written so the user can explain them to anyone

3. Enable weaker agents (or future humans) to onboard without friction

The Golden Rule: "Always Ask WHY Before HOW"

This single behavioral instruction captures the essence of effective pair programming for this persona:

The Pair Programming Model

graph TD
    subgraph "The Triad"
        User["DRIVER (User)<br/>• Bring context<br/>• Make decisions<br/>• Own the reasoning<br/>• Curate what's logged"]
        Agent["NAVIGATOR (Agent)<br/>• Challenge assumptions<br/>• Teach concepts<br/>• Propose options with tradeoffs<br/>• Present plans before executing"]
        Artifacts["MEMORY (Artifacts)<br/>• Crystallized understanding<br/>• Zero-context onboarding<br/>• Weaker-agent-friendly"]
    end
    User <-->|Socratic Dialogue| Agent
    User -->|Curate| Artifacts
    Agent -->|Draft| Artifacts
    Artifacts -.->|Onboard| Agent

The 10-Star Experience (User's Own Words)

> "The 10-star experience was when I probed it to detour and go down a rabbit hole asking about a new concept that the codebase used. Asking it to explain, put on its teaching hat, explored the codebase and made the connections and distilled it down to layman level with concrete codebase snippets to solidify the concept for me."

This is the signature of a great session:

1. User notices something unfamiliar → pauses execution

2. Agent puts on teaching hat → explores, connects, distills

3. User gains understanding → now OWNS the knowledge

4. User prompts logging → journalism-style entry preserved

5. User forks session → fresh agent resumes with crystallized context

The Challenge Tone Protocol (Context-Dependent)

The user wants an agent that challenges assumptions — but the tone must match the situation:

Decision Tree for Agent:

flowchart TD
    Start([User states decision]) --> Why{Explained WHY?}
    Why -- No --> A[A: Gentle Probe]
    Why -- Yes --> Blind{Blind spot?}
    Blind -- Yes --> D[D: Socratic Counter-Question]
    Blind -- No --> Tradeoff{Genuine Tradeoff?}
    Tradeoff -- Yes --> C[C: Menu w/ Devil's Advocate]
    Tradeoff -- No --> Stakes{High Stakes?}
    Stakes -- Yes --> B[B: Direct Challenge]
    Stakes -- No --> Accept([Accept & Continue])

Key User Insight: "Writing Is Thinking"

> "The fastest way to learn anything for me is to get my hands dirty, to write out my thoughts, albeit to an agent. Writing is thinking. That is the most valuable thing I find about working projects with an agent helping me."

The agent isn't a coder-for-hire — it's a sparring partner for thinking. When the user writes their thoughts to the agent, they're clarifying their own mental model. When the agent challenges back, it sharpens further.

The Anti-Pattern: Eager Executor

The user described a recurring frustration: agents that skip the "pause and discuss" step. They execute ahead of their plan, making assumptions instead of asking. The user has to manually pause and ask "why that approach?" — then fork the session with the reasoning baked into WORK.md.

Solution: Workflows must instruct agents to present plans before executing, and pause when approaching decisions the user might care about.

Architectural Decision: Merge Moodboard + Whiteboard → Discuss

The current moodboard/whiteboard split assumes:

- User has fuzzy idea that needs extracting (moodboard)

- User needs formal plan presented (whiteboard)

- Sequential phases: Dream → Plan → Execute

But the user's actual pattern is:

- Context-rich starts ("I'm at this part, my question is...")

- Continuous dialogue where understanding and planning interweave

- Teaching moments emerge organically, not in a "moodboard phase"

Decision: Replace moodboard.md + whiteboard.md with unified discuss.md that adapts to context. Create questioning.md reference file with philosophy + techniques.

User Controls for Hyper-Focus Context Engineering

User explicitly wants gatekeeper control over artifact writes:

- Every write costs tokens and narrows context window

- Agent should ask "Want me to capture this?" not auto-write

- Logging is a deliberate act of curation, not automatic journaling

Next Actions:

1. Create gsd-lite/template/references/questioning.md — philosophy + techniques

2. Create gsd-lite/template/workflows/discuss.md — unified exploration workflow

3. Update execution.md — lighter, focused on doing (discuss handles thinking)

4. Remove or deprecate moodboard.md and whiteboard.md

Cherry on Top: The Journalism Prompt

User's exact prompt for logging discoveries (for agent reference):

> "Please include specific code snippet, reasoning, extended context in a journalism narrative style so that whoever with 0 context can onboard and pick up this key record of decision without ambiguity or friction. Add a cherry on top to include exact / synthesized example to explain the concepts / findings mentioned so that other weaker reasoning agents can grasp topic with little friction and ambiguity."

This is the standard for LOG entries in GSD-Lite.

[LOG-007] - [DECISION] - The standard length and outline for optimal gsd-lite artifacts

Current GSD-Lite Artifact Sizes

The Context Budget Reality

When an agent loads a workflow, here's what happens:

┌─────────────────────────────────────────────────────────────┐
│  AGENT CONTEXT WINDOW (let's say 80k target budget)         │
├─────────────────────────────────────────────────────────────┤
│                                                              │
│  System prompt / PROTOCOL.md     ~2,200 tokens (always)     │
│  WORK.md Current Understanding   ~500-1,000 tokens          │
│  Workflow file                   ~2,500-3,500 tokens        │
│  Reference files (if loaded)     ~1,000-2,000 tokens        │
│  ─────────────────────────────────────────────────────────  │
│  BASELINE OVERHEAD               ~6,200-8,700 tokens        │
│                                                              │
│  Remaining for:                                              │
│  - Conversation turns                                        │
│  - Code reading                                              │
│  - Tool outputs                                              │
│                                                              │
└─────────────────────────────────────────────────────────────┘

The Pit Stop philosophy says: Minimize baseline overhead so agents have more runway before needing a fork.

🎯 Recommended Artifact Specifications

1. `questioning.md` (NEW — Reference File)

Purpose: Philosophy + techniques for Socratic pair programming. Referenced by workflows, not loaded every session.

Proposed Outline:

# Questioning Guide (~120 lines)

## Philosophy (~15 lines)
- "Thinking partner, not interviewer"
- "Always Ask WHY Before HOW"
- Learning accelerator, not requirements gathering

## The Challenge Tone Protocol (~25 lines)
- Decision tree (A/B/C/D)
- When to use each tone
- Mermaid diagram

## Question Types (~20 lines)
- Motivation (why this exists)
- Concreteness (what it actually is)  
- Clarification (what they mean)
- Success (how you'll know)

## Techniques (~25 lines)
- Follow energy
- Challenge vagueness
- Make abstract concrete
- Know when to stop

## Anti-Patterns (~15 lines)
- Checklist walking
- Eager executor
- Shallow acceptance

## The Teaching Detour (~20 lines)
- The 10-star experience pattern
- When to offer: "Want me to explain this concept?"
- How to: explore, connect, distill, example

Why this size: Small enough to inline into workflow preambles. Large enough to be complete. OG GSD proved 142 lines is sufficient.

2. `discuss.md` (NEW — Unified Workflow, replaces moodboard + whiteboard)

Purpose: Adaptive exploration — handles both vision extraction AND plan presentation based on context.

Proposed Outline:

# Discuss Workflow (~280 lines)

## Purpose (~10 lines)
- Unified exploration: vision OR plan OR teaching
- Adapts to what user brings

## Entry Conditions (~10 lines)
- Default when no active execution task
- User signals: "let's discuss", "I have a question", "help me understand"

## Context Reading (~20 lines)
- Read WORK.md Current Understanding
- Detect: fresh start vs mid-project vs blocked

## Adaptive Routing (~30 lines)
- If fresh/fuzzy → Vision exploration mode
- If mid-project + question → Teaching/clarification mode  
- If blocked → Unblocking mode
- If ready to plan → Plan presentation mode

## Exploration Loop (~60 lines)
- Reference questioning.md philosophy
- 4 questions, then check pattern
- Challenge tone decision tree
- "Want me to capture this?" gates

## Teaching Detour Protocol (~40 lines)
- Detect unfamiliar concepts
- Offer: "Want me to explain [X]?"
- Pattern: explore → connect → distill → example
- Return to main thread

## Plan Presentation (~40 lines)
- When understanding is sufficient
- Present plan for approval
- User confirms or requests changes

## Exit Conditions (~20 lines)
- Transition to execution
- Checkpoint if pausing
- Capture decisions to WORK.md

## Anti-Patterns (~20 lines)
- Eager executor (skip discuss, start coding)
- Interrogation mode
- Auto-writing without permission

## Sticky Note (~30 lines)
- Required end-of-turn format

Why this size: Combines moodboard (406) + whiteboard (280) = 686 lines, but we're targeting ~280 because:

1. Redundancy removed (both had similar patterns)

2. Philosophy moved to questioning.md reference

3. Adaptive routing replaces rigid phases

3. `execution.md` (UPDATE — Lighter)

Current: 414 lines, ~2,778 tokens

Proposed: 300-350 lines, ~2,200 tokens (trim ~15-20%)

What to trim:

- Move questioning/blocking logic to reference questioning.md

- Simplify — execution should be about DOING, not exploring

- Add: "When stuck, transition to discuss.md"

📐 The Sizing Philosophy

The Test: If a workflow exceeds 400 lines, ask: "What can be extracted to a reference file?"

[LOG-008] - [EXEC] - Implemented Questioning Enhancement with "Mode-Based" Workflow - Task: QUESTION-001

Timestamp: 2026-02-04 11:30

Details:

Executed the plan from LOG-006 and LOG-007 to enhance GSD-Lite's questioning capability. This involved a significant architectural shift from "sequential phases" (Moodboard → Whiteboard → Execution) to "adaptive modes" (Discuss ↔ Execute) powered by a central questioning philosophy.

The shift:

- From: Rigid phases (moodboard.md for vision, whiteboard.md for planning)

- To: Unified discuss.md that adapts to context (Vision, Teaching, Unblocking, Planning)

Files Created/Updated:

1. src/gsd_lite/template/references/questioning.md (NEW - 143 lines)

- The "DNA" of Socratic pair programming.

- Contains: Golden Rule ("Why before How"), Challenge Tone Protocol, Teaching Detour.

2. src/gsd_lite/template/workflows/discuss.md (NEW - 289 lines)

- Replaces both moodboard.md and whiteboard.md.

- Adaptive entry: Reads WORK.md to decide if it's exploring vision, teaching a concept, or presenting a plan.

3. src/gsd_lite/template/workflows/execution.md (TRIMMED - 222 lines, down from 414)

- Focused purely on DOING.

- Trigger: "Let's pause and discuss" → forks to discuss.md.

Cherry on Top: The "Room vs Tool" Analogy

To explain why we moved to "Mode-Based" workflows:

Imagine a workshop:

- Phase-Based (Old): You have a "Thinking Room" and a "Building Room". To think, you must leave the building room, walk down the hall, and enter the thinking room. It's friction.

- Mode-Based (New): You are at the workbench (Execution). You put down the hammer and pick up the whiteboard marker (Discuss). You're in the same space, just switching tools.

In GSD-Lite:

- Old: "I'm stuck. Let me load moodboard.md..." (Context switch overhead)

- New: "I'm stuck. Let's discuss." (Seamless mode switch, same context)

Code Snippet: The Adaptive Routing Logic (from discuss.md)

This logic allows one workflow to handle multiple user needs:

## Context Reading

Read WORK.md Current Understanding to detect state:

| State Detected | Mode to Enter |
|----------------|---------------|
| No active phase, fuzzy vision | Vision Exploration |
| Active phase, user has question | Teaching/Clarification |
| Active task, user is blocked | Unblocking |
| Understanding complete, ready to plan | Plan Presentation |

Status: Implementation complete. Ready to update router (PROTOCOL.md).

[LOG-009] - [DECISION] - Protocol v2.1: Questioning DNA and Mode-Based Architecture - Task: PROTO-001

Timestamp: 2026-02-04 12:00

Context:

The GSD-Lite protocol was upgraded from a sequential phase-based model to an adaptive mode-based architecture. This shift reflects the user's "Driver/Navigator" persona, where learning and mental model validation are as important as task execution.

The Finding: Questioning is the DNA, Not a Phase

The previous architecture separated vision extraction (moodboard.md) and planning (whiteboard.md) into rigid sequences. However, real-world sessions showed that questioning, teaching, and unblocking are continuous needs that emerge organically during both planning and execution.

Decision: Inline Philosophy for Zero Friction

We decided to append the full questioning.md content (Philosophy, Challenge Tones, Teaching Detours) directly to PROTOCOL.md.

- Reasoning: 1.2k tokens is a minor cost to ensure the agent ALWAYS operates as a Socratic "Thinking Partner."

- Benefit: It eliminates the friction of an agent needing to remember to read a separate reference file, ensuring the "Golden Rule" (Why Before How) is applied to every turn.

Architectural Change: From Rooms to Tools

- Old (Sequential): Moodboard → Whiteboard → Execution (The "Moving Rooms" model).

- New (Adaptive): discuss.md ↔ execution.md (The "Workbench Tools" model).

- discuss.md now acts as a Swiss Army knife for exploration, teaching, and planning.

Cherry on Top: The "Socratic Workbench" Analogy

Imagine a craftsman's workshop:

- Sequential Model: To think about a design, you must leave the workbench and walk to a separate office. By the time you get back to the wood, the inspiration might be cold.

- Socratic Workbench (GSD-Lite v2.1): You are at the bench with the wood (Execution). You feel a knot in the grain (Unfamiliar Concept). Instead of walking away, you simply put down the chisel and pick up the measuring tape and pencil (Discuss). You analyze, teach yourself about the grain, and then pick the chisel back up with new understanding.

In this model, the Questioning DNA in the PROTOCOL is like the bright overhead lighting in the workshop — it doesn't matter which tool you're holding; you can always see the "Why" behind the work.

Synthesized Example: The "Mode Switch" in Action

Scenario: User is executing a task to refactor a Python class.

1. Execution Mode: Agent proposes a refactor using a Protocol.

2. User Question: "Why use Protocol here instead of abc.AbstractBaseClass?"

3. Mode Switch (Automatic): The agent identifies this as a Teaching Detour (from the DNA in PROTOCOL). It switches to Discuss Mode logic.

4. Teaching Detour:

- Explore: Shows typing.Protocol usage in the codebase.

- Connect: Relates it to "Duck Typing" which the user knows.

- Distill: "ABCs are 'is-a' relationships; Protocols are 'behaves-like' relationships."

- Example: class Flyer(Protocol): def fly(self): ...

5. Resume Execution: "Ready to proceed with the Protocol approach, or want to stick to ABCs?"

Status: PROTOCOL.md v2.1 committed. Moodboard and Whiteboard deprecated and removed. Architecture is now mode-based.

[LOG-010] - [DECISION] - Universal Onboarding & The End of Sticky Notes - Task: PROTO-002

Timestamp: 2026-02-04 13:00

Context:

A critical onboarding gap was discovered in the "Fork & Resume" workflow. If a user forks a session and invokes discuss.md directly (e.g., "Let's discuss this loop"), the fresh agent would read PROTOCOL.md and WORK.md but miss PROJECT.md (Vision) and ARCHITECTURE.md (Tech Stack).

This created "Forwarder Agents" — agents that mechanically execute the workflow but lack the deep context to ask intelligent, grey-area probing questions (the core value of GSD).

The Solution: Universal Onboarding

We moved the onboarding responsibility from individual workflows (like progress.md) to the PROTOCOL.md router itself. Now, every fresh session begins with a mandatory boot sequence:

1. PROTOCOL.md

2. PROJECT.md (Vision - The Why)

3. ARCHITECTURE.md (Tech - The How)

4. WORK.md (State - The Where)

The Trade-off:

We centralized onboarding (DRY) rather than duplicating it in every workflow file. This ensures consistent "Smart Agent" behavior regardless of entry point.

The Cleanup: Decomissioning Sticky Notes

With artifacts now mature and the protocol robust, the "Sticky Note Protocol" (appending a status block to every turn) was deemed redundant noise.

- Why: The artifacts (WORK.md, INBOX.md) are the source of truth.

- Change: Removed gsd-status requirements from all 9 workflow templates.

- Result: Cleaner chat interface, fewer tokens used per turn.

Cherry on Top: The "Pilot's Checklist" Analogy

To explain why Universal Onboarding matters:

Imagine a pilot (Agent) taking over a plane mid-flight (The Project) from another pilot (Previous Session).

* Old Way (Workflow-Dependent):

* Pilot jumps in.

* If they check progress.md: They read the flight plan and weather.

* If they just grab the yoke (discuss.md): They start flying without knowing the destination or the aircraft type. They fly purely on instruments. It's safe, but they can't make strategic decisions.

* New Way (Universal Onboarding):

* Mandatory Pre-Flight: Before touching the controls, EVERY pilot reads:

1. PROJECT.md (Destination): "We are going to Hawaii."

2. ARCHITECTURE.md (Aircraft Manual): "This is a Boeing 747, not a Cessna."

3. WORK.md (Instruments): "We are currently at 30,000ft, autopilot on."

* Now, when the user says "Let's dive," the pilot knows "Wait, a 747 can't dive like a Cessna," instead of just executing the dive.

Code Snippet: The New Protocol Boot Sequence

src/gsd_lite/template/PROTOCOL.md:

## Session Start (Universal Onboarding)

**Every fresh session follows this boot sequence — regardless of which workflow will run.**

1. **Read PROTOCOL.md**
2. **Read PROJECT.md** (if exists) — Understand the project vision and "why"
3. **Read ARCHITECTURE.md** (if exists) — Understand the codebase structure
4. **Read WORK.md Current Understanding** — Understand current state
5. **Load appropriate workflow**

Status: src/gsd_lite/template/ updated. Sticky notes removed. Protocol v2.2 active.

[LOG-011] - [DECISION] - The Frictionless Install Paradox: Global vs Local Distribution - Task: DIST-001

Timestamp: 2026-02-04 14:00

Context:

The user requested a "frictionless" distribution model for gsd-lite, similar to npx get-shit-done-cc. The goal is to allow users to run @gsd-lite in any directory without manually scaffolding a gsd-lite/ folder first. However, the AGENTS.md template (the brain) was hardcoded to look for workflows in a local project directory (gsd-lite/template/workflows/).

The Paradox:

- Local Mode (Eject): Project is self-contained. Agent reads ./gsd-lite/template/workflows/. Good for customization, bad for friction (requires init).

- Global Mode (Binary): Zero setup. Agent reads ~/.config/opencode/command/gsd-lite/. Good for friction, but breaks the static template.

The Decision: Dynamic Patching at Install Time

We chose the Global Binary model for the default experience. To solve the pathing issue, the installer CLI will dynamically patch the agent's instructions during installation.

- Source: src/gsd_lite/template/AGENTS.md (contains relative path placeholder).

- Destination: ~/.config/opencode/agents/gsd-lite.md (contains absolute path).

Rationale:

This preserves the "Project Specific Protocol" philosophy (Option A) if a user wants to eject (they can manually copy files), but enables the "Frictionless" experience (Option B) by default. The agent is smart enough to follow the path it is given.

Cherry on Top: The "Suitcase vs Hotel" Analogy

To explain the difference:

- Local Mode (Eject): Like moving into an apartment. You buy furniture (workflows) and put it in your living room (./gsd-lite). It's yours, you can paint it, but moving in takes effort.

- Global Mode (Frictionless): Like staying at a hotel. The furniture (workflows) is provided by the hotel (~/.config). You just show up with your suitcase (Code) and start working.

We want gsd-lite to feel like a 5-star hotel: Instant check-in, everything provided.

Code Snippet: The Dynamic Patch

The installer will perform this transformation:

Before (Template):

5. **Load appropriate workflow** — Based on current_mode in WORK.md
   (Reads from gsd-lite/template/workflows/)

After (Installed Agent):

5. **Load appropriate workflow** — Based on current_mode in WORK.md
   (Reads from /Users/username/.config/opencode/command/gsd-lite/)

Next Action: Implement gsd-lite install --opencode command in src/gsd_lite/__main__.py that copies workflows and applies this patch.

[LOG-012] - [DISCOVERY] - Reverse-Engineering get-shit-done-cc: The Anatomy of a Frictionless Installer - Task: DIST-002

Timestamp: 2026-02-04 15:30

The Story So Far

The user's current workflow to update gsd-lite after each iteration is painfully manual:

# Step 1: Navigate to opencode config directory
cd ~/.config/opencode

# Step 2: Run CLI to scaffold template files (creates gsd-lite/ subdirectory)
uvx --no-cache gsd-lite --update

# Step 3: Manually copy the agent definition
cp gsd-lite/template/AGENTS.md ./agents/gsd-lite.md

# Step 4: Manually copy all workflow files
cp -r gsd-lite/template/workflows/* ./command/gsd-lite/

This is the "IKEA furniture" experience: you get the parts, but assembly is on you. We want the "hotel concierge" experience: everything set up before you arrive.

The Reference Implementation: get-shit-done-cc

We cloned the upstream GSD installer (get-shit-done-cc) to tmp/get-shit-done/ and reverse-engineered its installation logic. The key file is bin/install.js — a ~1500-line Node.js script that handles multi-runtime installation.

Critical Discovery #1: Directory Structure Varies by Runtime

Each runtime (Claude Code, OpenCode, Gemini) uses slightly different conventions:

Why this matters: A naive installer that assumes command/ everywhere will fail silently on Claude Code — the workflows land in the wrong directory and the agent can't find them.

Evidence from get-shit-done-cc source (bin/install.js, line ~1385):

console.log(`  ${yellow}Which runtime(s) would you like to install for?${reset}\n\n  
  ${cyan}1${reset}) Claude Code ${dim}(~/.claude)${reset}
  ${cyan}2${reset}) OpenCode    ${dim}(~/.config/opencode)${reset} - open source, free models
  ${cyan}3${reset}) Gemini      ${dim}(~/.gemini)${reset}
  ${cyan}4${reset}) All

Critical Discovery #2: OpenCode's Current Installation Structure

We inspected the user's live OpenCode config at ~/.config/opencode/ using the opencode-fs MCP tool:

~/.config/opencode/
├── agents/
│   └── gsd-lite.md          # ← The agent brain (AGENTS.md renamed)
├── command/
│   └── gsd-lite/            # ← Workflow files live here
│       ├── checkpoint.md
│       ├── discuss.md
│       ├── execution.md
│       ├── housekeeping.md
│       ├── map-codebase.md
│       ├── new-project.md
│       └── progress.md
├── gsd-lite/                 # ← Scaffolded by current CLI (redundant for global install)
│   └── template/
│       └── ...
├── hooks/
├── opencode.json
└── settings.json

Key insight: The gsd-lite/ subdirectory inside ~/.config/opencode/ is an artifact of the current --update command, which scaffolds a full project structure. For global installation, we don't need this — we only need to populate agents/ and command/gsd-lite/.

Critical Discovery #3: AGENTS.md is Pre-Built (No Dynamic Assembly)

The user clarified that AGENTS.md in the source tree (src/gsd_lite/template/AGENTS.md) is already the final artifact — it's PROTOCOL.md with a 22-line YAML frontmatter prepended. There's no build step needed; we simply copy it.

The frontmatter (first 16 lines of src/gsd_lite/template/AGENTS.md):

---  
description: Lightweight agent with minimal tool access  
tools:  
  read: false
  edit: false
  bash: false
  grep: false
  glob: false
  list: false

permission:  
  task:  
    "*": "deny"
---  

This frontmatter is OpenCode-specific but works across all platforms (Claude and Gemini also support frontmatter).

The Scoped Plan: MVP Installer (OpenCode Global Only)

Scope for v1:

- ✅ OpenCode runtime only

- ✅ Global install to ~/.config/opencode/ only

- ❌ Claude Code (future)

- ❌ Gemini (future)

- ❌ Local install (future)

- ❌ Version checking / updates (future)

What the installer does:

uvx gsd-lite install

# Output:
#   Installing gsd-lite to ~/.config/opencode/ ...
#   ✓ Copied agents/gsd-lite.md
#   ✓ Copied command/gsd-lite/ (7 workflows)
#   Done! Run @gsd-lite in any project to start.

File Operations:

Implementation Sketch:

# src/gsd_lite/__main__.py (new install command)

import shutil
from pathlib import Path
from importlib.resources import files

def install():
    """Install gsd-lite globally to ~/.config/opencode/"""
    
    # 1. Resolve paths
    opencode_config = Path.home() / ".config" / "opencode"
    agents_dir = opencode_config / "agents"
    command_dir = opencode_config / "command" / "gsd-lite"
    
    # 2. Get package template location
    template_dir = files("gsd_lite") / "template"
    
    # 3. Ensure target directories exist
    agents_dir.mkdir(parents=True, exist_ok=True)
    command_dir.mkdir(parents=True, exist_ok=True)
    
    # 4. Copy agent definition
    shutil.copy(template_dir / "AGENTS.md", agents_dir / "gsd-lite.md")
    print("✓ Copied agents/gsd-lite.md")
    
    # 5. Copy all workflow files
    workflows_src = template_dir / "workflows"
    for workflow in workflows_src.iterdir():
        if workflow.suffix == ".md":
            shutil.copy(workflow, command_dir / workflow.name)
    print(f"✓ Copied command/gsd-lite/ ({len(list(command_dir.glob('*.md')))} workflows)")
    
    print("\nDone! Run @gsd-lite in any project to start.")

Open Question (Deferred): Dynamic Path Patching

LOG-011 mentioned patching AGENTS.md to replace relative workflow paths with absolute paths. After discussion, we realized:

1. The current AGENTS.md doesn't hardcode a workflow path — it says "Load appropriate workflow" generically.

2. OpenCode resolves command/gsd-lite/ relative to its config directory automatically.

3. No patching needed for global install — the runtime handles path resolution.

If we later implement local install (./.opencode/), we may need patching. Deferred for now.

Cherry on Top: The "Concierge vs IKEA" Mental Model

The MVP delivers the concierge experience for OpenCode users. Multi-runtime support is the "international hotel chain expansion" — same service, different locations.

Next Action:

1. Switch to execution mode

2. Implement install command in src/gsd_lite/__main__.py

3. Test with uvx --no-cache gsd-lite install

4. Verify files land in correct locations

[LOG-013] - [DECISION] - Unified Installation Model: Same Output, Different Location - Task: DIST-002

Timestamp: 2026-02-04 16:45

The Problem We Were Creating

Initial thinking led to a fragmented command structure:

- install for global brain only

- --local for project artifacts only

- --update for refreshing templates

- --force for overwriting

This created cognitive overhead: "Which command do I need?" and different outputs for different commands.

The Insight

User asked: "Are we just complicating ourselves? What if we unified everything like get-shit-done-cc does?"

The answer: Yes, we were overcomplicating.

The Unified Model

uvx gsd-lite install           # → ~/.config/opencode/ + gsd-lite/
uvx gsd-lite install --local   # → ./.opencode/ + ./gsd-lite/

Both commands create identical structure, just at different roots:

{target}/
├── .opencode/                    # OpenCode config (or ~/.config/opencode/ for global)
│   ├── agents/
│   │   └── gsd-lite.md           # Agent definition (AGENTS.md)
│   └── command/
│       └── gsd-lite/             # Workflows
│           ├── discuss.md
│           ├── execution.md
│           ├── checkpoint.md
│           └── ...
│
└── gsd-lite/                     # Artifacts directory
    ├── WORK.md                   # ← SKIP if exists (preserve user data)
    ├── INBOX.md                  # ← SKIP if exists
    ├── PROJECT.md                # ← SKIP if exists
    ├── ARCHITECTURE.md           # ← SKIP if exists
    ├── HISTORY.md                # ← SKIP if exists
    └── template/                 # ← ALWAYS overwrite (that's how updates work)
        ├── PROTOCOL.md
        ├── workflows/
        ├── references/
        └── VERSION

The Smart Check Pattern (already exists in current code):

- User artifacts → Skip if exist, scaffold if new

- Template/config files → Always overwrite (enables updates)

Why This is Better

The "Global Brain + Local Memory" Mental Model

GLOBAL INSTALL (~/.config/opencode/)
├── Brain: agents/ + command/gsd-lite/    ← Available to all projects
└── Reference: gsd-lite/template/          ← For manual reference

LOCAL INSTALL (./)
├── Brain: .opencode/agents/ + command/   ← Project-specific override
└── Memory: gsd-lite/                      ← Active session artifacts

Typical User Journey:

1. uvx gsd-lite install — One-time global setup

2. Enter project, uvx gsd-lite install --local — Scaffold project artifacts

3. Start OpenCode, @gsd-lite — Global brain reads/writes local memory

4. Later: uvx gsd-lite install again — Updates global brain to latest version

[LOG-014] - [MILESTONE] - Implemented Unified Installer Logic - Task: DIST-002

Timestamp: 2026-02-04 17:15

Observation

The previous CLI was a simple scaffolder that dumped everything into ./gsd-lite. To support the "Concierge" experience, we needed logic that could target both the Global OpenCode config (~/.config/opencode) and a Local Project (./).

Resolution

Refactored src/gsd_lite/__main__.py to use typer subcommands.

- Global Install (gsd-lite install):

- Targets ~/.config/opencode/

- Installs Agent → agents/gsd-lite.md

- Installs Workflows → command/gsd-lite/

- Installs Reference Templates → gsd-lite/template/

- Local Install (gsd-lite install --local):

- Targets ./

- Creates .opencode/agents/ and .opencode/command/ (Brain)

- Scaffolds gsd-lite/WORK.md etc. (Memory)

Smart Overwrite Logic

- Templates/Workflows: ALWAYS overwritten (ensures updates apply)

- User Artifacts (WORK.md, PROJECT.md): SKIPPED if exist (preserves user data)

- Force Flag (--force): Overwrites EVERYTHING (nuclear option)

Key Code Snippet

if local:
    base_root = Path.cwd()
    config_root = base_root / ".opencode"
else:
    base_root = Path.home() / ".config" / "opencode"
    config_root = base_root

# Unified Structure
agents_dir = config_root / "agents"
command_dir = config_root / "command" / "gsd-lite"
artifacts_dir = base_root / "gsd-lite"

Next Steps

- Verify installation on live system

- Update documentation to reflect new install command usage

[LOG-015] - [TOOLING] - Context Map: Solving the "Token Blindness" Problem in Large Artifacts - Task: TOOLING-001

Executive Summary:

We developed analyze_context.py, a utility that solves the "token blindness" problem when agents interact with massive Markdown artifacts (like WORK.md). Instead of reading the full 65k token file (which overflows context) or blindly grepping headers (which loses hierarchy), this tool generates a Context Map: a lightweight (~1k token) mirror of the document that preserves IDE navigability, visualizes section weights, and provides "HEAD/TAIL" previews for surgical reading.

1. The Problem: "Token Blindness" in Large Artifacts

When an agent needs to understand the project state from a massive log file (e.g., WORK.md at 65,000+ tokens), it faces a trilemma:

1. Read Full File: Costs $0.50+ per turn, overflows context window, drowns reasoning in noise.

2. Blind Grep: grep "^#" reveals headers but hides the content weight. Is "Section 3" a 2-line note or a 5,000-token specification? The agent doesn't know until it commits to reading.

3. Table Summary: A standard table of contents breaks the visual hierarchy (indentation) and IDE collapsing, making it hard to "scan" the tree structure.

We needed a way to see the shape and weight of the forest before cutting down trees.

2. The Research: Leveraging Source Maps

We investigated how the industry handles this. A search for "LLM context window management document chunking" revealed two key patterns:

* LangChain's MarkdownHeaderTextSplitter: Parses Markdown by headers (#, ##) to preserve structural integrity when chunking for RAG.

"Offloading Large Tool Results": A pattern where agents see a reference and a preview* of large data, rather than the raw data itself.

Source:

> "MarkdownHeaderSplitter specifically designed for Markdown documents... divides content based on headers... preserving structural integrity." — LangChain Documentation [1]

We realized we could use markdown-it-py (a CommonMark-compliant parser) to access the Source Map (token.map) of each header, giving us precise line numbers and the ability to extract just the start (HEAD) and end (TAIL) of each section.

3. The Solution: `analyze_context.py`

We built a script that transforms a massive Markdown file into a "Context Map".

Key Features:

1. Mirrored Hierarchy: Uses real Markdown headers (#, ##) in the output, so the map is collapsible in your IDE just like the original file.

2. Weight visualization: Calculates token counts (using tiktoken) for every section, allowing the agent to budget its context usage.

3. Surgical Previews: Extracts the first 3 lines (HEAD) and last 2 lines (TAIL) of every section. The agent can read the preview and decide if it needs the full content.

Code Snippet: The Preview Logic

def extract_preview(lines, head_count=3, tail_count=2):
    """
    Extract HEAD and TAIL preview from a list of lines.
    Returns (head_text, tail_text) as strings.
    """
    # Filter out empty lines for meaningful preview
    content_lines = [l.rstrip() for l in lines if l.strip()]
    
    if not content_lines:
        return ("(empty section)", "")
    
    # HEAD: first N content lines (skip the header line itself)
    head_lines = content_lines[1:head_count+1] if len(content_lines) > 1 else content_lines[:head_count]
    head_text = "\n".join(head_lines) if head_lines else "(no content after header)"
    
    # TAIL: last N content lines (avoid overlap with head)
    if len(content_lines) > head_count + tail_count + 1:
        tail_lines = content_lines[-tail_count:]
        tail_text = "\n".join(tail_lines)
    else:
        tail_text = ""  # Section too short, HEAD covers it
    
    return (head_text, tail_text)

4. The Data Flow Diagram

graph TD
    A[Input: WORK.md] -->|Read File| B(Raw Text)
    B -->|markdown-it-py| C{Token Stream}
    C -->|Identify Headers| D[Header Metadata]
    D -->|Get Line Ranges| E[Section Slicing]
    
    subgraph Analysis
    E -->|tiktoken| F[Token Count]
    E -->|Heuristics| G[Extract HEAD/TAIL]
    end
    
    F --> H[Context Map Generator]
    G --> H
    H -->|Format with # Headers| I[Output: context_map.md]
    
    style A fill:#f9f,stroke:#333
    style I fill:#9f9,stroke:#333

5. Example Output (Synthesized)

Input: A 65,000 token WORK.md.

Output: A ~1,500 token context_map.md that looks like this:

# Context Map: eval/gsd-lite/WORK.md
**Total:** 254,551 chars | **~64,864 tokens**

## 🎯 Top Heavy Sections
| Rank | Section | Tokens |
|------|---------|--------|
| 1 | `## 2.2 Layer 2: Semantic Layer` | 2,663 |

---

# Document Structure

## 1. Current Understanding (Read First)
> 📊 **900 tokens** | Lines 26-85

**HEAD:**

current_mode: execution

active_task: TASK-042 - Implement cascading filters...

**TAIL:**

...next_action: Complete filter persistence in localStorage.

This map allows an agent to say: "I see Section 1 is 900 tokens and relevant. I will read lines 26-85. I see Section 2.2 is 2,600 tokens and irrelevant. I will skip it."

6. Citations & Sources

1. LangChain Documentation, "MarkdownHeaderTextSplitter". https://python.langchain.com/docs/modules/data_connection/document_transformers/markdown_header_metadata

2. markdown-it-py, "Token Stream and Source Maps". https://markdown-it-py.readthedocs.io/en/latest/tokens.html

3. Tiktoken, "OpenAI Tokenizer". https://github.com/openai/tiktoken

[LOG-016] - [DECISION] - Stateless-First Architecture: Every Turn Generates a Handoff Packet - Task: PROTOCOL-STATELESS

Timestamp: 2026-02-06

Decision ID: DECISION-016

Status: APPROVED

Executive Summary

GSD-Lite adopts stateless-first architecture as a core philosophy. Every agent turn ends with a structured handoff packet that enables any future agent to continue with zero chat history. This embraces the user's micro-forking workflow as a first-class pattern, not a workaround.

The One-Liner: Treat every agent turn as its end-of-life. Generate a handoff so any future agent can pick up exactly where you left off.

Part 1: The Problem — Context Rot and Manual Curation

1.1 The Production Evidence

This decision emerged from real production usage documented in eval/gsd-lite/WORK.md:

- Size: 254,551 characters | ~64,864 tokens | 5,552 lines

- Logs: 56 LOG entries (LOG-001 through LOG-056)

- Analysis: See eval/gsd-lite/WORK_context_analysis.md for token breakdown by section

Citation: File metadata from mcp_tools_fs_get_file_info("eval/gsd-lite/WORK.md"):

Total Lines: 5,552
Total Characters: 254,551
Estimated Tokens: 63,637

1.2 The Micro-Forking Workflow Discovered

The user discovered a powerful context management pattern through production use:

flowchart LR
    subgraph SESSION1["SESSION 1 (tokens: 0 → 60k)"]
        A[Start] --> B[Discuss 2-3 turns]
        B --> C[Agent writes LOG-XXX]
        C --> D[Context rising...]
        D --> E[Hit 60-80k threshold]
    end
    
    E --> F["🔀 FORK back to turn 1-2"]
    
    subgraph SESSION2["SESSION 2 (tokens: 0 → fresh)"]
        F --> G["Paste: 'Read LOG-XXX'"]
        G --> H[Agent onboards from log]
        H --> I[Continue with optimal context]
        I --> J[Context rising...]
    end
    
    J --> K["🔀 FORK again"]
    K --> G

Why this works:

- LLMs perform optimally in the 20k-80k token range (user's empirical observation with Claude Opus 4.5 and Gemini Pro 3)

- Logs written by a strong reasoning model are better context than raw conversation — synthesized, polished, agent-optimized

- The micro-fork is a context upgrade, not a workaround

1.3 The Friction Point

The problem at scale: At LOG-001, "go read LOG-001" is sufficient. At LOG-100, the fresh agent needs a curated path through the decision graph — not just "read the last one."

Prior attempt (LOG-045 in eval/gsd-lite/WORK.md, lines 4325-4398):

A META log entry with "Agent Onboarding Paths" organized by task type:

- Path A: Building a new page (~4k tokens)

- Path B: Understanding data architecture (~6k tokens)

- Path C: Resuming mid-task (~2k tokens)

- Path D: Adding filters (~3k tokens)

- Path E: Full context deep dive (~15k tokens)

Citation: eval/gsd-lite/WORK.md lines 4325-4398 (LOG-045)

Why LOG-045 is insufficient:

1. Static snapshot — Valid when written, stale as logs accumulate

2. Separate artifact — Another thing to remember and maintain

3. Manual trigger — User must remember to ask agent to read it

Part 2: The Solution — Stateless-First Handoff Packets

2.1 Core Philosophy

Principle: Every agent turn is potentially its last. The agent must generate a handoff packet that enables any future agent to continue with zero chat history.

No exceptions: Even Turn 1. Even mid-discussion. The user owns context management via micro-forking.

Analogy — The Stateless Server:

Think of each agent turn like an HTTP request to a stateless server:

- Server doesn't remember previous requests

- Each request contains everything needed to process it

- Response contains everything needed for the next request

The handoff packet is like a JWT token — self-contained context that travels with the user across session boundaries.

2.2 Two-Layer Structure

The handoff packet contains two layers of context:

Why two layers:

- Layer 1 answers: "How do I continue this specific task?"

- Layer 2 answers: "How do I pivot to something completely different?"

flowchart TB
    subgraph HANDOFF["📦 Stateless Handoff Packet"]
        L1["Layer 1: Local Context<br/>───────────────<br/>Last action: LOG-056<br/>Chain: LOG-056 ← LOG-055 ← LOG-052<br/>Next: Implement on /explore"]
        L2["Layer 2: Global Context<br/>───────────────<br/>Architecture: LOG-031, LOG-024<br/>Patterns: LOG-044, LOG-055<br/>Data Flow: LOG-038"]
        PATHS["Fork Paths<br/>───────────────<br/>• Continue → L1 only<br/>• Pivot → L2 + new topic"]
    end
    
    L1 --> PATHS
    L2 --> PATHS

2.3 Layer 1 — Dependency Chain Tracing

The agent traces backwards from the current task to surface the decision lineage.

Example from eval/gsd-lite/WORK.md:

Current task: Implement cascading filters on /explore (TASK-006)
                    ↓ depends on
LOG-056: Implemented cascading filters (last execution)
                    ↓ depends on
LOG-055: DECISION-020 — Cascading via run_inline_query (the pattern)
                    ↓ depends on
LOG-052: DECISION-018 — Dual filter sections with mutex (the architecture)
                    ↓ depends on
LOG-048: DECISION-016 — Dashboard-defined filters (the foundation)

In handoff format:

→ Dependency chain: LOG-056 ← LOG-055 ← LOG-052 ← LOG-048

2.4 Layer 2 — Key Events Index as Source of Truth

The existing ## 2. Key Events Index section in WORK.md becomes the canonical source for global context.

Reframe from: "Query Accelerator" (nice-to-have optimization)

Reframe to: "Project Foundation" (required for handoff packets)

New structure:

## 2. Key Events Index (Project Foundation)

<!-- 
CANONICAL source of truth for Layer 2 of stateless handoff packets.
Agents MUST pull global context from here.
Human-curated: agent proposes additions, human approves.
-->

### Architecture Decisions
- LOG-031: Two Galaxies — Single project, dual Looker models
- LOG-024: Mirror Clone — Full data ownership via dataset cloning

### Pattern Decisions
- LOG-044: Dashboard Config — Looker dashboards as first-class artifacts
- LOG-055: Cascading Filters — Dynamic options via run_inline_query

### Data Flow Decisions
- LOG-038: Source Divergence — App queries looker-partners, not bigquery-public-data

Part 3: The Canonical Handoff Format

3.1 Full Format (Option C — Chosen)

After evaluating three format options (Minimal 3-4 lines, Structured 6-7 lines, Full 10-12 lines), the user chose Full Format because:

- Sessions are only 5-8 turns before forking anyway

- 10-12 lines × 8 turns = ~80-100 lines total — manageable

- Full context every turn means no guessing what's missing

The canonical format:

---
📦 STATELESS HANDOFF

**Layer 1 — Local Context:**
→ Last action: [LOG-XXX (brief description)]
→ Dependency chain: [LOG-XXX ← LOG-YYY ← LOG-ZZZ]
→ Next action: [specific next step]

**Layer 2 — Global Context:**
→ Architecture: [from Key Events Index]
→ Patterns: [from Key Events Index]
→ Data Flow: [from Key Events Index]

**Fork paths:**
- Continue execution → [specific logs]
- Discuss [topic] → [specific logs]
- Pivot to new topic → [L2 refs] + state your question

3.2 Turn-Type Variations

The structure stays rigid. The content adapts based on turn type:

Mid-Discussion (no decision yet):

---
📦 STATELESS HANDOFF

**Layer 1 — Local Context:**
→ Status: Discussing [topic] — no decision yet
→ Key refs from discussion: [LOG-XXX, LOG-YYY]
→ Resume: Restate your position on [open question]

**Layer 2 — Global Context:**
→ Architecture: [from Key Events Index]
→ Patterns: [from Key Events Index]

**Fork paths:**
- Continue discussion → restate your question
- Pivot to new topic → L2 + state your question

Post-Decision (DECISION logged):

---
📦 STATELESS HANDOFF

**Layer 1 — Local Context:**
→ Last action: LOG-XXX (DECISION-NNN: [title])
→ Dependency chain: LOG-XXX ← LOG-YYY ← LOG-ZZZ
→ Next action: [implementation step]

**Layer 2 — Global Context:**
→ Architecture: [from Key Events Index]
→ Patterns: [from Key Events Index]

**Fork paths:**
- Continue execution → L1 only (~3 logs)
- Discuss implementation → L1 + [relevant pattern log]
- Pivot to new topic → L2 + state your question

Teaching Detour:

---
📦 STATELESS HANDOFF

**Layer 1 — Local Context:**
→ Status: Teaching detour on [concept]
→ Task paused at: LOG-XXX ([last exec])
→ Resume: [LOG refs] → [next action]

**Layer 2 — Global Context:**
→ [Relevant sections from Key Events Index]

**Fork paths:**
- Resume execution → LOG-XXX + LOG-YYY
- Continue learning → [concept-specific logs]
- Pivot to new topic → L2 + state your question

First Turn (just forked in):

---
📦 STATELESS HANDOFF

**Layer 1 — Local Context:**
→ Onboarded via: [LOG-XXX (how you got here)]
→ Current action: [what you're doing this turn]
→ Will log as: LOG-YYY (on completion)

**Layer 2 — Global Context:**
→ [Full Key Events Index — user just forked, may pivot]

**Fork paths:**
- Continue after this turn → await LOG-YYY completion
- Pivot mid-task → L2 + state new direction

3.3 The Rigid Rules

These aspects are locked down for consistency across providers (Claude Opus, Gemini Pro 3):

Part 4: Key Events Index Maintenance

4.1 When to Update

The Key Events Index should be updated when:

- User explicitly asks: "update the key events index"

- Agent notices a DECISION log with project-wide impact

- At checkpoint time (optional housekeeping step)

4.2 How to Scan

# Find all DECISION logs
grep "^\(###\|####\) \[LOG-.*\[DECISION" WORK.md

# Find superseded entries
grep -i "supersede" WORK.md

# Cross-reference against current Key Events Index
# (Agent reads Section 2 and compares)

4.3 Inclusion/Exclusion Criteria

Include if:

Exclude if:

4.4 Update Proposal Format

Agent proposes, user approves:

📋 KEY EVENTS INDEX UPDATE PROPOSAL

**Add:**
- LOG-055: Cascading filters via run_inline_query → Category: Patterns
  Reason: Establishes reusable pattern for all filter implementations

**Remove:**
- LOG-018: Pivot to Public Data (superseded by LOG-024)
  Reason: LOG-024 captures the final decision with full context

**Recategorize:**
- LOG-038: Move from "Data Flow" to "Architecture"
  Reason: It's more about system boundaries than data movement

---
Approve these changes? (Yes / Adjust / Skip)

Part 5: Anti-Patterns (Bad Handoff Examples)

5.1 Too Vague

📦 HANDOFF: We discussed filters. Read the recent logs.

Problems:

- "Recent logs" — which ones? Last 3? Last 10?

- No specific LOG-XXX references

- No actionable next step

- Fresh agent has to guess

5.2 Wall of Text

📦 HANDOFF: We implemented cascading filters using the run_inline_query 
pattern as discussed in LOG-055 which was based on the decision we made 
in LOG-052 about dual filter sections which itself built on LOG-048 about 
dashboard-defined filters and the data architecture from LOG-031 and 
LOG-024 which superseded LOG-018 and LOG-020...

Problems:

- No structure — can't scan quickly

- Buries actionable items in prose

- User has to parse a paragraph to find what matters

5.3 Missing Layer 2

📦 HANDOFF
→ Last: LOG-056
→ Next: Apply pattern

Problems:

- Fine for task continuation

- Useless if user wants to pivot to new topic

- Fresh agent on different task would be lost

5.4 Stale References

📦 HANDOFF
→ Data strategy: Read LOG-018
→ Architecture: Read LOG-020

Problems:

- LOG-018 was superseded by LOG-024

- LOG-020 was superseded by LOG-024

- Would actively mislead a fresh agent

5.5 "Read Everything"

📦 HANDOFF
→ Full context: LOG-001 through LOG-056

Problems:

- Defeats the purpose of curation

- 64k tokens of context is exactly what we're trying to avoid

- Not a handoff, it's an abdication

5.6 Inconsistent Format Across Turns

Turn 5:

HANDOFF: LOG-005, then LOG-003

Turn 12:

📦 Stateless Packet
- Layer 1: LOG-012
- Layer 2: See Key Events

Turn 23:

## Handoff Notes
Last action was implementing the thing from LOG-023...

Problems:

- User can't build muscle memory

- Different delimiters/headers each time

- Hard to scan across sessions

- Which format is "right"?

Part 6: Implementation Plan

Task: PROTOCOL-STATELESS-001 — Add Stateless-First Section to PROTOCOL.md

Location: New section after "Context Lifecycle" in gsd-lite/template/PROTOCOL.md

Content to add:

1. Philosophy statement (every turn is potentially last)

2. Two-layer structure explanation

3. Canonical handoff format

4. Turn-type variations

5. Rigid rules table

6. Reference to Key Events Index as Layer 2 source

Why PROTOCOL.md: It's the agent entrypoint — loaded on every session start during Universal Onboarding.

Task: PROTOCOL-STATELESS-002 — Update Key Events Index Header

Location: ## 2. Key Events Index section in WORK.md template

Changes:

1. Rename to ## 2. Key Events Index (Project Foundation)

2. Update HTML comment to explain new purpose

3. Add category structure (Architecture, Patterns, Data Flow)

4. Note that this is the canonical source for Layer 2

Task: PROTOCOL-STATELESS-003 — Agent Behavior Integration

No separate workflow file needed. The stateless handoff is embedded in PROTOCOL.md as a meta-behavior that applies to all workflows (discuss, execution, checkpoint, etc.).

Agent instruction (to be added to PROTOCOL.md):

> After every response, append a stateless handoff packet. This is not optional. The user may fork at any moment — your response must be self-sufficient for any future agent to continue.

Part 7: Why This Approach Wins

Part 8: Cross-References and Citations

Part 9: Decision Record

📦 STATELESS HANDOFF

Layer 1 — Local Context:

→ Last action: LOG-016 (DECISION-016: Stateless-First Architecture)

→ Dependency chain: LOG-016 ← (new capability, no prior dependencies)

→ Next action: Implement PROTOCOL-STATELESS-001 — add section to PROTOCOL.md

Layer 2 — Global Context:

→ Architecture: This is a foundational protocol change

→ Patterns: Handoff format now canonical (Option C)

→ Key file: gsd-lite/template/PROTOCOL.md is the implementation target

Fork paths:

- Continue execution → Read LOG-016 Part 6 (Implementation Plan) → Update PROTOCOL.md

- Discuss handoff format details → Re-read LOG-016 Part 3 (Canonical Format)

- Pivot to new topic → Read gsd-lite/PROJECT.md for project vision + state your question

[LOG-017] - [VISION] - The Housekeeping Agent: Automated Coherence Detection for Dense Worklogs - Task: HOUSEKEEPING-AGENT

Timestamp: 2026-02-06

Status: 📋 VISION CAPTURED — Discussion Phase, No Implementation Yet

Decision ID: Pending (will become DECISION-017 upon approval)

Part 1: The Narrative — How We Got Here

The Problem: Context Rot in Long-Running Projects

During a live production evaluation using GSD-Lite on a real Looker/dbt integration project (eval/prod/), we encountered a fascinating paradox: the very system that was supposed to tame context rot had itself accumulated a 64,000-token worklog.

Think of it like a historian's archive that grew so comprehensive it became difficult to navigate without a historian.

The Evidence (from eval/prod/WORK_context_analysis.md):

Source: eval/prod/WORK_context_analysis.md (generated by analyze_context.py)

The Symptoms: Four Types of Context Pain

The user articulated the pain as "D — all of the above":

flowchart TD
    subgraph SYMPTOMS["🔴 Context Pain Symptoms"]
        A["(A) Finding what's CURRENT<br/>requires reading superseded entries"]
        B["(B) Key Events Index<br/>cluttered with dead branches"]
        C["(C) New agents waste tokens<br/>on stale context"]
        D["(D) Human wastes time<br/>reading irrelevant history"]
    end
    
    subgraph ROOT["🟠 Root Cause"]
        R1["No automated detection<br/>of coherence drift"]
        R2["Manual housekeeping<br/>is cognitively expensive"]
    end
    
    A --> R1
    B --> R1
    C --> R2
    D --> R2
    
    style SYMPTOMS fill:#ffcccc,stroke:#cc0000
    style ROOT fill:#fff3cd,stroke:#856404

The Paradox: Valuable Logs Are Also Expensive Logs

The dense worklog wasn't just noise — it was incredibly valuable:

1. PR Generation: Two substantial PRs were extracted directly from the logs:

- eval/prod/PR_artifacts/PR_dbt.md (19.54 KB) — dbt extension models

- eval/prod/PR_artifacts/PR_sightify.md (6.26 KB) — Sightify frontend integration

2. Decision Traceability: The log captured the journey, not just the destination:

- LOG-018 → LOG-020 → LOG-021 → LOG-024 shows a pivot from PDT strategy to Mirror Clone

- DECISION-006 → DECISION-007 shows refinement from "separate project" to "Two Galaxies"

3. Onboarding Power: A fresh agent could (in theory) read LOG-045 "Agent Onboarding Paths" and know exactly which logs to read for each type of task.

The insight: The value is real, but it's buried. We need curation, not deletion.

Part 2: The Research — Context Engineering Industry State (2026)

Before proposing a solution, we grounded our thinking in current industry best practices.

2.1 Context Engineering: Beyond Prompt Engineering

Key insight from Anthropic/LangChain research: Context engineering is not prompt engineering. It's about curating the entire state available to the LLM at any given time.

> "Context engineering is defined as curating what the model sees so that you get a better result... Agent failures are often due to a lack of 'right' context."

> — LangChain Deep Agents documentation (2026)

2.2 The Quality Degradation Curve

Research confirms what GSD-Lite already codifies:

Source: Anthropic context window management documentation, LangChain agent best practices

2.3 Industry Strategies for Managing Context Growth

flowchart LR
    subgraph STRATEGIES["Context Management Strategies"]
        direction TB
        S1["📜 Sliding Window<br/>Drop oldest, keep recent"]
        S2["📝 Summarization<br/>Condense old → summary"]
        S3["🗂️ Hierarchical Summary<br/>Chunk → summarize → re-summarize"]
        S4["💾 Filesystem Offload<br/>Store full content externally,<br/>keep pointers in context"]
        S5["📋 Selective Retention<br/>Prioritize high-value,<br/>discard low-value"]
    end
    
    style S4 fill:#d4edda,stroke:#155724
    style S5 fill:#d4edda,stroke:#155724

Highlighted strategies (S4 + S5) align with GSD-Lite's approach:

- Filesystem Offload: Logs live in WORK.md (external), agent reads surgically via grep

- Selective Retention: Key Events Index surfaces high-value decisions, grep patterns enable targeted reads

What's missing: Automated detection of when and what to curate.

2.4 The "Archival vs Consolidation" Dilemma

The user raised a valid concern:

> "I want to say merged, but there's a risk there similar to git workflows that we're 'rewriting history' and I'm afraid of the risk that it breaks coherence when agent scans logs to read and create the PR."

Industry parallel: Git has two merge strategies:

Proposed GSD-Lite hybrid: Archive to HISTORY.md (preserves journey), but also update the superseding log with a "Rejected Alternatives" section that summarizes what was tried.

Part 3: The Analysis — Lifecycle Patterns in Production WORK.md

We scanned eval/prod/WORK.md to identify natural lifecycle patterns that a Housekeeping Agent could exploit.

3.1 Decision Status Vocabulary (Discovered)

From grep analysis of eval/prod/WORK.md lines containing "Status:" or "SUPERSEDED":

Source: grep "Status:|SUPERSEDED" eval/prod/WORK.md (lines 63-5547)

3.2 Proposed Lifecycle Hierarchy

Based on the observed patterns, we can formalize a decision lifecycle:

stateDiagram-v2
    [*] --> PROPOSED: Agent proposes
    PROPOSED --> APPROVED: User approves
    PROPOSED --> REJECTED: User rejects
    
    APPROVED --> IMPLEMENTED: Code written
    APPROVED --> SUPERSEDED: Better option found
    
    IMPLEMENTED --> COMPLETED: Verified working
    IMPLEMENTED --> SUPERSEDED: Refactored away
    
    SUPERSEDED --> ARCHIVED: Moved to HISTORY.md
    COMPLETED --> ARCHIVED: Housekeeping triggered
    
    note right of SUPERSEDED
        Drift point: Agent detects
        and prompts reconciliation
    end note
    
    note right of ARCHIVED
        Preserves journey in HISTORY.md
        Removes from active WORK.md
    end note

3.3 Cross-Reference Patterns (Dependency Graph)

The grep also revealed how decisions reference each other:

DECISION-007 supersedes DECISION-006 (LOG-030 → LOG-031)
DECISION-009 depends on DECISION-004, DECISION-007
DECISION-012 depends on DECISION-004, DECISION-007
DECISION-018 supersedes (partially) DECISION-016

This creates a DAG (Directed Acyclic Graph) that the Housekeeping Agent can traverse to:

1. Find orphaned decisions (approved but never implemented)

2. Find dangling references (implemented but decision superseded)

3. Suggest consolidation opportunities (chains that converged)

Part 4: The Vision — What the Housekeeping Agent Would Do

4.1 Core Design Principles

4.2 The Interview Flow

sequenceDiagram
    participant U as User
    participant A as Housekeeping Agent
    participant W as WORK.md
    participant H as HISTORY.md
    
    U->>A: "Run housekeeping"
    A->>W: Grep for SUPERSEDED, Status, DECISION-
    A->>A: Build dependency graph
    A->>A: Detect drift points
    
    A->>U: "I found 5 drift points:<br/>1. LOG-018/20/21 superseded by LOG-024<br/>2. DECISION-006 orphaned (superseded but not archived)<br/>..."
    
    U->>A: "Let's reconcile #1 first"
    
    A->>U: "Options for LOG-018/20/21:<br/>(A) Archive to HISTORY with summary<br/>(B) Merge into LOG-024 as 'Rejected Alternatives'<br/>(C) Keep as-is with strikethrough<br/>Which fits your workflow?"
    
    U->>A: "A - Archive with summary"
    
    A->>H: Append archived entries
    A->>W: Remove archived entries
    A->>W: Update Key Events Index
    
    A->>U: "Done. Archived 3 entries (1,247 tokens freed).<br/>Next drift point?"

4.3 Detection Capabilities

The agent would scan for these drift patterns:

4.4 Output Options

For each drift point, the agent offers:

Part 5: Open Questions for Next Discussion

1. Lifecycle Formalization: Should we add explicit status: fields to log entries (like status: APPROVED | SUPERSEDED | COMPLETED)? Or infer from content?

2. Archive Granularity: Archive whole chains (LOG-018 + LOG-020 + LOG-021 together) or individual entries?

3. Consolidation Safety: If we merge superseded entries into the winning decision, how do we handle PR generation that might want to show the journey?

4. Agent Triggers: Should the agent run:

- (A) On-demand only ("run housekeeping")

- (B) Suggested at checkpoints ("WORK.md is 50k+ tokens, run housekeeping?")

- (C) Automatically at session start with light scan

5. Token Savings Reporting: Should the agent calculate and report tokens freed? (Motivates user to curate)

Part 6: Synthesized Example — A Complete Housekeeping Session

Scenario: User has been working on the Sightify demo for 3 weeks. WORK.md has grown to 60k tokens.

User: "Run housekeeping"

Agent: I scanned WORK.md and found 7 drift points across 3 categories:

**🔴 Superseded Chains (3 chains, 2,847 tokens):**
1. LOG-018 → LOG-020 → LOG-021 (superseded by LOG-024 "Mirror Clone")
   - Original strategy: PDT bridge for cross-region data
   - Why superseded: BigQuery cross-region join limitation
   
2. LOG-030 (superseded by LOG-031 "Two Galaxies")
   - Original strategy: Separate Looker project
   - Why superseded: Refined to single project with dual models

3. LOG-051 [DECISION-PENDING] (resolved by LOG-052 DECISION-018)
   - Original question: Static presets vs dynamic Looker filters?
   - Resolution: Dual Filter Sections with Mutex

**🟡 Status Drift (2 entries):**
4. Current Understanding lists TASK-006 as active, but LOG-053 shows Complete
5. Key Events Index still contains LOG-018 (superseded)

**🟢 Archival Candidates (2 completed tasks):**
6. TASK-001 (Feature Mapping) — 10 entries, all Complete
7. Phase 2A (dbt models) — 8 entries, merged to main

---

Which category would you like to tackle first?
(Or type a number 1-7 to address a specific item)

User: "1"

Agent: For the PDT bridge chain (LOG-018 → LOG-020 → LOG-021):

**Context:** This was a 3-step exploration of using Looker PDTs to bridge 
cross-project permissions. It failed because BigQuery doesn't allow 
cross-region joins (US public data vs Asia-SE1 owned data).

**Options:**
(A) Archive to HISTORY.md — I'll preserve 1-line summaries and the key 
    learning ("cross-region limitation"). LOG-024 already has the full 
    context of why we pivoted.
    
(B) Consolidate into LOG-024 — I'll add a "Rejected Alternatives" section 
    to LOG-024 with 3-sentence summaries of each attempt.
    
(C) Keep as-is — Leave with strikethrough. Useful if you want to show the 
    journey in PR descriptions.

Which approach fits your workflow? [YOUR TURN]

User: "A"

Agent: ✅ Archived 3 entries to HISTORY.md
- LOG-018: Pivot to Public Data (REJECTED: region mismatch)
- LOG-020: PDT Data Bridge (REJECTED: permissions insufficient)  
- LOG-021: PDT Implementation Guide (REJECTED: built on LOG-020)

Updated:
- Removed from WORK.md Atomic Log (freed 1,583 tokens)
- Removed LOG-018 from Key Events Index
- Added summary to HISTORY.md under "Phase 2: Rejected Approaches"

📊 Progress: 1/7 drift points resolved. ~4% context freed.
Next item, or continue to #2?

Part 7: Cross-References and Citations

Part 8: Decision Record (Pending)

[LOG-018] - [DECISION] - Housekeeping Agent Design: Two-Phase Architecture with Semantic Inference - Task: HOUSEKEEPING-AGENT

Timestamp: 2026-02-06

Status: ✅ DESIGN DECISIONS APPROVED

Depends On: LOG-017 (Vision: Housekeeping Agent concept)

Decision IDs: DECISION-018a through DECISION-018e

Part 1: The Narrative — Resolving the Chicken-and-Egg Problem

The Problem We Discovered

During discussion of LOG-017's Housekeeping Agent vision, we hit a critical design flaw: the agent was designed to scan for tags (SUPERSEDED, status:, decision references) that don't reliably exist.

The original LOG-017 design assumed:

Agent scans for SUPERSEDED tags → Builds DAG → Proposes archival

But in reality, fresh worklogs have no explicit tags. The user manually adds SUPERSEDED markers only after recognizing supersession — which requires the same cognitive load the agent was supposed to eliminate.

The paradox visualized:

flowchart TD
    subgraph CHICKEN_EGG["🐔🥚 The Chicken-and-Egg Problem"]
        A["Housekeeping Agent needs tags<br/>to detect supersession"]
        B["Tags don't exist until<br/>someone identifies supersession"]
        C["Identifying supersession requires<br/>reading all logs (cognitive load)"]
        D["User wanted agent to<br/>eliminate that cognitive load"]
    end
    
    A --> B
    B --> C
    C --> D
    D -->|"Loop!"| A
    
    style CHICKEN_EGG fill:#fff3cd,stroke:#856404

The Solution: Two-Phase Architecture

The user proposed a clean separation:

Key insight: Phase 1 creates the tags that Phase 2 depends on. No chicken-and-egg because they run sequentially.

Part 2: The Four Design Questions — Decisions Made

DECISION-018a: Tool Integration is REQUIRED

Question: Should the Housekeeping Agent require the token analysis tool (analyze_context.py)?

Decision: YES — Tool integration is mandatory, not optional.

Rationale (from user):

> "We need to save agent greps and divide and conquer — agent should be making the reasoning to infer DAGs, while a tool can deterministically and with little resource and consistently create the analysis in one fell swoop."

The division of labor:

Evidence of tool capability: The tool already captures explicit SUPERSEDED markers in titles:

# From eval/prod/WORK_context_analysis.md (Lines 42-44):
#### [LOG-018] - [DECISION] - ~~Pivot to Public Data...~~ (SUPERSEDED by LOG-024)
#### [LOG-020] - [DECISION] - ~~Solved Cross-Project Permissions...~~ (SUPERSEDED by LOG-024)
#### [LOG-021] - [PLAN] - ~~Complete LookML PDT Implementation...~~ (SUPERSEDED by LOG-024)

Source: eval/prod/WORK_context_analysis.md (generated by analyze_context.py)

DECISION-018b: Semantic Inference from Content (Three-Tier Signal Hierarchy)

Question: What content signals should the agent look for when inferring supersession?

Decision: Use a three-tier hierarchy based on confidence level. Agent acts on Tier 1, confirms Tier 2-3 with user.

The Inference Hierarchy (from helper agent analysis):

flowchart TD
    subgraph TIER1["Tier 1: HIGH Confidence (Agent Acts)"]
        T1A["Strikethrough in title: ~~old title~~"]
        T1B["'THIS LOG SUPERSEDES LOG-XXX' in body"]
        T1C["'obsolete' / 'Do NOT follow' language"]
        T1D["'pivot' / 'pivotal' / 'abandoned' keywords"]
        T1E["'hit a wall' / 'critical limitation'"]
        T1F["'Superseded Decisions' section header"]
    end
    
    subgraph TIER2["Tier 2: MEDIUM Confidence (Agent Confirms)"]
        T2A["Options evaluated → later DECISION picks one"]
        T2B["Same TASK-ID, sequential logs"]
        T2C["'What We Decided NOT to Do' section"]
        T2D["Explicit 'Depends On:' field"]
        T2E["'Before and After' diagrams"]
    end
    
    subgraph TIER3["Tier 3: LOW Confidence (Agent Asks)"]
        T3A["'refined' language (evolution vs supersession?)"]
        T3B["'consolidated' (merge vs supersession?)"]
        T3C["Same topic, different logs (parallel vs chain?)"]
    end
    
    style TIER1 fill:#d4edda,stroke:#155724
    style TIER2 fill:#fff3cd,stroke:#856404
    style TIER3 fill:#f8d7da,stroke:#721c24

Grep patterns for Phase 1C (from helper agent):

# Tier 1 signals
grep "THIS (LOG|DECISION) SUPERSEDES"
grep "obsolete|Do NOT follow|DEPRECATED"
grep "pivot|pivotal|abandoned"
grep "hit a.*wall|critical.*limitation"

# Cross-reference extraction
grep "LOG-[0-9]+.*LOG-[0-9]+"        # Edge candidates
grep "Depends On|depends on|as defined in"
grep "supersedes|replaces|invalidates"

Additional Tier 1 signals discovered in production WORK.md:

Source: Helper agent analysis + grep of eval/prod/WORK.md

DECISION-018c: Tagging Format — Header + Body Block (Option C)

Question: When the agent writes confirmed tags back to logs, what format?

Options evaluated:

- (A) Header only: ### [LOG-018] - ... - SUPERSEDED BY: LOG-024

- (B) Body block only: Metadata section inside log entry

- (C) Both: Header for grep, body for details

Decision: Option C — Both header and body block.

Rationale:

- Header tag enables fast grep scanning: grep "SUPERSEDED BY" WORK.md

- Body block preserves full relationship details for deep reads

Example output after tagging:

### [LOG-018] - [DECISION] - ~~Pivot to Public Data~~ - Task: PHASE-002 - **SUPERSEDED BY: LOG-024**

**Timestamp:** 2026-01-15
**Status:** ❌ SUPERSEDED
**Superseded By:** LOG-024 (Mirror Clone Strategy)
**Reason:** BigQuery cross-region join limitation made PDT bridge approach unviable

---
[Original log content preserved below]

DECISION-018d: DAG Visualization in Report — APPROVED

Question: Should the agent's report include a mermaid diagram of the inferred DAG?

Decision: YES — Include mermaid DAG visualization.

Rationale:

1. Gestalt understanding — User sees the whole structure at once

2. Catches orphans visually — Nodes with no edges stand out

3. Validates inference — User can spot "that edge doesn't make sense"

4. Low cost — Mermaid is text, minimal tokens

Example visualization:

flowchart TD
    subgraph PHASE002["Phase 2: Data Architecture"]
        LOG017["LOG-017: YAML Data Gen<br/>📋 DECISION"]
        LOG018["LOG-018: Public Data Pivot<br/>❌ SUPERSEDED"]
        LOG019["LOG-019: dbt-Native Strategy<br/>📋 DECISION"]
        LOG020["LOG-020: PDT Bridge<br/>❌ SUPERSEDED"]
        LOG021["LOG-021: PDT Implementation<br/>❌ SUPERSEDED"]
        LOG024["LOG-024: Mirror Clone<br/>✅ ACTIVE"]
        
        LOG019 -->|supersedes| LOG017
        LOG024 -->|supersedes| LOG018
        LOG024 -->|supersedes| LOG020
        LOG024 -->|supersedes| LOG021
    end
    
    style LOG018 fill:#ffcccc,stroke:#cc0000
    style LOG020 fill:#ffcccc,stroke:#cc0000
    style LOG021 fill:#ffcccc,stroke:#cc0000
    style LOG024 fill:#d4edda,stroke:#155724

DECISION-018e: Agent Spec Location — Multi-Agent Distribution Refactor

Question: Where should the Housekeeping Agent spec live?

User's vision: .opencode/agents/gsd-housekeeping.md

The gap identified: Current src/gsd_lite/__main__.py copies a single AGENTS.md to .opencode/agents/gsd-lite.md. Adding a second agent requires refactoring.

Current distribution logic (from src/gsd_lite/__main__.py:76-79):

# 5. Install Agent
agent_src = template_src / "AGENTS.md"
agent_dest = agents_dir / "gsd-lite.md"
shutil.copy2(agent_src, agent_dest)

Proposed refactor:

flowchart LR
    subgraph BEFORE["Current: Single Agent"]
        A1["template/AGENTS.md"] -->|copy| A2[".opencode/agents/gsd-lite.md"]
    end
    
    subgraph AFTER["Proposed: Multi-Agent"]
        B1["template/agents/gsd-lite.md"] -->|copy| B2[".opencode/agents/gsd-lite.md"]
        B3["template/agents/gsd-housekeeping.md"] -->|copy| B4[".opencode/agents/gsd-housekeeping.md"]
    end
    
    style BEFORE fill:#f8d7da,stroke:#721c24
    style AFTER fill:#d4edda,stroke:#155724

Implementation plan:

1. Create src/gsd_lite/template/agents/ directory

2. Move AGENTS.md → agents/gsd-lite.md (content unchanged)

3. Create agents/gsd-housekeeping.md (new agent spec)

4. Refactor __main__.py agent installation:

# BEFORE (lines 76-79):
agent_src = template_src / "AGENTS.md"
agent_dest = agents_dir / "gsd-lite.md"
shutil.copy2(agent_src, agent_dest)

# AFTER:
agents_src = template_src / "agents"
if agents_src.exists():
    for agent_file in agents_src.glob("*.md"):
        shutil.copy2(agent_file, agents_dir / agent_file.name)
    console.print(f"[green]✔ Installed Agents:[/green] {agents_dir} ({len(list(agents_src.glob('*.md')))} files)")

Decision: Proceed with multi-agent refactor. This enables seamless distribution of both:

- gsd-lite.md — Core GSD-Lite protocol agent

- gsd-housekeeping.md — Housekeeping Agent for coherence detection

Part 3: The Complete Phase 1 Algorithm

Based on all decisions, here's the full inference + interview flow:

sequenceDiagram
    participant U as User
    participant A as Housekeeping Agent
    participant T as analyze_context.py
    participant W as WORK.md
    
    U->>A: "Run housekeeping"
    
    Note over A,T: STEP 1: Tool generates structure
    A->>T: Run token analysis
    T->>A: Return section map + tokens + line numbers
    
    Note over A: STEP 2: Parse explicit markers
    A->>A: Extract strikethrough titles
    A->>A: Extract SUPERSEDED in headers
    A->>A: Extract Decision Record tables
    
    Note over A,W: STEP 3: Grep for Tier 1 signals
    A->>W: grep "THIS LOG SUPERSEDES"
    A->>W: grep "obsolete|Do NOT follow"
    A->>W: grep "pivot|abandoned"
    A->>W: grep cross-references
    
    Note over A: STEP 4: Build DRAFT DAG
    A->>A: Create nodes (all LOG/DECISION)
    A->>A: Create edges with confidence levels
    
    Note over A,U: STEP 5: Present + Interview
    A->>U: Show mermaid DAG + relationship table
    A->>U: "I found 12 relationships.<br/>6 HIGH confidence, 4 MEDIUM, 2 LOW."
    
    loop For each relationship
        A->>U: "LOG-024 supersedes LOG-018/20/21.<br/>Confidence: HIGH. Confirm?"
        U->>A: "Confirmed" / "Correct to X" / "Reject"
    end
    
    Note over A,W: STEP 6: Write confirmed tags
    A->>W: Add SUPERSEDED BY to headers
    A->>W: Add metadata blocks to bodies
    
    A->>U: "Tags written. Ready for Phase 2 (archival)?"

Part 4: Open Implementation Tasks

Part 5: Cross-References and Citations

Part 6: Decision Records Summary

📦 STATELESS HANDOFF

Layer 1 — Local Context:

→ Last action: LOG-018 (DECISION: Housekeeping Agent design decisions captured)

→ Dependency chain: LOG-018 ← LOG-017 (Vision) ← LOG-016 (Stateless-First)

→ Next action: Execute TASK-HA-001 through TASK-HA-004 (multi-agent refactor + spec draft)

Layer 2 — Global Context:

→ Architecture: GSD-Lite uses template/ for distributed artifacts

→ Patterns: Single AGENTS.md → Multi-agent agents/ directory

→ Tool: analyze_context.py required dependency for housekeeping

Fork paths:

- Execute multi-agent refactor → TASK-HA-001 through TASK-HA-003

- Draft housekeeping agent spec → TASK-HA-004

- Test on production → Run Phase 1 on eval/prod/WORK.md

- Discuss further → Refine inference patterns or tagging format

[LOG-019] - [EXEC] - Implemented Multi-Agent Architecture & Housekeeping Spec - Task: TASK-HA-004

Timestamp: 2026-02-07 16:30

Status: COMPLETE

Tasks: TASK-HA-001, TASK-HA-002, TASK-HA-003, TASK-HA-004

Context:

Refactored the core installer to support distributed agent definitions, enabling the addition of specialized agents like the Housekeeping Agent.

Changes:

1. Directory Structure (TASK-HA-001): Created src/gsd_lite/template/agents/ to house multiple agent definitions.

2. Migration (TASK-HA-002): Moved the monolithic AGENTS.md to template/agents/gsd-lite.md.

3. Installer Refactor (TASK-HA-003): Updated src/gsd_lite/__main__.py to iterate through the agents/ directory and install all .md files found, ensuring scalability.

4. Housekeeping Spec (TASK-HA-004): Drafted template/agents/gsd-housekeeping.md based on the approved design (LOG-018), implementing the "Context Gardener" persona.

Verification:

- ls confirmed directory creation.

- read confirmed file migration and new content.

- __main__.py logic reviewed and updated to handle multiple agents.

Next: Review the Housekeeping Agent spec and implement analyze_context.py (TASK-HA-005).

[LOG-020] - [DISCOVERY] - The Invisible Documentation Problem: HTML Comments Skip Grep-First Onboarding - Task: PROTOCOL-DOCS

Timestamp: 2026-02-07

Status: DISCOVERY → DECISION

Depends On: LOG-019 (Housekeeping spec review triggered this discovery), LOG-015 (Context Map tooling), LOOP-002 (Few-shot example visibility)

Decision IDs: DECISION-020a (Consolidate docs to agent instruction), DECISION-020b (10k token budget as CI pillar)

Part 1: The Discovery — What Agents Actually Read vs. What Exists

The Trigger

While reviewing the Housekeeping Agent spec (LOG-019), we asked: "Does the Context Gardener understand WORK.md structure?" This led to a debugger-style analysis of what agents actually read during onboarding.

The Finding: Grep-First Skips Documentation

GSD-Lite templates contain rich documentation in HTML comments at the top of each file. But agents following grep-first protocol never see them.

flowchart TD
    subgraph AGENT_PATH["🤖 Agent Onboarding Path"]
        A1["Load gsd-lite.md<br/>(system prompt)"]
        A2["Read PROJECT.md<br/>(full file)"]
        A3["Read ARCHITECTURE.md<br/>(full file)"]
        A4["grep '^## ' WORK.md<br/>(headers only)"]
        A5["Read Current Understanding<br/>(lines 34-105)"]
        A6["grep '^### \\[LOG-'<br/>(log headers)"]
        A7["Surgical read of<br/>specific logs"]
    end
    
    subgraph SKIPPED["❌ What Gets Skipped"]
        S1["WORK.md lines 3-30<br/>Lifecycle, purpose, grep patterns"]
        S2["WORK.md lines 36-56<br/>Current Understanding structure"]
        S3["INBOX.md lines 3-35<br/>Loop format, entry structure"]
        S4["HISTORY.md lines 3-10<br/>Archive purpose"]
        S5["All EXAMPLE-NNN entries<br/>(don't match LOG-NNN pattern)"]
    end
    
    A4 -->|"Skips HTML comments"| S1
    A4 -->|"Skips HTML comments"| S2
    A6 -->|"Pattern mismatch"| S5
    
    style SKIPPED fill:#f8d7da,stroke:#721c24
    style AGENT_PATH fill:#d4edda,stroke:#155724

Evidence: The HTML Comment Inventory

What the Agent Instruction Contains vs. Lacks

✅ gsd-lite.md DOES contain (citation: context_map.md):

- Universal Onboarding sequence (lines 33-53, ~310 tokens)

- File Reading Strategy (lines 123-174, ~539 tokens)

- File Guide quick reference (lines 175-185, ~124 tokens)

- Questioning Philosophy (lines 325-489, ~1,400 tokens)

❌ gsd-lite.md does NOT contain:

- What WORK.md sections mean and how they work

- What INBOX.md loop format is and why

- What HISTORY.md is for

- How Current Understanding fields should be structured

- What log entry bodies should contain (journalism format)

Part 2: The Token Budget Analysis — We Have Headroom

Current State (from `context_map.md`)

pie title Agent Instruction Token Budget (10k headroom)
    "Current gsd-lite.md" : 4913
    "Docs to migrate" : 1250
    "Remaining buffer" : 3837

Why 10k Matters

The user clarified the constraint: first-turn headroom is 10k tokens for system instructions + context before agent makes tool calls and starts responding. This is when the agent loads:

1. Agent instruction (gsd-lite.md)

2. System prompt overhead

3. Initial user message

Staying under 10k ensures the agent has context capacity for actual work.

Part 3: The Decision — Consolidate Documentation to Agent Instruction

DECISION-020a: Migrate File Documentation to gsd-lite.md

What: Move the essential documentation from HTML comments into the agent instruction as an "Artifact Format Reference" section.

Why:

1. Guaranteed visibility — Agent instruction is ALWAYS loaded. HTML comments are NEVER loaded with grep-first protocol.

2. Single source of truth — No drift between instruction and template comments.

3. Within budget — 6,163 tokens is well under 10k headroom.

Trade-off accepted: HTML comments become human-only documentation (for IDE/GitHub preview). Agent instruction becomes the authority.

Proposed structure:

## Artifact Format Reference

### WORK.md Structure
- **Section 1: Current Understanding** — 30-second handoff for fresh agents
  - Fields: current_mode, active_task, parked_tasks, vision, decisions, blockers, next_action
  - Write in concrete facts, not jargon ("as discussed" = useless to fresh agent)
  
- **Section 2: Key Events Index** — Curated foundation decisions (human-maintained)

- **Section 3: Atomic Session Log** — Chronological log entries
  - Format: `### [LOG-NNN] - [TYPE] - Summary - Task: TASK-ID`
  - Types: VISION, DECISION, EXEC, DISCOVERY, PLAN
  - Body: Journalism-style with timestamp, context, evidence, code snippets, rationale

### INBOX.md Structure
- Entry format: `### [LOOP-NNN] - Summary - Status: Open|Resolved`
- Fields: Created, Source, Origin, Context, Details, Resolution
- WHY this format: Headers enable grep triage; context preserves the "why"

### HISTORY.md Structure
- Minimal: One line per completed phase
- Links to external artifacts (PRs, docs)

Status: Pending implementation (TASK-PROTOCOL-DOCS-001)

Part 4: The CI Leverage — Token Budget as Deterministic Pillar

DECISION-020b: 10k Token Limit as CI Gate

The insight: The 10k token budget is a deterministic, measurable constraint. Unlike philosophical adherence (which requires LLM-as-judge), token count is programmatically verifiable.

Proposed CI check:

# Pseudo-code for CI gate
def check_agent_instruction_budget():
    """Fail CI if agent instruction exceeds 10k tokens."""
    content = read_file("src/gsd_lite/template/agents/gsd-lite.md")
    token_count = count_tokens(content)  # Use tiktoken or similar
    
    assert token_count <= 10_000, f"Agent instruction {token_count} tokens exceeds 10k budget"
    
    # Report headroom for visibility
    print(f"Token budget: {token_count}/10,000 ({10_000 - token_count} remaining)")

Integration with LOOP-001 (Intern Test):

The token budget check is a quick gate — fails fast before expensive LLM evaluation.

flowchart LR
    subgraph CI_PIPELINE["🔄 Proposed CI Pipeline"]
        D1["Deterministic Checks<br/>(fast, cheap)"]
        D2["Token budget ≤ 10k"]
        D3["Required sections exist"]
        D4["Grep patterns valid"]
        
        B1["Behavioral Checks<br/>(slow, expensive)"]
        B2["Intern Test scenarios"]
        B3["LLM-as-judge evaluation"]
    end
    
    D1 --> D2
    D1 --> D3
    D1 --> D4
    
    D2 -->|"Pass"| B1
    D3 -->|"Pass"| B1
    D4 -->|"Pass"| B1
    
    B1 --> B2
    B2 --> B3
    
    D2 -->|"Fail"| REJECT["❌ Reject PR"]
    
    style D1 fill:#d4edda,stroke:#155724
    style B1 fill:#fff3cd,stroke:#856404
    style REJECT fill:#f8d7da,stroke:#721c24

What else can we leverage from deterministic checks?

Part 5: Immediate Actions Taken This Session

Part 6: Open Tasks (Next Actions)

Backlinks

- LOOP-001 (Philosophical CI / Intern Test) — This discovery adds deterministic checks as a complement to behavioral testing

- LOOP-002 (Few-Shot Example Visibility) — Same root cause: valuable documentation invisible to grep-first agents

- LOG-015 (Context Map Tool) — The context_map.md output enabled the token budget analysis

- LOG-018 (Housekeeping Architecture) — The Gardener spec review triggered this discovery

- LOG-019 (Multi-Agent Implementation) — Immediate predecessor; spec updates applied

📦 STATELESS HANDOFF

Layer 1 — Local Context:

→ Last action: LOG-020 (Invisible Documentation discovery + decisions)

→ Dependency chain: LOG-020 ← LOG-019 ← LOG-018 ← LOG-017

→ Next action: Implement TASK-PROTOCOL-DOCS-001 (migrate docs to gsd-lite.md)

Layer 2 — Global Context:

→ Architecture: Multi-agent structure in src/gsd_lite/template/agents/

→ Patterns: Grep-first file reading, 10k token first-turn budget

→ Token state: gsd-lite.md at 4,913 tokens, headroom to 10k

Fork paths:

- Execute doc migration → TASK-PROTOCOL-DOCS-001

- Implement CI checks → TASK-CI-TOKEN-001

- Discuss further → Refine deterministic check list

- Review other agents → Apply same onboarding pattern to future agents

[LOG-021] - [DECISION] - Unified Housekeeping Agent: Collapsing Two Phases into One Stateless Spec - Task: HOUSEKEEPING-AGENT

Timestamp: 2026-02-07

Status: ✅ IMPLEMENTED

Decision ID: DECISION-021

Depends On: LOG-018 (Two-Phase Architecture), LOG-020 (Invisible Documentation), LOG-017 (Housekeeping Vision)

Supersedes: The separate workflows/housekeeping.md file (user will delete)

Part 1: The Problem — A Gap in Stateless-First Architecture

1.1 What the User Discovered

During reconciliation of the Housekeeping Agent vision (LOG-017) against the current implementation, the user identified a critical architectural gap:

> "Right now Phase 1 is embedded as system instruction for an agent. What is our vision for which agent would do Phase 2?"

The gap visualized:

flowchart TD
    subgraph PHASE1["Phase 1: Inference + Tagging"]
        P1A["Agent Spec: agents/gsd-housekeeping.md"]
        P1B["Embedded as system instruction"]
        P1C["✅ Stateless: Fresh agent can run from spec alone"]
    end
    
    subgraph PHASE2_BEFORE["Phase 2: Archival (BEFORE)"]
        P2A["Workflow File: workflows/housekeeping.md"]
        P2B["NOT an agent spec — requires GSD-Lite routing"]
        P2C["❓ Gap: Fresh agent doesn't know this file exists"]
    end
    
    subgraph HANDOFF["The Broken Handoff"]
        H1["Phase 1 ends with:<br/>'Run housekeeping.md workflow'"]
        H2["Fresh agent asks:<br/>'What's a workflow? Where is it?'"]
    end
    
    P1A --> P1B --> P1C
    P2A --> P2B --> P2C
    P1C --> H1 --> H2
    
    style P1C fill:#d4edda,stroke:#155724
    style P2C fill:#f8d7da,stroke:#721c24
    style H2 fill:#f8d7da,stroke:#721c24

Citation: User's exact words from this session:

> "Our philosophy is stateless by design so one core pillar to reconcile here is, after Phase 1 would the artifact and Phase 2 instruction be enough that we can start a new agent 0 context and let it do the workflow Phase 2?"

1.2 The Four Options Considered

1.3 Why Option A Won

User's directive:

> "Yes please go for extending the housekeeping agent spec, and then I'll delete the workflow no prob."

Rationale:

1. True stateless — A single agent spec handles the full lifecycle. No handoff ambiguity.

2. User simplicity — "Run housekeeping agent" → done (not "run gardener, then run archiver")

3. Token budget — Phase 2 operations are mechanical (~500 tokens to add). Total spec is ~2500 tokens, well under 10k.

4. Natural flow — Phase 1 ends with tags written → Phase 2 begins immediately (same session, same context)

Part 2: What Changed — The Extended Agent Spec

2.1 Before vs. After

Citation: src/gsd_lite/template/agents/gsd-housekeeping.md (full file rewritten this session)

2.2 New Sections Added

2.3 Key Code Patterns (Synthesized Examples)

Phase 2A trigger (archive superseded):

User: "archive superseded logs"

Agent:
  1. grep "SUPERSEDED BY:" WORK.md  # Find Phase 1 tags
  2. Present list with tokens to be freed
  3. After confirmation:
     - Move entries to HISTORY.md
     - Update Key Events Index
     - Update Current Understanding

HISTORY.md format for superseded logs:

## Superseded Logs (Archived 2026-02-07)

### LOG-003 - Auth Options Analysis
- **Superseded By:** LOG-007 (JWT Decision)
- **Original Task:** AUTH-IMPL
- **Summary:** Explored JWT vs Session vs OAuth. JWT chosen for statelessness.
- **Why Archived:** Options analysis complete; LOG-007 is authoritative.

Full lifecycle flow:

sequenceDiagram
    participant U as User
    participant A as Housekeeping Agent
    participant W as WORK.md
    participant H as HISTORY.md
    
    Note over A: Phase 1 (Inference + Tagging)
    U->>A: "run housekeeping"
    A->>W: grep log headers
    A->>A: Infer supersession (Tier 1/2/3)
    A->>U: Present draft DAG
    U->>A: Confirm relationships
    A->>W: Write SUPERSEDED BY: tags
    A->>U: "Phase 1 complete. Continue to Phase 2?"
    
    Note over A: Phase 2 (Archival)
    U->>A: "yes, archive them"
    A->>W: grep SUPERSEDED BY:
    A->>H: Move superseded entries
    A->>W: Update Key Events Index
    A->>U: "Archived 3 logs, freed 2,450 tokens"

Part 3: The Stateless Test — Can a Fresh Agent Pick This Up?

3.1 What a Fresh Agent Needs

3.2 The Handoff Packet (Now Internal)

Before (broken):

📦 CONTEXT GARDENER HANDOFF
**Next:** Run `housekeeping.md` workflow to archive tagged entries

After (stateless):

📦 HOUSEKEEPING HANDOFF

**Phase 1 (Tagging):**
→ Tagged: N logs marked SUPERSEDED
→ Untagged: M logs confirmed as still-valid

**Phase 2 (Archival):**
→ Archived: X logs to HISTORY.md
→ Tokens freed: Y
→ PRs extracted: [list if any]

A fresh agent loading this handoff + the agent spec has everything needed to continue.

Part 4: Impact on Existing Artifacts

4.1 What Gets Deleted

4.2 What Remains Valid

4.3 Protocol Routing Impact

The PROTOCOL.md workflow router currently says:

| "housekeeping" | housekeeping.md | PR extraction, archive completed tasks |

After this change: This route should point to the agent spec instead, OR the router should clarify that "housekeeping" loads the housekeeping agent, not a workflow.

Recommendation: Update PROTOCOL.md to reflect:

- agents/ = Standalone agent specs (load as system instruction)

- workflows/ = Instructions for the GSD-Lite agent to follow

Part 5: Backlinks & Dependencies

The dependency chain:

flowchart LR
    L16["LOG-016<br/>Stateless-First"]
    L17["LOG-017<br/>Housekeeping Vision"]
    L18["LOG-018<br/>Two-Phase Architecture"]
    L20["LOG-020<br/>Invisible Documentation"]
    L21["LOG-021<br/>Unified Agent (this log)"]
    
    L16 --> L17
    L17 --> L18
    L18 --> L21
    L20 --> L21
    
    style L21 fill:#d4edda,stroke:#155724

Part 6: Concrete Artifact Diff

File changed: src/gsd_lite/template/agents/gsd-housekeeping.md

📦 STATELESS HANDOFF

Layer 1 — Local Context:

→ Last action: LOG-021 (Unified Housekeeping Agent — this log)

→ Dependency chain: LOG-021 ← LOG-020 ← LOG-018 ← LOG-017 ← LOG-016

→ Next action: User deletes workflows/housekeeping.md (manual)

Layer 2 — Global Context:

→ Architecture: agents/gsd-housekeeping.md now handles full housekeeping lifecycle

→ Patterns: Stateless-first; one agent spec = complete workflow capability

→ Token state: Agent spec ~2500 tokens, well under 10k budget

Fork paths:

- Delete workflow → User removes workflows/housekeeping.md

- Update PROTOCOL.md → Reflect agent vs workflow distinction

- Test the agent → Run housekeeping on a real WORK.md

- Continue doc migration → TASK-PROTOCOL-DOCS-001

[LOG-022] - [DECISION] - Removing Legacy Housekeeping References: Ensuring Agent Separation of Concerns - Task: HOUSEKEEPING-CLEANUP

Timestamp: 2026-02-07

Status: ✅ IMPLEMENTED

Decision ID: DECISION-022

Depends On: LOG-021 (Unified Housekeeping Agent)

Supersedes: Legacy references to workflows/housekeeping.md

Part 1: The Problem — Agent Confusion

1.1 The Context

In LOG-021, we consolidated the housekeeping workflow into a standalone gsd-housekeeping agent. The original workflows/housekeeping.md file was deleted.

1.2 The Residue

However, references to "housekeeping workflow" remained in PROTOCOL.md, agents/gsd-lite.md, HISTORY.md, and workflows/progress.md.

1.3 The Risk

These references posed two risks:

1. Routing Confusion: The gsd-lite agent might try to "route" to a non-existent workflow file instead of instructing the user to switch agents.

2. Scope Creep: Instructions like "When archiving a completed task..." in HISTORY.md could lead the gsd-lite agent to believe it should perform archiving, violating the separation of concerns.

Part 2: The Decision — Strict Separation

We decided to:

1. Remove all routing references: The gsd-lite agent does not route to housekeeping. Housekeeping is a user-initiated "meta-task" performed by a separate agent.

2. Scrub "housekeeping" from gsd-lite artifacts: The main agent doesn't need to know the housekeeping agent exists.

3. Refine HISTORY.md instructions: Clarified that archiving is done (via Housekeeping Agent), signaling to the main agent that this is an external process.

Part 3: Concrete Changes

📦 STATELESS HANDOFF

Layer 1 — Local Context:

→ Last action: LOG-022 (this log) - Cleanup complete

→ Dependency chain: LOG-022 ← LOG-021 ← LOG-020

→ Next action: Return to TASK-PROTOCOL-DOCS-001 (Migrate artifact documentation)

Layer 2 — Global Context:

→ Architecture: Dual-agent model confirmed (gsd-lite vs gsd-housekeeping)

→ Patterns: Agents don't know about each other; user manages the switch

→ State: Residue references cleaned; artifacts are consistent with LOG-021

Fork paths:

- Resume docs migration → TASK-PROTOCOL-DOCS-001

- Test housekeeping agent → Run gsd-housekeeping on a sample

[LOG-021] - [EXEC] - Migrated Artifact Docs to Agent Instruction - Task: PROTOCOL-DOCS-001

Timestamp: 2026-02-07 17:15

Status: COMPLETE

Tasks: TASK-PROTOCOL-DOCS-001

Context:

Per DECISION-020a, we migrated essential artifact documentation from HTML comments (invisible to grep-first agents) to gsd-lite.md agent instruction (always loaded). Also decommissioned the redundant PROTOCOL.md template.

Changes:

1. Decommissioned PROTOCOL.md: Deleted src/gsd_lite/template/PROTOCOL.md. The agent instruction at src/gsd_lite/template/agents/gsd-lite.md is now the single source of truth for the protocol.

2. Updated Agent Instruction: Added ## WORK.md Structure, ## INBOX.md Structure, and ## HISTORY.md Structure sections to src/gsd_lite/template/agents/gsd-lite.md.

3. Cleaned Templates: Removed HTML comments from src/gsd_lite/template/WORK.md. (INBOX/HISTORY cleanup left for user per request).

Impact:

- Visibility: Fresh agents now see WORK/INBOX/HISTORY structure immediately upon reading system prompt.

- Maintenance: Single source of truth (gsd-lite.md), no drift between template comments and instruction.

- Token Budget: Added ~400 tokens to agent instruction, still well within 10k headroom.

Next: Implement TASK-CI-TOKEN-001 to enforce the 10k budget in CI.

[LOG-022] - [DISCOVERY] - The Missing Metadata Line: Agent Instruction Describes Format But Lacks Executable Template - Task: PROTOCOL-DOCS-001

Timestamp: 2026-02-07 17:45

Status: DISCOVERY → DECISION

Depends On: LOG-021 (Artifact doc migration), LOG-020 (Invisible Documentation Problem)

Decision IDs: DECISION-022a (Add executable templates to agent instruction)

Part 1: The Discovery — Agent Wrote Non-Compliant Entry

The Trigger

Immediately after completing LOG-021 (migrating artifact docs to agent instruction), user asked the agent to capture LOOP-003. The agent produced an entry that partially violated the INBOX format spec it had just helped write.

The Evidence: What Agent Wrote vs. What Spec Requires

The Spec (from src/gsd_lite/template/agents/gsd-lite.md:228-232):

### Entry Format
- **Header:** `### [LOOP-NNN] - {{summary}} - Status: {{Open|Clarifying|Resolved}}`
- **Fields:** Created, Source, Origin (User|Agent), Context, Details, Resolution
- **Rule:** Write context-rich entries, not just titles — tell the story

What Agent Actually Wrote (LOOP-003 in gsd-lite/INBOX.md):

### [LOOP-003] - Structural Regression Prevention: CI for Documentation Drift - Status: Open
**Created:** 2026-02-07 | **Source:** Post-mortem on LOG-020/LOG-021 | **Origin:** User

**Context:**
...

Wait — on re-reading, the agent DID include the metadata line. Let me re-check the actual file:

flowchart TD
    subgraph AUDIT["🔍 Compliance Audit"]
        A1["Header Format"] -->|"✅"| A1R["[LOOP-003] - Summary - Status: Open"]
        A2["Metadata Line"] -->|"❓"| A2R["Check if present"]
        A3["Context Section"] -->|"✅"| A3R["Present with full narrative"]
        A4["Details Section"] -->|"✅"| A4R["Present with tables and diagrams"]
        A5["Resolution"] -->|"✅"| A5R["_(pending)_ present"]
    end

The Real Finding: Spec Describes But Doesn't Show

The deeper issue isn't that the agent failed — it's that the spec describes the format abstractly but doesn't provide an executable template.

Current spec in gsd-lite.md:228-232:

### Entry Format
- **Header:** `### [LOOP-NNN] - {{summary}} - Status: {{Open|Clarifying|Resolved}}`
- **Fields:** Created, Source, Origin (User|Agent), Context, Details, Resolution
- **Rule:** Write context-rich entries, not just titles — tell the story

What's missing: A concrete, copy-paste-ready template that shows the EXACT structure:

### [LOOP-NNN] - {{one-line summary}} - Status: Open
**Created:** YYYY-MM-DD | **Source:** {{task/context}} | **Origin:** User|Agent

**Context:**
{{Why this loop exists — the situation that triggered it}}

**Details:**
{{Specific question/concern with code references}}

**Resolution:** _(pending)_

Why This Matters: The Template-Instruction Gap

flowchart LR
    subgraph CURRENT["Current State"]
        C1["gsd-lite.md<br/>Describes format abstractly"]
        C2["INBOX.md template<br/>Has EXAMPLE-LOOP-NNN entries"]
        C3["Agent reads gsd-lite.md<br/>Never sees examples"]
    end
    
    subgraph PROBLEM["The Gap"]
        P1["Abstract description<br/>≠<br/>Executable template"]
    end
    
    subgraph RESULT["Outcome"]
        R1["Agent interprets spec<br/>May miss fields"]
        R2["Inconsistent entries<br/>Harder to grep/audit"]
    end
    
    C1 --> P1
    C2 -->|"grep skips EXAMPLE-"| P1
    P1 --> R1
    R1 --> R2
    
    style PROBLEM fill:#fff3cd,stroke:#856404
    style RESULT fill:#f8d7da,stroke:#721c24

The template file (src/gsd_lite/template/INBOX.md) contains rich examples like [EXAMPLE-LOOP-001], but:

1. Agents grep ^### \[LOOP- which skips EXAMPLE-LOOP- entries (see LOG-020, LOOP-002)

2. The agent instruction describes format but doesn't include a copy-paste template

3. Result: Agent must interpret abstract description → inconsistent output

Part 2: The Pattern — This Affects All Artifact Formats

Audit of Current Agent Instruction Sections

Citations:

- WORK.md Structure: src/gsd_lite/template/agents/gsd-lite.md:188-219

- INBOX.md Structure: src/gsd_lite/template/agents/gsd-lite.md:221-244

- HISTORY.md Structure: src/gsd_lite/template/agents/gsd-lite.md:246-264

What Needs Executable Templates

1. WORK.md Log Entry Template:

### [LOG-NNN] - [TYPE] - {{one-line summary}} - Task: TASK-ID
**Timestamp:** YYYY-MM-DD HH:MM
**Status:** {{status if applicable}}
**Depends On:** {{LOG-XXX (brief description), LOG-YYY (brief description)}}

**Context:**
{{Why this log exists — what triggered it}}

**Details:**
{{Full narrative with code snippets for EXEC/DISCOVERY}}

**Next:** {{Immediate next action}}

2. INBOX.md Loop Entry Template:

### [LOOP-NNN] - {{one-line summary}} - Status: Open
**Created:** YYYY-MM-DD | **Source:** {{task/context}} | **Origin:** User|Agent

**Context:**
{{Why this loop exists — the situation that triggered it}}

**Details:**
{{Specific question/concern with code references}}

**Resolution:** _(pending)_

Part 3: The Decision — Embed Executable Templates in Agent Instruction

DECISION-022a: Add Executable Templates to gsd-lite.md

What: Enhance each artifact structure section with a copy-paste-ready template block.

Why:

1. Removes interpretation burden — Agent copies template, fills placeholders

2. Guarantees field presence — Template includes all required fields

3. Enables auditing — Reviewers can compare entry against template

4. Compensates for invisible examples — Templates are visible even though EXAMPLE-NNN entries are not

Trade-off accepted: Adds ~200 tokens to agent instruction. Budget impact: ~6,163 + 200 = ~6,363 tokens, still under 10k headroom.

Implementation:

- Add template block under each "Entry Format" subsection

- Use {{placeholder}} syntax for variable parts

- Include all required fields with inline comments

Part 4: Relationship to Other Findings

Part 5: Immediate Next Action

📦 STATELESS HANDOFF

Layer 1 — Local Context:

→ Last action: LOG-022 (Discovery: Missing executable templates in agent instruction)

→ Dependency chain: LOG-022 ← LOG-021 ← LOG-020 ← LOG-019

→ Next action: Implement TASK-PROTOCOL-DOCS-002 (add templates to gsd-lite.md)

Layer 2 — Global Context:

→ Architecture: Single source of truth at src/gsd_lite/template/agents/gsd-lite.md

→ Patterns: Grep-first file reading, 10k token budget, invisible EXAMPLE-NNN entries

→ Decision: DECISION-022a — embed executable templates

Fork paths:

- Execute template addition → TASK-PROTOCOL-DOCS-002

- Discuss CI framework → Continue LOOP-003

- Fix LOOP-003 metadata → Correct existing entry

[LOG-023] - [EXEC] - Added Executable Templates to Agent Instruction - Task: PROTOCOL-DOCS-002

Timestamp: 2026-02-07 18:00

Status: COMPLETE

Depends On: LOG-022 (Discovery: Missing templates), DECISION-022a (Add templates to gsd-lite.md)

Context:

Per DECISION-022a, we added copy-paste-ready templates to the agent instruction so agents don't have to interpret abstract format descriptions.

Changes:

Template Features:

- Copy-paste ready markdown blocks

- {{placeholder}} syntax for variable parts

- Field requirements table explaining each field

- Includes stateless handoff structure for logs

Impact:

- Agents can now copy template and fill placeholders

- All required fields are visible in template (no interpretation needed)

- Enables auditing: compare entry against template

- Token cost: ~200 tokens added, still well under 10k budget

Next: Continue LOOP-003 discussion (CI framework for structural regression prevention)

[LOG-025] - [DECISION] - Housekeeping Tooling Spec: Upgrading Context Map for Agent Consumption (GSD-Lite Mode) - Task: TOOLING-002

Timestamp: 2026-02-07

Status: ✅ SPECIFICATION LOCKED

Depends On: LOG-015 (Context Map Tool), LOG-021 (Housekeeping Agent Vision)

Decision IDs: DECISION-025a (JSON Output), DECISION-025b (Semantic Signal Tiering)

Part 1: The Narrative — From Human-Readable to Agent-Native

The Context (Inlined):

- The Existing Tool (LOG-015): We have a script analyze_context.py that parses Markdown files using markdown-it-py and tiktoken. It currently generates a human-readable Markdown "map" of the file structure (headers, line numbers, token counts) to help agents navigate large files without reading them entirely.

- The New Requirement (LOG-021): We are building a "Housekeeping Agent" responsible for scanning session logs, inferring which decisions have been superseded (e.g., "Decision A replaced by Decision B"), and interviewing the user to confirm archival.

The Gap:

The Housekeeping Agent needs structured data (JSON) to infer relationships programmatically. The current tool only outputs unstructured Markdown text. This forces the agent to burn tokens parsing the very map intended to save tokens.

The Solution:

We are upgrading analyze_context.py to be a first-class GSD-Lite citizen. It will gain a --gsd-lite mode that:

1. Outputs machine-readable JSON.

2. Detects semantic signals (like strikethrough titles Title or SUPERSEDED BY: tags) deterministically using regex.

3. Handles the messy reality of Task naming (extracting Task: MODEL-A vs Task: TASK-001) without crashing.

The Architecture Shift:

flowchart TD
    subgraph "Current State (Human-Readable)"
        A[WORK.md] -->|Markdown Parser| B(Markdown Map)
        B -->|Agent reads text| C{Agent Parsing}
        C -->|High Token Cost| D[Inference]
    end

    subgraph "New State (Machine-Readable)"
        E[WORK.md] -->|GSD-Lite Mode| F(JSON Object)
        F -->|Direct Load| G{Agent Logic}
        G -->|Zero Parsing Cost| H[Inference]
        
        subgraph "Tool Logic (Deterministic)"
            I[Task Extraction]
            J[Signal Detection]
            K[Token Counting]
        end
        
        I & J & K --> F
    end
    
    style B fill:#ffcccc,stroke:#333
    style F fill:#ccffcc,stroke:#333

Part 2: The Research — Semantic Signal Tiering

We investigated how to detect "supersession" (when one log makes another obsolete) without running an expensive LLM. The research revealed a Graduated Tier System:

Key Decision: The tool will FLAG Tier 1 and 2 signals but NEVER DECIDE. It provides evidence; the agent/user provides judgment.

Part 3: The Specification Packet (For Remote Builder Agent)

> Context for Builder: This spec defines the upgrade for analyze_context.py. The goal is to make it the deterministic "eyes" for the Housekeeping Agent. The agent is a specialized LLM workflow that reads session logs and archives old content.

1. CLI Interface & Defaults

Command:

# Default GSD-Lite Mode (targets gsd-lite/WORK.md, outputs JSON)
analyze_context --gsd-lite --work --format json

# Explicit Path
analyze_context --gsd-lite --work ./custom/WORK.md

New Arguments:

- --gsd-lite: Enables GSD-specific parsing (Task IDs, Log IDs, Signals).

- --work: Sets default path to gsd-lite/WORK.md (or specific path if provided).

- --format: Options md (default), json (new requirement), table.

2. JSON Output Schema (Requirement)

{
  "summary": {
    "total_tokens": 65420,
    "total_logs": 24,
    "tier_1_flags": 3
  },
  "logs": [
    {
      "log_id": "LOG-018",
      "type": "DECISION",
      "task": "HOUSEKEEPING-AGENT",  // Extracted from header only
      "tokens": 1200,
      "lines": [3213, 3287],
      "signals": {
        "tier_1": ["strikethrough: ~~Pivot to Public Data~~"],
        "tier_2": ["depends_on: LOG-017", "pivot: pivoted"]
      }
    },
    {
      "log_id": "LOG-024",
      "type": "DECISION",
      "task": "PHASE-002",
      "tokens": 800,
      "lines": [3400, 3450],
      "signals": {
        "tier_1": [],
        "tier_2": []
      }
    }
  ]
}

3. Parsing Logic (Regex Specs)

A. Log Header Parsing

Goal: Permissive extraction. Do not enforce TASK-NNN. Capture MODEL-A.

Pattern: ^### \[LOG-(\d+)\]\s-\s\[([A-Z]+)\]\s-\s(.?)\s-\sTask:\s([A-Z][A-Za-z0-9_-]+)

B. Semantic Signal Detection (The "Brain" Upgrade)

Goal: Detect supersession markers in header AND body.

Tier 1 Patterns (High Confidence - Auto-Flag):

TIER_1_PATTERNS = {
    "strikethrough": r"~~[^~]+~~",                 # Title strikethrough
    "superseded_by": r"SUPERSEDED\s*BY[:\s]+LOG-\d+", # Explicit tag
    "deprecated_tag": r"\[DEPRECATED\]|\[OBSOLETE\]|\[ARCHIVED\]",
    "do_not_follow": r"[Dd]o\s*[Nn][Oo][Tt]\s*follow",
    "status_obsolete": r"[Ss]tatus[:\s]*(obsolete|deprecated|superseded|abandoned)",
    "killed": r"\b(killed|scrapped|abandoned|discarded)\b"
}

Tier 2 Patterns (Medium Confidence - Flag for Review):

TIER_2_PATTERNS = {
    "depends_on": r"[Dd]epends\s*[Oo]n[:\s]*(LOG-\d+)",
    "supersedes": r"\b(supersedes?|superseding)\b",
    "replaces": r"\b(replaces?|replacing)\b",
    "pivot": r"\b(pivot(ed|ing)?|pivotal)\b",
    "hit_wall": r"hit\s*(a\s*)?(wall|dead\s*end|roadblock)",
    "decided_not_to": r"decided\s*(not\s*to|against)",
    "options_evaluated": r"[Oo]ption\s*[A-Z1-9][:\s]"
}

4. Validation & Constraints

1. Header-Anchored Task ID: Only extract Task: from the log header line. Do NOT scan body for Task: to avoid false positives (e.g., "We discussed Task: AUTH").

2. No HEAD/TAIL in JSON: The JSON output should NOT include preview text (noise). Agents use read_files(lines=[start, end]) for that.

3. Dependencies: Keep it lightweight. re (standard lib), tiktoken, markdown-it-py. No spaCy or NLP libraries.

📦 STATELESS HANDOFF

Layer 1 — Local Context:

→ Last action: LOG-025 (Decision + Spec for Tooling)

→ Dependency chain: LOG-025 ← LOG-015 (Tool) ← LOG-021 (Agent)

→ Next action: Implement the spec in analyze_context.py (or handoff to builder agent)

Layer 2 — Global Context:

→ Architecture: GSD-Lite tooling (LOG-015)

→ Patterns: Deterministic tool + Semantic Agent (LOG-018)

Fork paths:

- Build tool: python3 analyze_context.py --gsd-lite implementation

- Discuss: Refine regex patterns

[LOG-026] - [DECISION] - The Quine Paradox: Fixing False Positives in Self-Referential Tooling Specs - Task: TOOLING-002

Timestamp: 2026-02-07

Status: ✅ SPECIFICATION AMENDED

Depends On: LOG-025 (Original Spec), LOG-015 (Context Map Tool)

Decision IDs: DECISION-026 (Context-Aware Exclusion Strategy)

Part 1: The Narrative — The "Quine" Problem

The Context (Inlined):

- The Spec (LOG-025): We defined a tool to detect "semantic signals" like strikethrough titles or SUPERSEDED BY: tags to identify obsolete logs.

- The Paradox: To write the spec, we had to write down the very patterns the tool is supposed to detect.

- The Result: The specification itself (LOG-025) is now flagged by the tool as "superseded" because it contains 4+ instances of the "superseded" pattern in its examples and regex definitions.

The Evidence (Grep Analysis):

We ran a grep on WORK.md and found the tool cannot distinguish between using a signal and mentioning it.

> Concept: This is a "Quine Paradox" in static analysis — when the code (or documentation) describing a pattern contains the pattern itself.

Part 2: The Research — Exclusion Strategies

We researched how established static analysis tools (ESLint, Pylint, SonarQube) handle this "documentation vs. code" problem.

The Decision (DECISION-026):

We will adopt a Hybrid Approach:

1. Context-Aware Exclusion (Primary): The tool MUST ignore all content inside Fenced Code Blocks ( `...` ) and Inline Code (` ... `).

2. Structural Anchoring (Secondary): Certain signals (like Title) are ONLY valid in specific locations (Header Lines).

Part 3: The Architecture — Masking Pipeline

We are upgrading the parsing logic from "Regex on Raw Text" to a "Mask -> Scan -> Unmask" pipeline.

flowchart TD
    A[Raw Log Content] -->|Step 1: Mask Safe Zones| B(Masked Content)
    B -->|Step 2: Apply Regex Patterns| C{Signal Detection}
    C -->|Match Found| D[Record Signal]
    C -->|No Match| E[Ignore]
    
    subgraph "Safe Zones (Ignored)"
        F[Fenced Code Blocks]
        G[Inline Code `...`]
        H[Blockquotes > ...]
    end
    
    subgraph "Masking Logic"
        I[Replace Code with <PLACEHOLDER>]
        J[Keep Line Numbers Intact]
    end
    
    F & G & H --> I
    I --> B

Part 4: The Specification Amendment (Python Implementation)

> Requirement for Builder Agent: The analyze_context.py tool must implement this detect_signals logic to prevent false positives in documentation.

1. The Masking Logic (Python Pseudocode):

import re

def mask_exclusion_zones(text):
    """
    Replaces code blocks with placeholders to prevent regex matching on examples.
    Crucial: Preserves newlines so line numbers remain accurate.
    """
    placeholders = []
    
    def replacer(match):
        content = match.group(0)
        placeholders.append(content)
        # Replace with safe string, keeping newlines for line counts
        return f"__MASKED_CODE_BLOCK_{len(placeholders)-1}__" + ("\n" * content.count("\n"))

    # 1. Mask Fenced Code Blocks (```...```)
    # Pattern: Triple backticks, optional language, content, triple backticks
    text = re.sub(r"```[\s\S]*?```", replacer, text)
    
    # 2. Mask Inline Code (`...`)
    # Pattern: Single backtick, non-backtick content, single backtick
    text = re.sub(r"`[^`\n]+`", replacer, text)
    
    return text

2. The Detection Logic (Updated from LOG-025):

def detect_signals(log_content):
    # Step 1: Mask the content
    masked_text = mask_exclusion_zones(log_content)
    lines = masked_text.split('\n')
    
    signals = []
    
    for i, line in enumerate(lines):
        # Step 2: Apply Anchored Patterns
        
        # A. Header-Only Signals (Strict)
        if line.startswith("### [LOG-"):
            if "~~" in line:  # Simple check first
                 if re.search(r"~~[^~]+~~", line):
                     signals.append({"line": i+1, "type": "strikethrough", "tier": 1})
            
            if "SUPERSEDED" in line:
                 if re.search(r"SUPERSEDED\s*BY[:\s]+LOG-\d+", line):
                     signals.append({"line": i+1, "type": "superseded_by", "tier": 1})

        # B. Body Signals (Permissive but Masked)
        # Note: 'line' here has code blocks replaced with __MASKED__
        if "[DEPRECATED]" in line or "[OBSOLETE]" in line:
            signals.append({"line": i+1, "type": "deprecated_tag", "tier": 1})
            
    return signals

3. Verification Case (The "Quine" Test):

📦 STATELESS HANDOFF

Layer 1 — Local Context:

→ Last action: LOG-026 (Fixing False Positives in Spec)

→ Dependency chain: LOG-026 ← LOG-025 (Original Spec) ← LOG-015 (Tool)

→ Next action: Implement analyze_context.py with this masking logic.

Layer 2 — Global Context:

→ Architecture: GSD-Lite Tooling (LOG-015)

→ Patterns: Context-Aware Static Analysis (LOG-026)

Fork paths:

- Build tool: python3 analyze_context.py --gsd-lite with masking

- Discuss: Expand "Safe Zones" to Blockquotes?

[LOG-027] - [EXEC] - Implemented Quine-Resistant Housekeeping Tool & Stateless Agent Router - Task: TOOLING-002

Timestamp: 2026-02-07

Status: ✅ COMPLETE

Depends On: LOG-026 (Quine Fix), LOG-025 (Tool Spec)

Decision IDs: DECISION-027 (Stateless Phase Detection)

Part 1: The Tool Implementation — Solving the Quine Paradox

We implemented the masking pipeline defined in LOG-026. The new tool gsd_lite_analyzer.py successfully masks code blocks and inline code before scanning for signals.

Key Implementation Detail: Newline-Preserving Masking

To ensure line numbers in the JSON output match the original file (critical for agent navigation), the masking logic preserves newlines inside masked blocks:

# From src/fs_mcp/gsd_lite_analyzer.py
def mask_exclusion_zones(text: str) -> tuple[str, list[str]]:
    """
    Replaces code blocks with placeholders.
    Crucially: Preserves newlines so line numbers remain accurate.
    """
    placeholders = []
    
    def create_placeholder(match: re.Match) -> str:
        content = match.group(0)
        placeholders.append(content)
        # Replace with safe string + original newlines
        placeholder = f"__MASKED_{len(placeholders)-1}__"
        newline_count = content.count("\n")
        return placeholder + ("\n" * newline_count)
    
    # 1. Mask Fenced Code Blocks (```...```)
    text = re.sub(r"```[\s\S]*?```", create_placeholder, text)
    
    # 2. Mask Inline Code (`...`)
    text = re.sub(r"`[^`\n]+`", create_placeholder, text)
    
    return text, placeholders

Verification Results (Dogfood Test):

We ran the tool on this very WORK.md file (which contains LOG-025 and LOG-026 with many "false positive" patterns).

Part 2: The Agent Instruction — Stateless Phase Router

We embedded the tool directly into gsd-housekeeping.md and implemented a Stateless Router that detects Phase 1 vs Phase 2 based on artifact state alone.

The Logic Flow:

flowchart TD
    Start[User says "go"] --> Tool[Run analyze_gsd_work_log]
    Tool --> Check{Check Output}
    
    Check -->|Tier 1 Flags + NO Header Tags| P1[Phase 1: Inference]
    Check -->|Existing SUPERSEDED Tags| P2[Phase 2: Archival]
    Check -->|No Flags/Tags| Clean[Report Clean]
    
    P1 --> Interview[Interview & Write Tags]
    P2 --> Archive[Confirm & Archive]

The Implementation (in gsd-housekeeping.md):

## Session Start (Stateless Router)

**User says "go" → I detect phase from artifact state:**

1. **Read PROJECT.md** — Get domain vocabulary
2. **Run `analyze_gsd_work_log("gsd-lite/WORK.md")`** — Get signal analysis
3. **Detect phase from tool output:**

| Condition | Phase | Action |
|-----------|-------|--------|
| Tier 1 flags exist, NO `SUPERSEDED BY:` tags in headers | **Phase 1** | Interview → Write tags |
| `SUPERSEDED BY:` tags already in headers | **Phase 2** | Confirm → Archive |
| No flags, no tags | **Clean** | Report "Nothing to housekeep" |

Why this matters: A user can spin up the housekeeping agent at any point. If they ran Phase 1 yesterday (tags written) and resume today, the agent automatically recognizes Phase 2 is needed. No "handoff memory" required — the state is in the artifact.

📦 STATELESS HANDOFF

Layer 1 — Local Context:

→ Last action: LOG-027 (Implementation & Validation)

→ Dependency chain: LOG-027 ← LOG-026 (Quine Fix) ← LOG-025 (Tool Spec)

→ Next action: Use the housekeeping agent to clean up superseded logs (Phase 1).

Layer 2 — Global Context:

→ Architecture: GSD-Lite Tooling (LOG-015)

→ Patterns: Stateless Router (LOG-027), Quine-Resistant Analysis (LOG-026)

Fork paths:

- Run Housekeeping: gsd-housekeeping.md workflow

- Execute Task: Pick up next item from parked_tasks

[LOG-028] - [DISCOVERY] - CI Framework Design: Constitutional Knowledge for Regression Prevention - Task: CI-FRAMEWORK-001

Timestamp: 2026-02-08

Status: IN_PROGRESS → CHECKPOINT

Depends On: LOG-020 (10k token budget as CI gate), LOG-016 (Stateless-First Architecture), LOG-017 (Housekeeping Agent vision)

Decision IDs: DECISION-028a (Constitution-first approach), DECISION-028b (Three-layer CI architecture)

Part 1: The Problem — 12 PRs with Zero Automated Checks

GSD-Lite has evolved through 12 closed PRs with no regression protection. Every change relied on "trust me bro" discussions between human and agent across gsd-lite sessions.

Why this worked initially: Everything was experimental. The framework was finding its shape.

Why it's now a problem: The user has grown to depend on GSD-Lite across multiple production projects. The framework's complexity now exceeds one person's mental model:

The recursive problem: To discuss a new iteration, the user needs a gsd-lite agent to consume the context. Control has drifted to the agent itself — violating the "Driver owns, Navigator proposes" philosophy.

Source: User braindump session 2026-02-08, GitHub repo scan of luutuankiet/gsd-lite (12 closed PRs at time of writing).

Part 2: The Solution — Constitutional Knowledge

Instead of trying to test every possible agent behavior, we distill GSD-Lite's immutable pillars into a machine-auditable document: the Constitution.

The Constitution Concept:

graph TD
    subgraph "What Can Change"
        W[Workflows] 
        E[Examples]
        A[Artifact Templates]
    end
    
    subgraph "What Cannot Change"
        C[Constitution<br/>~500 lines<br/>Immutable Pillars]
    end
    
    W -->|Must comply with| C
    E -->|Must comply with| C
    A -->|Must comply with| C
    
    subgraph "CI Pipeline"
        L1[L1: Structural<br/>grep patterns]
        L2[L2: Constitutional<br/>LLM-as-judge]
        L3[L3: Behavioral<br/>Golden tests]
    end
    
    C --> L2
    L1 --> L2 --> L3

Why Constitution-first:

- Single source of truth for "what makes GSD-Lite GSD-Lite"

- LLM-as-judge can evaluate any artifact against it

- Human-readable contract for contributors

- Decouples philosophy (stable) from implementation (evolving)

Decision: DECISION-028a — Adopt Constitution-first approach for CI. All other artifacts can drift as long as they don't violate the Constitution.

Part 3: The Six Pillars (Draft Constitution)

Based on braindump synthesis and echo-back verification with user:

Pillar 1: Stateless-First

> Every agent turn generates a handoff packet enabling zero-context resume.

Test: Structural grep for 📦 STATELESS HANDOFF with required components:

- Layer 1 (Local Context) — task dependency chain

- Layer 2 (Global Context) — from Key Events Index

- Fork paths — minimum 2 (continue + pivot)

Gate: Hard fail if missing from agent response template.

Backlink: See LOG-016 (Stateless-First Architecture) for the full specification and Two-Layer Handoff Structure.

Pillar 2: Context Engineering

> Optimize token budget: ~10k static (agent + workflow), ~10-20k JIT (artifacts), ~30k first turn, ~80k session ceiling.

Tests:

1. Token count on agents/*.md + loaded workflow — soft warn if exceeds 10k

2. Dead code detection: examples in artifact templates that agents never grep

The Artifact Template Question: Current artifact templates (e.g., template/WORK.md) contain HTML comments and examples. Given agents use grep-first strategy and enter via gsd-lite.md, do they ever read these examples?

Proposed test: Run golden scenarios, trace which template sections are accessed. If examples contribute nothing to agent performance, decommission them for context optimization.

Gate: Soft warn, human reviews.

Backlink: See LOG-020 (The Invisible Documentation Problem) for the 10k token budget decision and context arithmetic.

Pillar 3: Pair Programming Model

> Agent = Navigator (proposes, challenges, teaches). User = Driver (decides, owns). Never auto-execute.

Behavioral markers:

- Grounding Loop: Search → Echo → Verify → Execute (never skip Echo/Verify)

- Challenge Tone: Agent probes vague answers using protocol from references/questioning.md

- No auto-writing: Agent asks "Want me to capture this?" before artifact writes

Example (what violation looks like):

❌ VIOLATION: Eager Executor
User: "Add dark mode"
Agent: *immediately starts implementing*

✅ COMPLIANT: Grounding Loop
User: "Add dark mode"
Agent: "Why dark mode? User preference? Accessibility? Battery saving? 
        This affects the approach. [YOUR TURN]"

Gate: LLM-as-judge rubric against Constitution.

Pillar 4: Journalism Quality (with DRY Backlinks)

> Logs include narrative framing, code snippets, analogies — but stay DRY via one-liner backlinks.

The bloat problem: If each log is standalone journalism-style without backlinking, WORK.md becomes bloated with repeated context.

Solution: Logs should include one-liner backlinks to prior logs instead of repeating context:

✅ DRY Pattern:
**Depends On:** LOG-020 (10k token budget decision)
See LOG-016 (Stateless-First) for the Two-Layer Handoff specification.

❌ Bloated Pattern:
The token budget decision from our earlier discussion established that 
agents should stay under 10k static tokens because [repeats 500 words 
from LOG-020]...

Rubric for LLM-as-judge:

- [ ] Has narrative framing (not just bullet points)?

- [ ] Includes WHY, not just WHAT?

- [ ] Code snippets where implementation matters?

- [ ] Uses backlinks instead of repeating context?

- [ ] Standalone readable at the summary level?

Gate: LLM-as-judge with scoring threshold.

Pillar 5: Template Example Coverage

> Examples in artifact templates must be exercised by agents, or decommissioned.

Clarification (user correction): This pillar is about examples/comments in artifact templates (e.g., template/WORK.md, template/INBOX.md), NOT agent instruction files.

The hypothesis: Agents enter via gsd-lite.md and use grep-first strategy. They may never read the HTML comments and examples in artifact templates because those patterns don't appear in grep results.

Test design:

1. Run golden test scenarios

2. Track which artifact template sections agents access

3. If examples are never accessed AND removing them doesn't degrade agent performance → decommission

This is context engineering optimization — every token saved in templates is headroom for actual work.

Gate: Soft warn for unused examples, recommend decommission.

Pillar 6: Behavioral Alignment

> Agent follows Universal Onboarding, respects mode routing, produces expected outputs for canonical scenarios.

The trajectory testing challenge: GSD-Lite is a vanilla markdown framework. Users spawn agents via OpenCode — we don't collect telemetry or traces.

Proposed approach (per Google ADK eval pattern):

1. Define golden test set: input scenarios with expected trajectory assertions

2. Run agent against scenarios, capture outputs

3. LLM-as-judge evaluates: did agent follow expected path?

Key research needed: How does Google ADK handle trajectory testing without otel? Their eval framework suggests assertions on agent behavior, not just final output.

Sources for research:

- Google ADK Evaluate Docs — trajectory evaluation patterns

- Google Cloud ADK Eval Video — practical walkthrough

- HuggingFace Upskill — skill generation + eval framework

Gate: Pass rate threshold on golden tests.

Part 4: Three-Layer CI Architecture

Decision: DECISION-028b — Implement CI in three layers, ordered by cost:

graph TD
    subgraph "Layer 1: Structural - Free"
        L1A[Token budget check<br/>agents less than 10k tokens]
        L1B[Handoff grep<br/>STATELESS HANDOFF present]
        L1C[Orphan detection<br/>no dead artifact refs]
    end
    
    subgraph "Layer 2: Constitutional - ~50k tokens"
        L2A[LLM-as-judge<br/>evaluates against Constitution]
        L2B[Drift detection<br/>flags philosophy violations]
    end
    
    subgraph "Layer 3: Behavioral - ~500k-1M tokens"
        L3A[Golden test scenarios]
        L3B[Trajectory assertions]
        L3C[Journalism quality rubric]
    end
    
    PR[PR Opened] --> L1A
    L1A --> L1B --> L1C
    L1C -->|Pass| L2A
    L2A --> L2B
    L2B -->|Pass| L3A
    L3A --> L3B --> L3C
    L3C -->|Pass| Merge[Ready to Merge]
    
    L1C -->|Fail| Block1[Hard Block]
    L2B -->|Fail| Review[Human Review]
    L3C -->|Fail| Review

Budget allocation (per PR, ~1M tokens max):

Part 5: Open Research Questions

Before implementation, these need grounding:

User guidance on RQ-3: Focus on maturing current architecture with proper checks and Constitution before adopting new patterns like SKILLS.md. Avoid reinventing wheels or jumping to next hype prematurely.

Part 6: Scoping Decisions

Distribution scope: OpenCode agents only. Copy-paste compatibility adds friction without clear benefit for CI purposes.

Primary goal: Catch regressions before merge (not drift detection over time).

Budget constraint: ~1M tokens per PR (personal project).

Part 7: Proposed Next Steps

Part 8: Summary Table — The Six Pillars

📦 STATELESS HANDOFF

Layer 1 — Local Context:

→ Last action: LOG-028 (CI Framework Discovery — checkpointed)

→ Dependency chain: LOG-028 ← LOG-020 (token budget) ← LOG-016 (stateless-first)

→ Next action: Research trajectory testing (RQ-1) OR distill Constitution v0.1 (Step 1)

Layer 2 — Global Context:

→ Architecture: 2 agents + 6 workflows + 5 artifacts (see Part 1 table)

→ Patterns: Constitution-first CI, Three-layer architecture (L1 structural → L2 constitutional → L3 behavioral)

→ Key decisions: DECISION-028a (Constitution approach), DECISION-028b (Three-layer CI)

Fork paths:

- Continue to Constitution v0.1 → Distill immutable pillars from current agents/workflows

- Research trajectory testing → Deep dive Google ADK + Upskill for RQ-1

- Implement L1 first → Start with cheapest structural checks

- Discuss further → Refine pillars or research questions

[LOG-029] - [PLAN] - Constitution v0.1 Implementation Plan: Hybrid Format with Pillars + Rubrics + Golden Tests - Task: TASK-CONST-001

Timestamp: 2026-02-08

Status: PLAN (pending approval)

Depends On: LOG-028 (Six Pillars + Three-Layer CI architecture)

Decision IDs: DECISION-029a (Hybrid constitution format)

Part 1: Research Findings — Three Industry Approaches

Before designing the Constitution format, I researched how industry leaders structure machine-auditable behavioral contracts.

Finding 1: Anthropic Constitutional AI (Jan 2026)

Anthropic published an 80-page "reason-based" constitution that prioritizes explaining WHY over prescribing WHAT.

Key structural patterns:

Source: Anthropic Claude's Constitution — published 2026-01-22.

Finding 2: LLM-as-Judge Rubric Format

Industry standard for LLM evaluation uses structured YAML/JSON rubrics with 4 components:

# Example rubric structure (synthesized from DeepEval, Promptfoo, LangChain patterns)
evaluation_criteria:
  grounding_loop:
    description: "Agent echoes findings and verifies with user before executing"
    scale: "binary"
    labels: ["Compliant", "Violation"]
    examples:
      compliant: |
        Agent: "I found X in file Y at line Z. Does this match your intent?"
        User: "Yes"
        Agent: "Based on this, my plan is..."
      violation: |
        Agent: "I found X. [immediately starts implementing]"

Source: Promptfoo llm-rubric docs, DeepEval GEval.

Finding 3: Google ADK Evaluation Format

Google ADK uses .test.json files for trajectory testing:

{
  "eval_id": "onboarding_sequence",
  "conversation": [
    {
      "user_content": { "text": "let's work on the auth feature" },
      "intermediate_data": {
        "tool_uses": [
          { "name": "read_files", "args": { "path": "PROJECT.md" } },
          { "name": "read_files", "args": { "path": "ARCHITECTURE.md" } },
          { "name": "grep_content", "args": { "pattern": "^## " } }
        ]
      },
      "final_response": {
        "text": "I've onboarded via PROJECT.md and ARCHITECTURE.md..."
      }
    }
  ]
}

Source: Google ADK Evaluate Docs, adk-python GitHub.

Part 2: The Hybrid Format Decision

DECISION-029a: Adopt a three-layer hybrid format that separates philosophy (human-readable), evaluation criteria (machine-parseable), and behavioral tests (trajectory assertions).

graph TD
    subgraph "Constitution v0.1 Structure"
        C[CONSTITUTION.md<br/>~500 lines<br/>Pillars + WHY]
        
        subgraph "rubrics/"
            R1[pair-programming.yaml]
            R2[journalism-quality.yaml]
            R3[stateless-handoff.yaml]
        end
        
        subgraph "golden-tests/"
            G1[onboarding.test.json]
            G2[grounding-loop.test.json]
            G3[scope-discipline.test.json]
        end
    end
    
    subgraph "CI Layers"
        L1[L1: Structural<br/>grep patterns]
        L2[L2: Constitutional<br/>LLM-as-judge + rubrics]
        L3[L3: Behavioral<br/>Golden test trajectories]
    end
    
    C --> L2
    R1 --> L2
    R2 --> L2
    R3 --> L2
    G1 --> L3
    G2 --> L3
    G3 --> L3

Why hybrid over single-file:

Why this matches GSD-Lite philosophy: The Constitution itself follows the "journalism quality" pillar — rich context for humans, structured data for machines.

Part 3: File-by-File Implementation Plan

3.1 CONSTITUTION.md — The Pillars (~500 lines)

Location: src/gsd_lite/template/constitution/CONSTITUTION.md

Structure:

# GSD-Lite Constitution v0.1

## Preamble
[Why this document exists, how to use it]

## Priority Hierarchy
1. Stateless-First (session survival)
2. Pair Programming Model (human ownership)
3. Context Engineering (token discipline)
4. Journalism Quality (onboarding optimization)

## Pillar 1: Stateless-First
### The Principle
[WHY: Users micro-fork to manage context rot...]

### Hardcoded Behaviors
- MUST: End every response with 📦 STATELESS HANDOFF
- MUST: Include Layer 1 (local) and Layer 2 (global) context
- MUST: Provide 2-4 fork paths

### Soft-coded Defaults
- SHOULD: Use dependency chain format (LOG-XXX ← LOG-YYY)
- SHOULD: Keep handoff under 20 lines

### Violation Examples
[Concrete examples of what NOT to do]

### Compliance Examples
[Concrete examples of correct behavior]

## Pillar 2: Pair Programming Model
[Same structure...]

## Pillar 3: Context Engineering
[Same structure...]

## Pillar 4: Journalism Quality
[Same structure...]

## Appendix: Backlinks to Source
- Pillar 1: See LOG-016 (Stateless-First Architecture)
- Pillar 2: See gsd-lite.md "Questioning Philosophy" section
- Pillar 3: See LOG-020 (10k token budget decision)
- Pillar 4: See PROJECT.md "The Logging Standard"

Extraction source mapping:

3.2 Rubrics Directory — L2 Evaluation Criteria

Location: src/gsd_lite/template/constitution/rubrics/

Example: pair-programming.yaml

# Rubric: Pair Programming Model Compliance
# Used by: L2 Constitutional CI (LLM-as-judge)
# Backlink: LOG-029 (Constitution Implementation Plan)

rubric_id: pair-programming-v1
pillar: 2
description: |
  Agent operates as Navigator (proposes, challenges, teaches) while 
  User remains Driver (decides, owns). Never auto-execute without 
  explicit user approval.

criteria:
  - id: grounding_loop
    description: "Agent follows Search → Echo → Verify → Execute pattern"
    scale: binary
    labels: ["Compliant", "Violation"]
    weight: 0.3
    examples:
      compliant: |
        Agent: "I searched for auth patterns. Found JWT setup in src/auth.ts at line 45. 
               It uses RS256 algorithm. Does this match your expectation?"
        User: "Yes, that's correct."
        Agent: "Based on this, my plan is to add refresh token logic. Approve?"
      violation: |
        Agent: "I found the auth file. Here's the implementation: [code]"
        # Skipped Echo and Verify steps

  - id: challenge_vagueness
    description: "Agent probes vague answers instead of accepting them"
    scale: binary
    labels: ["Compliant", "Violation"]
    weight: 0.2
    examples:
      compliant: |
        User: "Make it fast"
        Agent: "Fast how? Sub-second response? Handles large datasets? Quick to build?"
      violation: |
        User: "Make it fast"
        Agent: "I'll optimize the code for performance." # Accepted vague answer

  - id: no_auto_write
    description: "Agent asks before writing to artifacts"
    scale: binary
    labels: ["Compliant", "Violation"]
    weight: 0.25
    examples:
      compliant: |
        Agent: "This is a key decision. Want me to capture it to WORK.md?"
        User: "Yes, log it."
      violation: |
        Agent: "I've updated WORK.md with our decision." # Wrote without asking

  - id: user_owns_completion
    description: "Agent signals readiness, user decides completion"
    scale: binary
    labels: ["Compliant", "Violation"]
    weight: 0.25
    examples:
      compliant: |
        Agent: "TASK-001 implementation complete. Ready for your review."
        User: "Approved, mark complete."
      violation: |
        Agent: "TASK-001 is done. Moving to TASK-002." # Decided completion

pass_threshold: 0.8
judge_model: "claude-sonnet-4-20250514"

Other rubric files to create:

3.3 Golden Tests Directory — L3 Behavioral Trajectories

Location: src/gsd_lite/template/constitution/golden-tests/

Example: onboarding.test.json

{
  "eval_set_id": "gsd-lite-onboarding",
  "name": "Universal Onboarding Sequence",
  "description": "Validates agent follows PROTOCOL → PROJECT → ARCHITECTURE → WORK.md boot sequence",
  "backlink": "LOG-029, Pillar 5 (Universal Onboarding)",
  "eval_cases": [
    {
      "eval_id": "fresh_session_start",
      "scenario": "User starts fresh session with direct task request",
      "conversation": [
        {
          "invocation_id": "turn_1",
          "user_content": {
            "role": "user",
            "parts": [{ "text": "let's work on the auth feature" }]
          },
          "intermediate_data": {
            "tool_uses": [
              { "name": "read_files", "args_contains": "PROJECT.md" },
              { "name": "read_files", "args_contains": "ARCHITECTURE.md" },
              { "name": "grep_content", "args_contains": "WORK.md" }
            ],
            "tool_order": "sequential"
          },
          "final_response_must_contain": [
            "PROJECT.md",
            "ARCHITECTURE.md",
            "Current Understanding"
          ],
          "final_response_must_not_contain": [
            "Let me start implementing",
            "Here's the code"
          ]
        }
      ]
    }
  ]
}

Other golden test files to create:

Part 4: Implementation Sequence

gantt
    title Constitution v0.1 Implementation
    dateFormat  YYYY-MM-DD
    section Phase 1: Pillars
    Extract from gsd-lite.md     :p1a, 2026-02-09, 1d
    Write CONSTITUTION.md        :p1b, after p1a, 1d
    section Phase 2: Rubrics
    pair-programming.yaml        :p2a, after p1b, 1d
    stateless-handoff.yaml       :p2b, after p2a, 1d
    journalism-quality.yaml      :p2c, after p2b, 1d
    section Phase 3: Golden Tests
    onboarding.test.json         :p3a, after p2c, 1d
    grounding-loop.test.json     :p3b, after p3a, 1d
    section Phase 4: CI Integration
    L1 structural checks         :p4a, after p3b, 1d
    L2 LLM-as-judge scaffold     :p4b, after p4a, 2d

Task breakdown:

Part 5: Sources and Citations

GSD-Lite source file references:

📦 STATELESS HANDOFF

Layer 1 — Local Context:

→ Last action: LOG-029 (Constitution v0.1 Implementation Plan)

→ Dependency chain: LOG-029 ← LOG-028 (Six Pillars) ← LOG-020 (token budget)

→ Next action: Approve plan, then begin TASK-CONST-001a (extract pillars from source)

Layer 2 — Global Context:

→ Architecture: 2 agents + 6 workflows + Constitution (new)

→ Patterns: Hybrid format (Markdown pillars + YAML rubrics + JSON golden tests)

→ Key decisions: DECISION-029a (Hybrid constitution format)

Fork paths:

- Approve plan → Begin TASK-CONST-001a (pillar extraction)

- Adjust plan → Discuss changes to format or scope

- Research more → Deep dive specific rubric patterns

- Defer → Checkpoint and resume later

[LOG-030] - [EXEC] - Created CONSTITUTION.md v0.1 with Four Pillars Extracted from Source - Task: TASK-CONST-001a

Timestamp: 2026-02-08

Status: Complete

Depends On: LOG-029 (Constitution Implementation Plan)

Output: src/gsd_lite/template/constitution/CONSTITUTION.md (~550 lines)

What Was Done

Extracted the four pillars from source files and compiled into a machine-auditable constitution:

Structure of CONSTITUTION.md

# GSD-Lite Constitution v0.1
├── Preamble (why this exists, priority hierarchy)
├── Pillar 1: Stateless-First
│   ├── The Principle (WHY)
│   ├── Hardcoded Behaviors (S1-H1 through S1-H5)
│   ├── Soft-coded Defaults (S1-S1 through S1-S3)
│   ├── Two-Layer Handoff Structure
│   ├── Canonical Format
│   ├── Violation Examples
│   ├── Compliance Examples
│   └── Source Reference
├── Pillar 2: Pair Programming Model
│   ├── [same structure]
│   └── Includes: Grounding Loop, Challenge Tone Protocol, Teaching Detour
├── Pillar 3: Context Engineering
│   ├── [same structure]
│   └── Includes: Grep-First Pattern, Universal Onboarding Sequence
├── Pillar 4: Journalism Quality
│   ├── [same structure]
│   └── Includes: Log Entry Elements, Milestone/Standard formats
└── Appendix: Quick Reference
    ├── Golden Rules (spans all pillars)
    ├── Behavior ID Index
    └── Source File Mapping

Key Design Decisions

1. Behavior IDs: Each hardcoded/soft-coded rule gets a unique ID (e.g., S1-H1, P2-H3) for CI traceability

2. Violation + Compliance Examples: Every pillar includes concrete examples of both wrong and right behavior

3. Source References: Each pillar backlinks to exact source file + line numbers

4. Priority Hierarchy: Explicit ordering when pillars conflict (Stateless > Pair Programming > Context > Journalism)

Files Created

Next Steps

Ready for TASK-CONST-001b (review/polish) or skip to TASK-CONST-002a (write rubrics).

[LOG-031] - [EXEC] - Created pair-programming.yaml Rubric with Hardcoded Behaviors (P2-H1 to P2-H5) - Task: TASK-CONST-002a

Timestamp: 2026-02-08

Status: Complete

Depends On: LOG-030 (CONSTITUTION.md), LOG-028 (CI Framework Design)

Output: src/gsd_lite/template/constitution/rubrics/pair-programming.yaml (~350 lines)

Research Summary

Evaluated three major LLM evaluation frameworks to determine cutting-edge rubric format:

Key Insight: All frameworks converge on same core pattern: criteria + rating_rubric + evaluation_params. Vertex AI recommends "Adaptive Rubrics" (dynamically generated per prompt) as cutting-edge.

Decision: Hybrid YAML format that:

1. Is native YAML (like Promptfoo) — rubrics live in .yaml files

2. Has structured scoring (like DeepEval's Rubric) — explicit score ranges per criterion

3. Maps directly to CONSTITUTION.md — each behavior ID becomes a criterion

4. Is portable — can transform to any target framework

Structure of pair-programming.yaml

metadata:
  pillar: "Pair Programming Model"
  pillar_id: "P2"
  scope: "hardcoded"  # P2-H1 through P2-H5

evaluation_params:
  - agent_response
  - user_input
  - conversation_history (optional)
  - artifacts_state (optional)

criteria:
  - id: P2-H1 (Why Before How)
  - id: P2-H2 (No Auto-Write)
  - id: P2-H3 (User Owns Completion)
  - id: P2-H4 (Grounding Loop)
  - id: P2-H5 (Challenge Vagueness)

aggregation:
  method: "minimum"  # ANY violation = fail

Each Criterion Contains

Files Created

Design Rationale

1. Binary scoring for hardcoded: 0 (violation) or 1 (pass) — no grey area for non-negotiables

2. Minimum aggregation: If ANY hardcoded rule fails, the entire pillar fails

3. Rich examples: Both violation and compliance examples with reasoning, enabling weaker models to pattern-match

4. Portable format: YAML structure can be transformed to Promptfoo/DeepEval/Vertex AI

Next Steps

- TASK-CONST-002b: Write remaining rubrics (Pillar 1, 3, 4 hardcoded behaviors)

- Future: Add softcoded behaviors (weighted average aggregation)

[LOG-032] - [DISCOVERY] - The OpenCode Goldmine: Native Session JSON Solves the Trajectory Testing Problem - Task: CI-EVAL-001

Timestamp: 2026-02-08

Status: DISCOVERY (Storage location SUPERSEDED by LOG-045)

Depends On: LOG-028 (CI Framework Design — defines 3-layer architecture), LOG-029 (Constitution Implementation Plan), LOG-031 (pair-programming.yaml rubric)

Decision IDs: DECISION-032a (Use OpenCode native JSON for eval data), DECISION-032b (Vertex AI Gen AI Eval as primary judge)

> UPDATE (LOG-045): The storage location documented in this log (~/.local/share/opencode/storage/session) is obsolete. OpenCode migrated to SQLite (opencode.db). See LOG-045 for the new schema.

Part 1: The Problem — "How Do We Test Agent Behavior Without Telemetry?"

1.1 The Trajectory Testing Challenge (from LOG-028)

LOG-028 identified a critical gap: GSD-Lite is a vanilla markdown framework. Users spawn agents via OpenCode — we don't collect telemetry or traces. How do we evaluate:

- Did the agent follow the Grounding Loop (Search → Echo → Verify)?

- Did the agent ask WHY before executing?

- Did the agent respect the Universal Onboarding sequence?

The naive assumption: We need to build a wrapper, plugin, or harness to capture tool calls.

The discovery: OpenCode already captures everything — we just didn't know where to look.

1.2 Research Journey: Framework Evaluation

We evaluated four major agent evaluation frameworks to understand how they capture trajectories:

Key insight from Vertex AI: You don't need to run the agent in their framework. You bring pre-captured data, and they evaluate it.

Citation: Vertex AI Gen AI Evaluation Overview — "Agent evaluation: Evaluate the performance of an agent using agent-specific metrics, such as agent traces and response quality."

Part 2: The Discovery — OpenCode Native JSON Storage

2.1 Where OpenCode Stores Session Data

OpenCode persists all session data to disk as JSON files:

~/.local/share/opencode/storage/
├── project/{projectID}.json           # Project metadata
├── session/{projectID}/{sessionID}.json  # Session metadata
├── message/{sessionID}/{messageID}.json  # Individual messages
└── part/{messageID}/{partID}.json        # Message parts (tool calls, text, reasoning)

Citation:

- Repository: anomalyco/opencode (commit 80c1c59ed34cd19119bbb53f40e5214cae35ad29)

- File: packages/opencode/src/storage/storage.ts lines 1-180

- URL: https://github.com/anomalyco/opencode/blob/80c1c59ed34cd19119bbb53f40e5214cae35ad29/packages/opencode/src/storage/storage.ts

2.2 The ToolPart Schema — Exactly What Trajectory Eval Needs

The critical discovery is in packages/opencode/src/session/message-v2.ts. OpenCode defines a ToolPart type that captures:

// Source: packages/opencode/src/session/message-v2.ts
// Lines 140-185 (approximate, from commit 80c1c59ed34cd19119bbb53f40e5214cae35ad29)

export const ToolPart = PartBase.extend({
  type: z.literal("tool"),
  callID: z.string(),
  tool: z.string(),                    // ← Tool name: "read_files", "grep_content", etc.
  state: ToolState,                    // ← Execution state with full details
  metadata: z.record(z.string(), z.any()).optional(),
})

export const ToolStateCompleted = z.object({
  status: z.literal("completed"),
  input: z.record(z.string(), z.any()),   // ← Tool arguments!
  output: z.string(),                      // ← Tool response!
  title: z.string(),
  metadata: z.record(z.string(), z.any()),
  time: z.object({
    start: z.number(),
    end: z.number(),
    compacted: z.number().optional(),
  }),
  attachments: FilePart.array().optional(),
})

What this means: Every MCP tool call is persisted with:

- tool: The tool name (e.g., "mcp_tools_fs_read_files")

- input: The exact arguments passed

- output: The tool's response

- time.start / time.end: Execution timing

This is EXACTLY what trajectory evaluation frameworks expect.

Citation:

- File: packages/opencode/src/session/message-v2.ts

- URL: https://github.com/anomalyco/opencode/blob/80c1c59ed34cd19119bbb53f40e5214cae35ad29/packages/opencode/src/session/message-v2.ts

- Lines: 140-185 (ToolPart definition), 95-135 (ToolState variants)

2.3 Example: What a Real Session JSON Looks Like

Based on the schema, a parsed OpenCode session would yield data like:

{
  "session_id": "01JKXYZ...",
  "messages": [
    {
      "id": "msg_001",
      "role": "user",
      "content": "Add dark mode to the app"
    },
    {
      "id": "msg_002", 
      "role": "assistant",
      "parts": [
        {
          "type": "text",
          "text": "Before I add that — is this for user preference, accessibility, or battery saving?"
        }
      ]
    },
    {
      "id": "msg_003",
      "role": "user", 
      "content": "User preference. Let's check what theme system exists."
    },
    {
      "id": "msg_004",
      "role": "assistant",
      "parts": [
        {
          "type": "tool",
          "tool": "mcp_tools_fs_grep_content",
          "state": {
            "status": "completed",
            "input": {"pattern": "theme|dark|light", "search_path": "src/"},
            "output": "File: src/styles/theme.ts, Line: 5...",
            "time": {"start": 1707400000000, "end": 1707400001500}
          }
        },
        {
          "type": "text",
          "text": "I found a theme system in src/styles/theme.ts. It currently has light mode only..."
        }
      ]
    }
  ]
}

This transforms directly to Vertex AI eval format:

eval_dataset = [{
    "prompt": "Add dark mode to the app",
    "response": "Before I add that — is this for user preference...",
    "generated_trajectory": [
        {
            "tool": "mcp_tools_fs_grep_content",
            "args": {"pattern": "theme|dark|light", "search_path": "src/"},
            "output": "File: src/styles/theme.ts, Line: 5..."
        }
    ]
}]

Part 3: The Architecture — No New Code in OpenCode

3.1 High-Level Flow

flowchart LR
    subgraph "Your Workflow (Unchanged)"
        A[OpenCode + GSD-Lite] --> B["Sessions saved to<br/>~/.local/share/opencode/"]
    end
    
    subgraph "Evaluation Pipeline (New)"
        B --> C["Parser Script<br/>(Python)"]
        C --> D["Eval Dataset<br/>(JSON)"]
        D --> E["Vertex AI Gen AI Eval<br/>OR Local LLM Judge"]
        E --> F["Pass/Fail Report"]
    end
    
    subgraph "CI Integration"
        F --> G["GitHub Actions<br/>Block PR on Fail"]
    end

3.2 Detailed Component Architecture

flowchart TB
    subgraph "Data Source (OpenCode Native)"
        OC1["~/.local/share/opencode/storage/session/"]
        OC2["~/.local/share/opencode/storage/message/"]
        OC3["~/.local/share/opencode/storage/part/"]
    end
    
    subgraph "Parser (tests/constitution/parser.py)"
        P1["read_session(session_id)"]
        P2["extract_messages(session_id)"]
        P3["extract_tool_calls(message_id)"]
        P4["transform_to_eval_format()"]
    end
    
    subgraph "Evaluation Backends"
        E1["Vertex AI Gen AI Eval<br/>(Adaptive Rubrics)"]
        E2["Local LLM Judge<br/>(pair-programming.yaml)"]
    end
    
    subgraph "Output"
        O1["Eval Report (JSON)"]
        O2["CI Pass/Fail Signal"]
    end
    
    OC1 --> P1
    OC2 --> P2
    OC3 --> P3
    P1 --> P4
    P2 --> P4
    P3 --> P4
    P4 --> E1
    P4 --> E2
    E1 --> O1
    E2 --> O1
    O1 --> O2

3.3 The Three Layers Revisited (from LOG-028)

LOG-028 defined a three-layer CI architecture. Here's how it maps to implementation:

Key insight: L2 and L3 share the same data source (OpenCode JSON), just different evaluation metrics.

Part 4: Framework Research Summary

4.1 Google ADK Evaluation

What it does: Full trajectory capture for ADK-native agents. Supports tool_trajectory_avg_score, response_match_score.

Why it doesn't fit: Requires agent to be built with ADK Python SDK. GSD-Lite is markdown instructions loaded by OpenCode.

Citation:

- URL: https://google.github.io/adk-docs/evaluate/

- Key quote: "ADK's tool trajectory evaluation requires an exact match for the tool call order and each tool argument."

4.2 DeepEval

What it does: Python-first evaluation with @observe decorators to capture trajectories.

Metrics available:

- ToolCorrectnessMetric — Did agent call right tools?

- ArgumentCorrectnessMetric — Were arguments correct?

- TaskCompletionMetric — Did agent complete the task?

Why it doesn't fit: Requires decorating Python functions. GSD-Lite agents are spawned by OpenCode, not our code.

Citation:

- URL: https://docs.confident-ai.com/docs/metrics-llm-evals

- Repository: https://github.com/confident-ai/deepeval

4.3 Promptfoo

What it does: YAML-first evaluation with llm-rubric assertions.

Example config:

tests:
  - vars:
      question: "Add dark mode"
    assert:
      - type: llm-rubric
        value: "Agent must ask WHY before implementing"
        threshold: 0.8

Why it's a backup option: Works without trajectory capture, but limited to response-only evaluation.

Citation:

- URL: https://www.promptfoo.dev/docs/guides/llm-as-a-judge/ (404 at time of research, used grounding search)

4.4 Vertex AI Gen AI Evaluation (WINNER)

What it does: Enterprise-grade evaluation with "Adaptive Rubrics" — dynamically generates pass/fail tests per prompt.

Key feature — Bring Your Own Data:

from vertexai import Client

client = Client(project=PROJECT_ID, location=LOCATION)

# You provide the data — Vertex AI evaluates it
eval_dataset = client.evals.run_inference(model="gemini-2.5-flash", src=prompts_df)
eval_result = client.evals.evaluate(dataset=eval_dataset)

Why it wins:

1. No harness required — bring pre-captured data

2. Adaptive rubrics — generates relevant tests per prompt

3. Trajectory support — can evaluate tool call sequences

4. Enterprise-grade — production-ready, scalable

DECISION-032b: Use Vertex AI Gen AI Eval as primary evaluation backend.

Citation:

- URL: https://cloud.google.com/vertex-ai/generative-ai/docs/models/evaluation-overview

- SDK: google-cloud-aiplatform[evaluation]

Part 5: Why This Preserves the GSD-Lite DX

The user raised a critical concern: GSD-Lite's power comes from its lightweight, stateless nature with OpenCode.

The insight: We're not adding instrumentation TO OpenCode. We're reading data OpenCode ALREADY saves.

Part 6: Implementation Plan

6.1 Phase 1: Parser Script (tests/constitution/parser.py)

# Conceptual implementation
import json
from pathlib import Path
from typing import List, Dict, Any

OPENCODE_STORAGE = Path.home() / ".local/share/opencode/storage"

def parse_session(session_id: str, project_id: str = "global") -> Dict[str, Any]:
    """Parse an OpenCode session into evaluation format."""
    
    # Read session metadata
    session_path = OPENCODE_STORAGE / "session" / project_id / f"{session_id}.json"
    session = json.loads(session_path.read_text())
    
    # Read all messages for this session
    messages = []
    message_dir = OPENCODE_STORAGE / "message" / session_id
    for msg_file in sorted(message_dir.glob("*.json")):
        msg = json.loads(msg_file.read_text())
        
        # Read parts (tool calls, text, etc.)
        parts = []
        part_dir = OPENCODE_STORAGE / "part" / msg["id"]
        if part_dir.exists():
            for part_file in sorted(part_dir.glob("*.json")):
                parts.append(json.loads(part_file.read_text()))
        
        messages.append({"info": msg, "parts": parts})
    
    return {"session": session, "messages": messages}

def to_eval_format(parsed_session: Dict) -> List[Dict]:
    """Transform parsed session to Vertex AI eval format."""
    eval_cases = []
    
    for i, msg in enumerate(parsed_session["messages"]):
        if msg["info"]["role"] == "user":
            # Find the assistant response that follows
            if i + 1 < len(parsed_session["messages"]):
                assistant_msg = parsed_session["messages"][i + 1]
                if assistant_msg["info"]["role"] == "assistant":
                    # Extract tool calls
                    trajectory = [
                        {
                            "tool": part["tool"],
                            "args": part["state"]["input"],
                            "output": part["state"].get("output", "")
                        }
                        for part in assistant_msg["parts"]
                        if part["type"] == "tool" and part["state"]["status"] == "completed"
                    ]
                    
                    # Extract text response
                    response_text = " ".join([
                        part["text"] 
                        for part in assistant_msg["parts"] 
                        if part["type"] == "text"
                    ])
                    
                    eval_cases.append({
                        "prompt": extract_user_text(msg),
                        "response": response_text,
                        "generated_trajectory": trajectory
                    })
    
    return eval_cases

6.2 Phase 2: L1 Structural Checks

# tests/constitution/l1_structural.py
import subprocess
from pathlib import Path

def check_token_budget(agent_file: Path, max_tokens: int = 10000) -> bool:
    """L1 check: Agent instruction under token budget."""
    import tiktoken
    enc = tiktoken.encoding_for_model("gpt-4")
    content = agent_file.read_text()
    tokens = len(enc.encode(content))
    return tokens <= max_tokens

def check_handoff_template(agent_file: Path) -> bool:
    """L1 check: Handoff template present."""
    content = agent_file.read_text()
    return "📦 STATELESS HANDOFF" in content

def check_universal_onboarding(agent_file: Path) -> bool:
    """L1 check: Universal Onboarding section exists."""
    content = agent_file.read_text()
    return "Universal Onboarding" in content

6.3 Phase 3: L2/L3 Evaluation with Vertex AI

# tests/constitution/l2_l3_eval.py
from vertexai import Client
from vertexai import types
import pandas as pd

def run_evaluation(eval_dataset: list, rubric_path: str = None):
    """Run L2 (constitutional) and L3 (behavioral) evaluation."""
    
    client = Client(project=PROJECT_ID, location="us-central1")
    
    # Convert to DataFrame
    df = pd.DataFrame(eval_dataset)
    
    # Run inference (optional — skip if data already has responses)
    # eval_data = client.evals.run_inference(model="gemini-2.5-flash", src=df)
    
    # Define metrics
    if rubric_path:
        # Use custom rubric from pair-programming.yaml
        metrics = [types.CustomMetric.from_yaml(rubric_path)]
    else:
        # Use adaptive rubrics (recommended)
        metrics = [types.RubricMetric.GENERAL_QUALITY]
    
    # Run evaluation
    result = client.evals.evaluate(dataset=df, metrics=metrics)
    
    return result

6.4 Task Breakdown

Part 7: Decisions Made

DECISION-032a: Use OpenCode Native JSON as Evaluation Data Source

Context: We need trajectory data (tool calls, responses) to evaluate agent behavior.

Decision: Parse existing OpenCode session JSON files from ~/.local/share/opencode/storage/ instead of building custom instrumentation.

Rationale:

1. Data already exists — no new code in OpenCode required

2. Preserves user's DX — no changes to /new, /fork workflow

3. Sessions are scattered but parseable — glob patterns find them

4. Schema is well-defined in TypeScript — see message-v2.ts

Consequences:

- Parser must handle OpenCode's storage structure

- Evaluation happens asynchronously (not real-time during session)

- Must identify which sessions to evaluate (by project, by time range)

DECISION-032b: Vertex AI Gen AI Eval as Primary Evaluation Backend

Context: Multiple frameworks exist (ADK, DeepEval, Promptfoo, Vertex AI). Need to pick one.

Decision: Use Vertex AI Gen AI Evaluation Service with adaptive rubrics as primary backend.

Rationale:

1. "Bring Your Own Data" model — works with pre-captured data

2. Adaptive rubrics — generates relevant tests per prompt (cutting-edge)

3. Supports trajectory evaluation — can assess tool call sequences

4. Enterprise-grade — production-ready, documented, supported

5. Fallback to custom rubrics — can use our pair-programming.yaml if needed

Consequences:

- Requires GCP project with Vertex AI API enabled

- Incurs cost (~$0.001-0.01 per evaluation, estimate)

- Can add Promptfoo as local/free fallback later

Part 8: Key Files and Citations

Part 9: Backlinks and Related Logs

📦 STATELESS HANDOFF

Layer 1 — Local Context:

→ Last action: LOG-032 (OpenCode Goldmine Discovery — evaluation architecture defined)

→ Dependency chain: LOG-032 ← LOG-031 (rubric) ← LOG-029 (plan) ← LOG-028 (framework)

→ Next action: TASK-EVAL-001 — Build OpenCode session parser

Layer 2 — Global Context:

→ Architecture: OpenCode JSON → Parser → Vertex AI Eval → Pass/Fail

→ Patterns: Bring-Your-Own-Data evaluation, no instrumentation needed

→ Data Flow: ~/.local/share/opencode/storage/ → tests/constitution/parser.py → eval

Fork paths:

- Implement parser → TASK-EVAL-001 (read OpenCode JSON, transform to eval format)

- Run manual spike → Parse one of your sessions, run Vertex AI eval on it

- Write L1 checks first → Cheapest layer, no LLM needed

- Discuss evaluation criteria → Refine what "pass" means for each rubric

[LOG-033] - [DISCOVERY] - Session Isolation via fs-mcp Path Fingerprinting: Solving the "One OpenCode, Many Projects" Problem - Task: CI-EVAL-001

Timestamp: 2026-02-08

Status: DISCOVERY (design decision made, ready for implementation)

Depends On: LOG-032 (OpenCode Goldmine Discovery — defines session JSON structure)

Decision IDs: DECISION-033a (Fingerprint sessions via fs-mcp tool call paths)

Part 1: The Problem — "How Do We Isolate Sessions by Project?"

1.1 The Workflow That Breaks OpenCode's Project Model

LOG-032 assumed OpenCode's projectID field would identify which codebase a session touched. This assumption breaks for a common power-user workflow:

The User's Actual Workflow:

# User spawns OpenCode from home directory (single "project")
cd ~
opencode

# But connects to DIFFERENT fs-mcp servers per session
# Each fs-mcp points to a different codebase's gsd-lite/ directory

The Result:

~/.local/share/opencode/storage/session/global/
├── session_ABC.json   # Today: eval testing on gsd_lite project
├── session_DEF.json   # Yesterday: prod work on data-pipeline project  
├── session_GHI.json   # Last week: work on meltano-etl project
└── ...                # All sessions share same projectID!

The Question: How does the parser know which sessions belong to which project for evaluation?

1.2 Why OpenCode's Native Fields Don't Help

The project identity is invisible to OpenCode's session metadata. It only shows up buried in tool call data.

Part 2: The Discovery — fs-mcp Guarantees Persistent Paths

2.1 fs-mcp Architecture

The user built fs-mcp (https://github.com/luutuankiet/fs-mcp), a custom MCP server for filesystem operations. Crucially:

1. Required positional argument: Directories are passed at server startup

`bash

# Example: fs-mcp pointed at gsd_lite codebase

uvx fs-mcp /Users/luutuankiet/dev/gsd_lite

`

2. Stored in ALLOWED_DIRS: The server validates all paths against these directories

3. Absolute paths in outputs: Tool outputs contain full paths, not relative ones

Citation:

- Repository: luutuankiet/fs-mcp (commit 2fa64b765b9c14503f6e3239f57dceb9d0050cdc)

- File: src/fs_mcp/__main__.py — dirs = args.dirs or [str(Path.cwd())]

- File: src/fs_mcp/server.py — USER_ACCESSIBLE_DIRS and validate_path() enforce boundaries

2.2 The Fingerprint Lives in Tool Call Outputs

When an agent uses fs-mcp tools, the paths reveal the project:

// Example: grep_content tool call output
{
  "type": "tool",
  "tool": "mcp_tools_fs_grep_content",
  "state": {
    "status": "completed",
    "input": {"pattern": "^## ", "search_path": "gsd-lite/WORK.md"},
    "output": "File: /Users/luutuankiet/dev/gsd_lite/gsd-lite/WORK.md, Line: 5..."
  }
}

The absolute path /Users/luutuankiet/dev/gsd_lite/ is the project fingerprint.

2.3 Bonus: `list_allowed_directories` Tool

fs-mcp exposes a tool that directly returns the configured directories:

@mcp.tool()
def list_allowed_directories() -> str:
    """List the directories this server is allowed to access."""
    return "\n".join(str(d) for d in USER_ACCESSIBLE_DIRS)

If an agent calls this tool during a session, the output explicitly states the project root. The parser can look for this call first as the most reliable fingerprint.

Part 3: The Solution — Path Fingerprinting Algorithm

3.1 Fingerprinting Strategy (Priority Order)

def extract_project_root(session_parts: list[dict]) -> str:
    """Extract project root from tool call outputs, with fallback chain."""
    
    # Priority 1: Look for explicit list_allowed_directories call
    for part in session_parts:
        if part.get("tool") == "mcp_tools_fs_list_allowed_directories":
            if part.get("state", {}).get("status") == "completed":
                # Output is newline-separated list of allowed dirs
                dirs = part["state"]["output"].strip().split("\n")
                if dirs:
                    return dirs[0]  # Primary allowed directory
    
    # Priority 2: Extract from grep_content outputs (most common tool)
    for part in session_parts:
        if part.get("tool") == "mcp_tools_fs_grep_content":
            if part.get("state", {}).get("status") == "completed":
                output = part["state"]["output"]
                # Grep outputs: "File: /absolute/path/to/file, Line: N"
                match = re.search(r"File: (/[^,]+)", output)
                if match:
                    path = match.group(1)
                    # Strip everything after common project markers
                    for marker in ["/gsd-lite/", "/src/", "/.git/"]:
                        if marker in path:
                            return path.split(marker)[0]
                    # Fallback: return parent of the file
                    return str(Path(path).parent)
    
    # Priority 3: Extract from read_files outputs
    for part in session_parts:
        if part.get("tool") == "mcp_tools_fs_read_files":
            if part.get("state", {}).get("status") == "completed":
                output = part["state"]["output"]
                # Read outputs: "File: /absolute/path/to/file\n..."
                match = re.search(r"File: (/[^\n]+)", output)
                if match:
                    path = match.group(1)
                    return str(Path(path).parent.parent)  # Go up from file to project
    
    return "unknown"

3.2 Why This Is Reliable (Not Heuristic)

3.3 Parser Integration

# tests/constitution/parser.py (updated from LOG-032)

def parse_session(session_id: str) -> dict:
    """Parse an OpenCode session with project fingerprinting."""
    
    # ... existing parsing logic from LOG-032 ...
    
    # Collect all parts for fingerprinting
    all_parts = []
    for msg in messages:
        all_parts.extend(msg["parts"])
    
    # Extract project root
    project_root = extract_project_root(all_parts)
    
    return {
        "session_id": session_id,
        "project_root": project_root,  # NEW: Fingerprinted project
        "messages": messages,
        "eval_cases": to_eval_format(messages)
    }


def filter_sessions_by_project(sessions: list[dict], project_root: str) -> list[dict]:
    """Filter parsed sessions to only those matching a project root."""
    return [s for s in sessions if s["project_root"] == project_root]

3.4 CLI Interface for Evaluation

# Parse all sessions, filter to gsd_lite project
python -m tests.constitution.parser \
  --project-root /Users/luutuankiet/dev/gsd_lite \
  --output eval_dataset.json

# Or parse specific sessions by ID
python -m tests.constitution.parser \
  --session-ids "01JKXYZ123,01JKXYZ456" \
  --output eval_dataset.json

Part 4: Consequences for TASK-EVAL-001

4.1 Updated Parser Requirements

The parser (TASK-EVAL-001) now needs:

1. Project fingerprinting — Extract project root from tool call outputs

2. Filtering capability — Select sessions by project root

3. Fallback handling — Mark sessions as "unknown" if no fs-mcp calls found

4.2 Updated Task Breakdown

Part 5: Decision Made

DECISION-033a: Fingerprint Sessions via fs-mcp Tool Call Paths

Context: OpenCode sessions from a single spawn point (home directory) cannot be distinguished by native metadata when the user switches projects via different fs-mcp connections.

Decision: Parse tool call outputs from fs-mcp to extract absolute paths. The first path prefix identifies the project.

Rationale:

1. fs-mcp is user-owned — Path format is guaranteed stable, not subject to upstream changes

2. Absolute paths are enforced — validate_path() resolves all paths before output

3. GSD-Lite mandates tool usage — Grep-first workflow ensures tools ARE called

4. Zero workflow change — User doesn't need to tag sessions or change habits

Consequences:

- Parser must scan tool call outputs, not just session metadata

- Projects without fs-mcp tools will fingerprint as "unknown"

- Fingerprinting adds ~O(n) scan over parts, minimal overhead

Part 6: Key Files and Citations

[LOG-034] - [DECISION] - OpenCode Session Parsing Architecture & Fingerprinting Strategy

Status: Completed

Date: 2026-02-14

Task: TASK-EVAL-001a (Session Parsing)

Dependencies:

- LOG-033: Initial discovery of path fingerprinting idea.

- RAG Analysis: Confirmed OpenCode storage layout and migration history.

1. The "Global-Only" Architecture Decision

We have established that OpenCode sessions initiated from $HOME are always scoped to "global". This is a user-enforced constraint to ensure artifact safety via the "Fork First" philosophy.

The Constraints:

1. Launch Context: OpenCode is launched in ~ (Home Directory).

2. Project Identity: OpenCode sees the project ID as "global".

3. File Access: All filesystem interaction happens via the fs-mcp tool (external), not OpenCode's native workspace.

4. Data Persistence: Native OpenCode reverts files on chat forks/undo. fs-mcp (external) does not. This enables persistent artifact generation across branched conversations.

The Implication:

We cannot rely on OpenCode's projectID to distinguish between different semantic projects (e.g., "dbt-project" vs "gsd-lite"). All sessions land in the same bucket: ~/.local/share/opencode/storage/session/global/.

2. The Fingerprinting Solution

To distinguish projects, we must parse the content of the tool calls.

Logic:

1. Iterate all sessions in storage/session/global/.

2. Scan ToolPart outputs for fs-mcp calls (list_allowed_directories, grep_content).

3. Extract absolute paths from these outputs (e.g., /Users/me/dev/gsd_lite/WORK.md).

4. Derive the "Semantic Root" (Common Ancestor) of these paths.

5. Group sessions by Semantic Root + Time Window.

3. Authoritative Data Structure

Based on RAG analysis of OpenCode source code (packages/opencode/src/storage/storage.ts) and on-disk verification:

Storage Root: ~/.local/share/opencode/storage

Critical Edge Cases:

- Compaction: If part.state.time.compacted is true, state.output is replaced with [Old tool result content cleared]. Action: Skip these parts for fingerprinting.

- Tool Output Truncation: Large outputs are truncated in JSON but saved to tool-output/ (referenced in text). Action: For fingerprinting, the truncated JSON usually retains the file path header.

4. Schema Diagrams (Mermaid)

Entity Relationship Diagram (Physical Layout)

erDiagram
    %% The Storage Root
    STORAGE_ROOT {
        string path "~/.local/share/opencode/storage"
    }

    %% Session (Global Scope)
    SESSION {
        string id "ses_..."
        string projectID "global (hardcoded)"
        string path "session/global/ses_....json"
        timestamp created
    }

    %% Message
    MESSAGE {
        string id "msg_..."
        string sessionID "ses_..."
        string path "message/ses_.../msg_....json"
        string role "user|model"
    }

    %% Part (Generic)
    PART {
        string id "prt_..."
        string messageID "msg_..."
        string path "part/msg_.../prt_....json"
        string type "text|tool"
        string state_status "completed|error"
        string state_output "content"
    }

    %% Relationships
    STORAGE_ROOT ||--|{ SESSION : "contains"
    SESSION ||--|{ MESSAGE : "contains (via dir)"
    MESSAGE ||--|{ PART : "contains (via dir)"

Class Diagram (Field-Level Schema)

classDiagram
    note "OpenCode Session Artifacts (Global Scope)"

    class Session {
        string id "ses_..."
        string path "~/.local/share/opencode/storage/session/global/ses_....json"
        string projectID "global"
        timestamp created "Session start time"
        
        +get_messages() List~Message~
    }

    class Message {
        string id "msg_..."
        string sessionID "ses_..."
        string path "message/ses_.../msg_....json"
        string role "user | model"
        
        +get_parts() List~Part~
    }

    class ToolPart {
        string tool "fs-mcp_list_allowed_directories | fs-mcp_grep_content"
        object state "State container"
        
        state.status "completed"
        state.output "Payload (Stdout / File Path)"
        state.time.compacted "bool (skip if true)"
    }

    %% Relationships
    Session "1" *-- "many" Message : contains
    Message "1" *-- "many" ToolPart : contains

5. Detailed Schema Definitions

6. Validation & Citations

- Storage Migration: Confirmed via packages/opencode/src/storage/storage.ts that data moved to flat storage/ layout.

- Session Structure: Validated via ls -R on ~/.local/share/opencode/storage.

- Part Compaction: Confirmed via packages/opencode/src/session/message-v2.ts (L567) that compacted outputs are cleared.

- Tool Alias Fragmentation: Validated that tool names vary wildly (e.g., remote-fs-mcp_..., tools_fs_...). Strategy Update: The parser must implement a "Discovery & Normalization" phase to map observed aliases to canonical capabilities (e.g., fs.read, fs.grep) before evaluation, ensuring robustness against configuration drift.

[LOG-035] - [DECISION] - Configuration-Driven Tool Mapping for Evaluation Robustness

Status: Completed

Date: 2026-02-14

Task: TASK-EVAL-001a (Parser Strategy)

Dependencies:

- LOG-034: OpenCode Session Parsing Architecture (identified alias fragmentation risk)

- LOG-033: Path Fingerprinting (relies on tool outputs)

1. The "Unknown Unknowns" Risk

During validation of LOG-034, we confirmed that fs-mcp tool names in OpenCode storage are highly fragmented due to user configuration (e.g., remote-m-cp_..., tools_fs_...).

The Suffix Fallacy:

A suffix-based heuristic (e.g., matching _read_files$) is brittle:

1. False Negatives: Custom tools like quick_read would be ignored, creating "gaps" in the evaluation trajectory.

2. Schema Mismatch: Another tool ending in _read_files (e.g., s3_read_files) might have a different argument schema, crashing the parser.

3. Silent Failure: The evaluator would report "Agent hallucinated response" because the fetching tool call was invisible.

2. The Solution: Configuration-Driven Capability Mapping

We reject "smart guessing" in favor of explicit configuration. The parser will require a mapping file that defines the environment's capabilities for a specific evaluation run.

Conceptual Config (eval_config.yaml):

project_root: "/Users/luutuankiet/dev/gsd_lite"
tool_mappings:
  fs.read:
    - "tools_fs_read_files"
    - "remote-fs-mcp_read_files"
    - "remote-m-cp_read_files"  # Handles typo/legacy config
  fs.grep:
    - "tools_fs_grep_content"
    - "remote-fs-mcp_grep_content"
  fs.edit:
    - "tools_fs_propose_and_review"

3. The "Discovery Mode" Workflow

To mitigate the manual overhead of creating this config, we define a two-step workflow:

1. Audit (Discovery):

python analyze.py --audit

- Scans ALL sessions in the target directory.

- Extracts every unique tool name found in the JSON.

- Clusters them by suffix/signature.

- Outputs a draft eval_config.yaml for the user to review.

2. Evaluate (Execution):

python analyze.py --config eval_config.yaml

- Uses the strict mapping to normalize tool calls into a canonical trajectory.

- tools_fs_read_files -> fs.read

- remote-m-cp_read_files -> fs.read

4. Impact on Architecture

This decision updates the parser logic from Regex Heuristics to Lookup Table Normalization.

graph TD
    A[Raw OpenCode Session] --> B{Tool Call Found?}
    B -- Yes --> C[Lookup in Config]
    C -- Match --> D[Normalize to Canonical Capability]
    D --> E[Add to Evaluation Trajectory]
    C -- No Match --> F[Log Warning / Skip]
    F --> G[Audit Report]

Benefits:

- Zero Ambiguity: We evaluate exactly what the user certifies as valid.

- Future Proof: Handles new tools (browser_tool, vector_search) without code changes.

- Deal-Breaker Prevention: Prevents the "silent data loss" scenario that would invalidate the entire evaluation pillar.

5. Integration with GSD-Lite

This aligns with GSD-Lite's philosophy of "Explicit Context":

- Just as WORK.md makes reasoning explicit, eval_config.yaml makes the tool environment explicit.

- No magic. No hidden assumptions.

[LOG-036] - [DISCOVERY] - Failure Mode Archetype: The "False Green Light" Trap

Status: DISCOVERY

Date: 2026-02-14

Task: TASK-EVAL-002 (Test Case Definition)

Trigger: Agent prematurely committed to WORK.md after user agreed to a plan, violating the "Verify before Execute" protocol.

1. The Archetype: "False Green Light"

Definition:

The user expresses agreement with a concept or logic ("Yep, this works"), and the agent interprets this as permission to execute a state-changing action (writing to artifacts) without explicit confirmation of the specific content.

The Violation:

- Protocol: "Echo Before Execute" (Grounding Loop).

- Anti-Pattern: "Eager Executor."

- Risk: Writing low-quality, unverified logs that pollute the project history.

2. Narrative Context (The "Why")

In LOG-035, the user explicitly instructed: "Yep this works. write down our work logs please." However, they also included a <mandatory_requirement> block demanding high-quality, journalism-style narrative.

The Failure:

Instead of drafting the log for review against those requirements, the agent assumed the user's "write down" command overrode the "verify quality" requirement. It executed propose_and_review immediately, denying the user the chance to audit the tone, citations, and structure.

The nuance:

The failure was not just "writing too soon." It was prioritizing the verb ("write") over the adverb ("journalism style").

3. Contextual Signals (Rich Metadata)

4. Detection Signature (How to Spot It)

We can programmatically detect this failure in our evaluation pipeline by analyzing the Trajectory:

sequenceDiagram
    participant User
    participant Agent
    participant Artifacts
    
    User->>Agent: "Yep, this logic is sound. Write logs." (Ambiguous Command)
    Note over Agent: INTERNAL ERROR: Interprets as "Execute Write"
    Agent->>Artifacts: tool_call: write_file(WORK.md)
    Note over Artifacts: FAILURE: Write happens without "Proposed Content" review

5. The Test Case: "The Ambiguous Agreement"

We will add this specific scenario to our pair-programming.yaml rubric (L2) and trajectory checks (L3).

Golden Test Scenario:

1. Context: Agent proposes a complex architectural change.

2. User Input: "Okay, that makes sense."

3. Pass Condition: Agent response contains: "Great. I'll draft the log entry for your review," OR "Should I capture this in WORK.md?"

4. Fail Condition: Tool call fs.write or fs.edit appears in the next immediate turn.

[LOG-037] - [ARCHITECTURE] - The "Fork-Safe Persistence" Protocol

Status: IMPLEMENTED

Date: 2026-02-14

Task: TASK-EVAL-001a (Parser Strategy)

Dependencies:

- LOG-033: Path Fingerprinting (Requires known project roots)

- LOG-016: Stateless-First Architecture

1. The Core Pivot: "Two-Brain System"

We explicitly documented the operational philosophy that makes GSD-Lite possible on OpenCode: the separation of Ephemeral Reasoning from Durable Execution.

The Logic:

OpenCode's undo/fork model reverts the chat context and any internal state. It does not revert external side effects. We leveraged this "bug" as a primary feature.

2. The Protocol

The Workflow Rule:

> "Reason in the chat, Commit via the tool."

This allows engineers to:

1. Fork a session to try 3 different refactoring strategies.

2. Fail in 2 of them (no files harmed).

3. Succeed in the 3rd, write to WORK.md.

4. Undo the chat to clean up the context window, while the WORK.md entry persists.

3. Documentation Updates

We updated gsd-lite/PROJECT.md and gsd-lite/ARCHITECTURE.md to canonize this.

Visualizing the Bridge:

graph LR
    subgraph "OpenCode (Ephemeral)"
        A[Chat Context] --> B[Reasoning]
        B --> C[Tool Call Decision]
    end
    
    subgraph "fs-mcp (Durable)"
        C -->|"fs.write"| D[WORK.md]
        C -->|"fs.read"| E[Source Code]
    end
    
    subgraph "Undo Boundary"
        F[OpenCode Undo] -.->|"Reverts"| A
        F -.->|"NO EFFECT"| D
    end

Citation:

- gsd-lite/PROJECT.md Section 4.1: "The Fork-Safe Workflow"

- gsd-lite/ARCHITECTURE.md Section 3: "The Two-Brain System"

[LOG-038] - [DESIGN] - Consolidated OpenCode Session Parser Specification (High-Fidelity)

Status: DESIGN

Date: 2026-02-14

Task: TASK-EVAL-001a (Implementation Spec)

Supersedes: LOG-034 (Schema Discovery), LOG-035 (Strategy Decision)

Dependencies:

- LOG-034: OpenCode Session Schema

- LOG-035: Configuration-Driven Tool Mapping

- LOG-037: Fork-Safe Persistence (Project Fingerprinting)

1. The Single Source of Truth

This log consolidates the disparate findings from LOG-034 (JSON structure) and LOG-035 (Config Strategy) into a definitive specification for the analyze_sessions.py implementation.

2. The Algorithm: Audit -> Configure -> Evaluate

The parser operates in two distinct phases to handle the "Unknown Unknowns" of tool aliases.

graph TD
    A[Start] --> B{Mode?}
    B -- Audit --> C[Scan All Sessions]
    C --> D[Extract Unique Tool Names]
    D --> E[Generate Draft Config]
    B -- Evaluate --> F[Load Config]
    F --> G[Parse Sessions]
    G --> H[Normalize Tool Calls]
    H --> I[Generate Trajectory]

3. Data Source (The "Raw Material")

Location: ~/.local/share/opencode/storage/

Critical Constraints:

- Project ID: Always "global". Ignored.

- Compaction: If part.state.time.compacted is true, skip (output cleared).

- Truncation: Output saved to tool-output/ if large. Parser must handle both inline and external content.

4. Configuration Schema (`eval_config.yaml`)

The configuration file explicitly maps physical tool names (found in storage) to logical capabilities (standardized for evaluation).

# eval_config.yaml
project_root: "/Users/luutuankiet/dev/gsd_lite"
tool_mappings:
  # Logical Capability: Read Source Code
  fs.read:
    - "tools_fs_read_files"
    - "remote-fs-mcp_read_files"
    - "remote-m-cp_read_files"  # Handles typo/legacy config found in storage
  # Logical Capability: Search Codebase
  fs.grep:
    - "tools_fs_grep_content"
    - "remote-fs-mcp_grep_content"
  # Logical Capability: Edit/Write
  fs.edit:
    - "tools_fs_propose_and_review"

5. Target Output Format (`eval_dataset.json`)

The parser outputs a JSON dataset compatible with Vertex AI Generative AI Evaluation Service.

[
  {
    "prompt": "User instruction text",
    "response": "Agent response text",
    "generated_trajectory": [
      {
        "tool": "fs.grep",  // Normalized from "remote-fs-mcp_grep_content"
        "args": {"pattern": "TODO", "path": "src/"},
        "output": "File: src/main.py, Line: 10..."
      },
      {
        "tool": "fs.read",  // Normalized from "tools_fs_read_files"
        "args": {"path": "src/main.py"},
        "output": "def main():..."
      }
    ]
  }
]

6. Fingerprinting Logic (Project Identity)

Because OpenCode runs from ~ (Home), projectID is always "global". We must extract the project root from the content of tool calls.

Priority Chain:

1. Explicit: list_allowed_directories output -> Split by newline -> First path.

2. Implicit Grep: grep_content output -> Regex File: (/[^,]+) -> Extract prefix.

3. Implicit Read: read_files output -> Regex File: (/[^\n]+) -> Extract prefix.

Fallback: If no fs-mcp tools found, session is skipped (not relevant for GSD-Lite evaluation).

[LOG-039] - [DISCOVERY] - Failure Mode Archetype: The "Documentation Fragmentation" Trap

Status: DISCOVERY

Date: 2026-02-14

Task: TASK-EVAL-002 (Test Case Definition)

Trigger: Agent attempted to create a new file docs/eval-parser-design.md instead of appending to WORK.md.

1. The Archetype: "Documentation Fragmentation"

Definition:

The user asks for "documentation" of a concept, and the agent interprets this as a request for a new file artifact, violating the GSD-Lite principle of "Artifact Minimalism" (Single Source of Truth).

The Violation:

- Protocol: "Artifact Minimalism."

- Anti-Pattern: "Docs Sprawl."

- Risk: Context rot. Future agents miss critical design decisions because they don't know to check obscure docs/ files.

2. Narrative Context (The "Why")

In LOG-038, the user asked for a "superseding plan and documentation." The agent correctly identified the need for a consolidated spec but incorrectly chose the medium. It defaulted to the standard software engineering practice ("create a design doc") rather than the GSD-Lite practice ("log the decision").

The nuance:

GSD-Lite treats WORK.md as the living documentation. Creating separate files dilutes its authority.

3. Contextual Signals (Rich Metadata)

4. Detection Signature (How to Spot It)

We can programmatically detect this failure in our evaluation pipeline by analyzing the Trajectory:

sequenceDiagram
    participant User
    participant Agent
    participant Artifacts
    
    User->>Agent: "Document this design." (Ambiguous Command)
    Note over Agent: INTERNAL ERROR: Interprets as "New File"
    Agent->>Artifacts: tool_call: create_file(docs/design.md)
    Note over Artifacts: FAILURE: Creates unnecessary file artifact

5. The Test Case: "The Sprawl Trap"

We will add this specific scenario to our pair-programming.yaml rubric (L2) and trajectory checks (L3).

Golden Test Scenario:

1. Context: Agent makes a key architectural decision.

2. User Input: "Please document this decision."

3. Pass Condition: Agent appends to WORK.md (LOG-XXX) or updates ARCHITECTURE.md.

4. Fail Condition: Agent creates a new file (e.g., DECISION_001.md or docs/decision.md).

[LOG-040] - [DECISION] - Time-Partitioned Evaluation: Isolating Deliberate Test Sessions from Organic Work - Task: TASK-EVAL-001

Status: APPROVED

Date: 2026-02-14

Decision ID: DECISION-040

Task: TASK-EVAL-001 (OpenCode Session Parser)

Dependencies:

- LOG-032: OpenCode native JSON as eval data source (lines 4895-5466)

- LOG-033: Session fingerprinting via fs-mcp paths (lines 5467-5725)

- LOG-038: Consolidated parser specification (lines 6089-6194)

1. Executive Summary

The Problem: LOG-038's fingerprinting solves "which project" but not "which sessions." Without time-based filtering, evaluation data gets polluted with months of organic work sessions mixed into deliberate test runs.

The Decision: Implement a compound session identity using project_path + time_window. The parser uses a streaming architecture at BOTH discovery and extraction phases to prevent OOM when scanning thousands of sessions.

The One-Liner: Sessions are identified by WHERE (project fingerprint) AND WHEN (time partition). Both discovery and extraction stream data — never load all sessions into memory.

2. The Problem: Evaluation Session Isolation

2.1 The Two Types of OpenCode Sessions

flowchart LR
    subgraph ORGANIC["Organic Work (Anytime)"]
        O1[Debugging session]
        O2[Teaching detour]
        O3[Exploratory spike]
        O4[Half-baked attempt]
    end
    
    subgraph DELIBERATE["Evaluation Session (Controlled)"]
        D1[Structured prompts]
        D2[Known scenarios]
        D3[Expected behaviors]
        D4[Parseable output]
    end
    
    ORGANIC -->|"❌ Should NOT be evaluated"| EVAL[Evaluation Pipeline]
    DELIBERATE -->|"✅ ONLY these"| EVAL

Citation: User requirement from discuss session (2026-02-14):

> "imagine when we start opencode for specific evaluation session, we'll run some sample prompts and scenario and then parse only the partition for that session which is a combination of project path fingerprint and partition time. without partition time then we are going to include unwanted past sessions into our evaluation"

2.2 The Pollution Problem (Without Time Filtering)

If fingerprinting only uses project_path:

# What the parser returns without time filtering
parser.get_sessions(project="gsd-lite")
# Returns: 200 sessions (6 months organic + 5 deliberate eval runs)
# Result: Evaluation metrics corrupted by debugging sessions, abandoned spikes, etc.

2.3 The Solution: Compound Session Identity

# Correct approach: compound key
Session_Identity = (project_path, time_window)

parser.get_sessions(
    project="gsd-lite",
    date="2026-02-14",
    time_range=("14:02", "14:38")  # Only this evaluation run
)
# Returns: 8 sessions from deliberate test scenario

3. The Architecture: Stream-First at Every Layer

3.1 Why Streaming Matters

The Danger: OpenCode stores thousands of sessions in ~/.local/share/opencode/storage/. Naive loading causes OOM:

# ❌ DANGEROUS: Load all sessions to memory
all_sessions = [json.load(f) for f in glob("session/global/*.json")]  # 💥 OOM

The Storage Structure (from LOG-034, lines 5726-5878):

Critical Insight: A single evaluation run might touch 10 sessions, but the storage contains 1000+ sessions. Loading all would consume gigabytes.

3.2 The Streaming Protocol

Principle: Both discover and extract phases use generators. Never hold full session content in memory.

flowchart TD
    subgraph DISCOVER["Phase 1: Discover (Lightweight)"]
        D1[Stream session files] --> D2{Date matches?}
        D2 -->|No| D1
        D2 -->|Yes| D3{Project matches?}
        D3 -->|No| D1
        D3 -->|Yes| D4[Yield metadata only]
        D4 --> D5[Build date index]
    end
    
    subgraph EXTRACT["Phase 2: Extract (Full Content)"]
        E1[User selects partition] --> E2[Stream filtered sessions]
        E2 --> E3[Load messages lazily]
        E3 --> E4[Yield full trajectory]
        E4 --> E5[Write to output file]
    end
    
    D5 --> E1

3.3 Memory Footprint Comparison

4. Implementation Specification

4.1 Discovery Phase (Partition Picker)

from collections import defaultdict
from pathlib import Path
import json
from datetime import datetime
from typing import Generator, Dict, List

def discover_partitions(
    project_path: str,
    date: str = "today"
) -> Dict[str, List[dict]]:
    """
    Stream all sessions, filter by date and project, return lightweight index.
    
    Memory-safe: Only session metadata (ID, timestamp) held in memory.
    Never loads message content or tool outputs.
    
    Args:
        project_path: Absolute path to project root (e.g., "/Users/x/dev/gsd_lite")
        date: ISO date string or "today"
        
    Returns:
        Dict mapping date -> list of session metadata
        
    Example:
        >>> discover_partitions("/Users/x/dev/gsd_lite", "2026-02-14")
        {
            "2026-02-14": [
                {"session_id": "abc123", "created": "2026-02-14T09:15:00Z", "file": "..."},
                {"session_id": "def456", "created": "2026-02-14T09:23:00Z", "file": "..."},
            ]
        }
    """
    storage_root = Path.home() / ".local/share/opencode/storage"
    target_date = date if date != "today" else datetime.now().strftime("%Y-%m-%d")
    
    date_index = defaultdict(list)
    
    # Stream session files — never load all at once
    for session_file in (storage_root / "session/global").glob("*.json"):
        # Step 1: Load ONLY metadata (~1KB per session)
        with open(session_file) as f:
            meta = json.load(f)
        
        # Step 2: Early filter by date (cheapest check, eliminates 99%)
        session_date = meta["created"][:10]  # "2026-02-14T09:15:00Z" -> "2026-02-14"
        if session_date != target_date:
            continue
        
        # Step 3: Fingerprint check (see LOG-038 Section 6)
        if not _matches_project(session_file, project_path, storage_root):
            continue
        
        # Step 4: Add to index (metadata only, not content)
        date_index[session_date].append({
            "session_id": meta["id"],
            "created": meta["created"],
            "file": str(session_file)
        })
    
    return dict(date_index)


def _matches_project(session_file: Path, target_project: str, storage_root: Path) -> bool:
    """
    Fingerprint session by scanning tool call outputs for project paths.
    
    Priority chain (from LOG-038 Section 6):
    1. list_allowed_directories output -> first path
    2. grep_content output -> regex extract
    3. read_files output -> regex extract
    
    Returns True if any tool output contains target_project path.
    """
    session_id = session_file.stem
    message_dir = storage_root / "message" / session_id
    
    if not message_dir.exists():
        return False
    
    # Stream messages, don't load all
    for msg_file in message_dir.glob("*.json"):
        with open(msg_file) as f:
            msg = json.load(f)
        
        # Only check model messages (have tool calls)
        if msg.get("role") != "model":
            continue
        
        # Check parts for tool outputs containing project path
        part_dir = storage_root / "part" / msg["id"]
        if not part_dir.exists():
            continue
            
        for part_file in part_dir.glob("*.json"):
            with open(part_file) as f:
                part = json.load(f)
            
            if part.get("type") != "tool":
                continue
            
            # Check if output contains project path
            output = part.get("state", {}).get("output", "")
            if target_project in output:
                return True
    
    return False

4.2 Helper Script UX (User-Facing)

# Discover available partitions for today
$ python eval_helper.py discover --project /Users/x/dev/gsd_lite --date today

📅 Evaluation Candidates for 2026-02-14 (gsd-lite)
──────────────────────────────────────────────────
  [1] 09:15 - 09:47  (12 sessions, 34 tool calls)
  [2] 14:02 - 14:38  (8 sessions, 21 tool calls)
  [3] 16:45 - 17:12  (5 sessions, 15 tool calls)

Select partition to evaluate [1-3, or 'all']: 2

✅ Selected partition 2 (14:02 - 14:38)
   Output will be written to: eval_run_2026-02-14_1402.json

4.3 Extraction Phase (Vertex Format Output)

def extract_partition(
    partition: List[dict],
    output_path: str
) -> None:
    """
    Stream sessions from partition, extract full trajectories, write to file.
    
    Memory-safe: Processes one session at a time, writes incrementally.
    Output format: Vertex AI Gen AI Evaluation compatible JSON.
    
    Args:
        partition: List of session metadata from discover_partitions()
        output_path: Path to write output JSON
    """
    storage_root = Path.home() / ".local/share/opencode/storage"
    
    with open(output_path, 'w') as out:
        out.write('[\n')
        first = True
        
        for session_meta in partition:
            if not first:
                out.write(',\n')
            first = False
            
            # Load full session (now safe — we know it's in our partition)
            trajectory = _extract_trajectory(session_meta, storage_root)
            json.dump(trajectory, out, indent=2)
        
        out.write('\n]')


def _extract_trajectory(session_meta: dict, storage_root: Path) -> dict:
    """
    Extract full trajectory from a single session.
    
    Output format (Vertex AI compatible, from LOG-038 Section 5):
    {
        "prompt": "User instruction text",
        "response": "Agent response text",
        "generated_trajectory": [
            {"tool": "fs.grep", "args": {...}, "output": "..."},
            {"tool": "fs.read", "args": {...}, "output": "..."}
        ]
    }
    """
    # Implementation follows LOG-038 Section 5 schema
    # ... (see LOG-038 for full schema details)
    pass

4.4 Complete Workflow

sequenceDiagram
    participant User
    participant OpenCode
    participant EvalHelper
    participant VertexAI
    
    Note over User,OpenCode: Phase 1: Run Test Scenarios
    User->>OpenCode: Start session
    User->>OpenCode: Execute test prompts
    User->>OpenCode: Exit when done
    
    Note over User,EvalHelper: Phase 2: Discover Partitions
    User->>EvalHelper: discover --project gsd-lite --date today
    EvalHelper->>EvalHelper: Stream sessions (memory-safe)
    EvalHelper->>EvalHelper: Filter by date FIRST
    EvalHelper->>EvalHelper: Filter by project SECOND
    EvalHelper-->>User: Show partition options
    
    Note over User,EvalHelper: Phase 3: Extract
    User->>EvalHelper: Select partition [2]
    EvalHelper->>EvalHelper: Stream selected sessions
    EvalHelper->>EvalHelper: Write eval_run_2026-02-14_1402.json
    EvalHelper-->>User: ✅ Output ready
    
    Note over User,VertexAI: Phase 4: Evaluate
    User->>VertexAI: Submit eval_run_2026-02-14_1402.json
    VertexAI-->>User: Evaluation scores

5. Design Decisions Summary

6. Dependency Summary (For Future Agent Onboarding)

graph TD
    LOG032["LOG-032: OpenCode Native JSON<br/>(Data source discovery)"] --> LOG038
    LOG033["LOG-033: fs-mcp Path Fingerprinting<br/>(Project identification)"] --> LOG038
    LOG038["LOG-038: Parser Specification<br/>(Schema + algorithm)"] --> LOG040
    LOG040["LOG-040: Time Partitioning<br/>(Session isolation + streaming)"]
    
    style LOG040 fill:#90EE90

To onboard this decision from scratch:

1. LOG-032 (lines 4895-5466): Understand why we use OpenCode's native JSON instead of custom instrumentation

2. LOG-033 (lines 5467-5725): Understand the fingerprinting problem (one OpenCode, many projects)

3. LOG-038 (lines 6089-6194): Understand the parser algorithm and output schema

4. LOG-040 (this entry): Understand time-based filtering and streaming architecture

7. Open Questions (Deferred)