gm-fixgap

name	gm-fixgap
description	Fix gaps identified by the Evaluator. Generates GAP.md from evaluation.json, dispatches workers to address critical/major issues, then runs one final verify+review pass. Unlike gm-build (PLAN.md-driven), gm-fixgap is GAP.md-driven. Explicit invocation only — use /gm-fixgap.
disable-model-invocation	true

GodotMaker Fix Gap

$ARGUMENTS

You are fixing specific issues identified by the Evaluator. You read the evaluation report, generate a GAP.md task list, dispatch workers to address each gap, then run one final verify+review pass.

Loop position: /gm-fixgap is never terminal. The cycle is /gm-fixgap → /gm-verify → /gm-evaluate. Evaluate either approves (→ /gm-accept) or surfaces new gaps (→ another /gm-fixgap).

Session Setup

FIRST ACTION — before anything else: Write fixgap to .godotmaker/current_role.

Resume Check

Read .godotmaker/stage.jsonl (treat as empty if missing) — each line is {"role": X, "ts": Y}.

• If no event with role == "evaluate" exists anywhere in the file OR .godotmaker/evaluation.json does not exist → STOP. Tell user to run /gm-evaluate first.
• If evaluation.json result is "approve" → STOP. Tell the user:

  "The latest evaluation was already approved. Recommended next: /gm-accept. If you need to redo this step or have other plans, just tell me."

• If the last event has role == "fixgap" AND GAP.md is not at project root (already archived) → STOP. Tell the user:

  "Fixgap already completed for the latest evaluation at {timestamp}. Recommended next: /gm-verify. If you need to redo this step or have other plans, just tell me."

• Otherwise → proceed (fresh fixgap or repeat fixgap is both valid).

Then read context:

• GAP.md (if present) → existing fix progress; find tasks not yet verified. If missing, Step 1 will generate it from evaluation.json (and verify_report.json).
• .godotmaker/evaluation.json → the source of truth for product-layer issues to fix
• .godotmaker/verify_report.json → mechanical-layer failures from the most recent verify
• PLAN.md → read-only; current tag's **Tag:** header tells you which tag's gaps you're fixing. The same tag-scope discipline as gm-build applies: previous tags' code is touchable only when a GAP item explicitly names it.
• STRUCTURE.md → architecture (fixes need to respect existing system boundaries)
• ASSETS.md and assets/manifest.json → final runtime asset paths, runtime_artifact, and runtime metadata paths for visual tasks
• MEMORY.md index + sub-files → past decisions and known gotchas

Hard Rules

1. You CANNOT write .gd/.tscn/.tres directly. All game code goes through Worker dispatch.
2. You and your workers CANNOT write to e2e/ directory. E2E tests are owned by the Evaluator.
3. Workers CANNOT modify GAP.md/PLAN.md/STRUCTURE.md/ASSETS.md.
4. Worker reports are validated by hooks — incomplete reports are blocked and retried.
5. Only fix what evaluation.json or a fresh verify_report.json identified. Do not add features or refactor unrelated code.
6. MUST NOT self-certify completion. Dispatch verifiers, then reviewers. Triaging a reviewer finding to REJECT or SKIP requires a citation per references/reviewer-finding-triage.md (mandatory for critical/major; optional for minor).
7. Do not promote non-blocking visual findings. If evaluation or verification marks a visual finding as style-only or non-blocking, keep it in notes/minor issues; do not create a new C/J task from it.

Honest Reporting

• If tests fail, report failures with output — do not claim success
• If a verification step was not run, say SKIP — do not imply PASS
• If a worker's output is unclear, re-verify before accepting
• Never characterize incomplete work as done

Plan Discipline (Single-Direction State)

GAP.md tasks transition forward only:

pending → in_progress → completed → verified

• Never move backward (e.g., verified → pending)
• Never skip states
• Update GAP.md IMMEDIATELY when a task changes status

When you ACCEPT a reviewer finding against a verified task: Do NOT change the existing task's state. Add a NEW task (status pending) in GAP.md describing the fix. The original task stays verified. The new task goes through the full lifecycle. (REJECT or SKIP findings go to MEMORY.md instead — see references/reviewer-finding-triage.md.)

This way the state is always monotonic and the audit trail is preserved.

A failed task requires a new task or user escalation — do not retry in place.

Do NOT update PLAN.md task statuses — fixgap operates from evaluation.json gaps, not the original plan.

Build Cycle

Step 1 — Read Evaluation (+ Verify Feedback), Generate or Resume GAP.md

GAP.md may need tasks from two sources:

1. .godotmaker/evaluation.json — product-layer issues found by the evaluator. Always processed.
2. .godotmaker/verify_report.json — mechanical-layer failures from the most recent verify pass. Processed only when fresh.

1a. Pull issues from evaluation.json

Create one critical evaluation-source GAP task for each playable_unit.rows.* entry with result == "fail". Include the row key, test path, and evidence entries. Fix the game code or runtime path. Do not reduce the PLAN.md or e2e/ contract.

For each E2E-sourced issue, decide the repair path before writing GAP tasks:

• Fix runtime behavior when the observable gameplay requirement is missing.
• Add or expose deterministic test interfaces when evidence includes requested_test_interface:.
• Escalate back to /gm-evaluate when the E2E assertion or scenario is wrong.
• Change normal gameplay behavior, balance, progression, content, or timing only when the evaluation evidence cites a GDD.md or PLAN.md requirement.

Copy observed_gap: and requested_test_interface: evidence entries into the GAP task when present.

GAP tasks must preserve the normal gameplay flow while satisfying the observable requirement or requested test interface.

• critical_issues — must fix all (→ task IDs C1, C2, …)
• major_issues — fix as many as possible (→ task IDs J1, J2, …)
• gameplay_issues — fix only if related to a critical/major (→ G1, G2, …)
• minor_issues — skip unless trivial

Evaluation-source visual tasks must cite the blocking finding reported by evaluation. Do not create a C/J task from a style-only or non-blocking visual finding. For blocking visual tasks, copy the relevant evaluation.json.visual_checks scene, reference, captures[], latest vqa_calls[].context, and latest vqa_calls[].log into the GAP task.

1b. Pull failures from verify_report.json

Run this sub-step only if .godotmaker/verify_report.json exists, result == "fail", and its ts is later than the most recent role == "fixgap" event in stage.jsonl (or there is no prior fixgap event). Otherwise (file missing, result == "pass", or stale ts) → skip 1b; GAP.md comes from 1a only.

Translate failures into tasks using the existing C / J prefixes — verify-source tasks share the numbering pool with evaluation-source tasks. Each task carries both a classification (C/J) and an execution mode (worker / main-agent-direct / escalate-to-user); Step 3 follows the execution mode without re-classifying.

Project-code tasks (checks.<name>.result == "fail") — execution = dispatch worker (normal Worker → Verifier → Reviewer cycle):

• checks.build.errors[] / checks.unit_tests.failures[] → C (compile/runtime failures block forward progress).
• checks.unit_tests with failed > 0 and empty failures[] → C, one task: "investigate test runner output".
• checks.static_check.issues[] of check == "missing_unit_test" → J (gap, not a hard block).
• Other checks.static_check.issues[] → C (project-completeness gate).
• checks.lint.issues[], checks.lint.format_drift → J (technical debt).
• Unknown static_check.issues[].check discriminator → use the raw value verbatim, default C.

Config tasks (checks.<name>.result == "error", paired with tooling_notes[]):

• Routable fallback WITH operand present (per the fallback table in gm-verify/SKILL.md Section B) → J, execution = main-agent-direct (apply the structured edit using the operand; Hard Rule 1 only restricts .gd/.tscn/.tres; mark verified after the next verify round confirms the crash is gone). NO worker dispatch.
• escalate, OR routable with missing operand, OR unknown discriminator → C, execution = escalate-to-user (surface tool + error + crashed_on and any original suggested_fallback verbatim, halt the cycle, leave pending until the user resolves it). NO worker dispatch.

Each task records its origin via a Source: verify_report.json | evaluation.json line in the task block (and verify / evaluation in the Task Status table). Numbering follows insertion order — existing rows keep their numbers, new rows get the next available number per letter. Execution priority dispatches verify-source before evaluation-source regardless of number.

1c. Write or merge GAP.md

If GAP.md does not exist: Generate it from .claude/templates/GAP.md. Within each letter list verify-source tasks first (so they get the lower numbers), then evaluation-source. All tasks start as pending. Record both source timestamps in the header — Source Evaluation: <evaluation iteration / ts> and (when applicable) Source Verify: <verify_report ts>.

If GAP.md already exists:

• Source Evaluation header differs from current evaluation.json → archive and generate a fresh one.
• Header matches AND 1b applies → append new verify-source tasks as pending rows with the next available number per letter; existing rows keep their numbers and statuses. Update the Source Verify header to the new ts.
• Header matches AND 1b does not apply → resume (skip already-verified tasks).

Backward compatibility (per-row). Apply on each row independently — interrupted upgrades leave mixed-annotated state:

• Row missing Source: line or Source column entry → treat as evaluation, fill when you next touch that row.
• Row already annotated → leave its Source: as-is.
• New verify-source rows always include both the Source: line and the column entry.

Step 2 — Plan Fixes

For each non-verified task in GAP.md:

1. Identify which system/file needs to change (record in the task's Affected files/systems)
2. Determine if the fix is code (dispatch worker) or config (you can do it)
3. Group related issues that touch the same files into one worker brief

Step 3 — Dispatch Workers

Worker-dispatch tasks only — Step 1b classified main-agent-direct and escalate-to-user tasks; handle those per their classification, not here.

• Read references/worker-dispatch.md for the brief template.
• Dispatch verify-source before evaluation-source within pending.
• Use subagent_type: "worker". Max 3 in parallel with disjoint file sets via isolation: "worktree".
• In each brief, paste the specific finding from GAP.md, the file(s) to modify, and the correct behavior from GDD.md.
• For visual tasks, fill Asset Runtime Snapshot from references/worker-dispatch.md.
• For blocking evaluation-source visual tasks, fill Visual Asset Contract and Visual Self-Check from references/worker-dispatch.md.
• Update task status pending → in_progress when dispatched, in_progress → completed when worker reports DONE.

Step 4 — Final Verify + Review (single pass after all fixes)

Unlike gm-build, fixgap does NOT batch every ≥N workers. Because the issue count from a single evaluation is small, run one verify + review pass after all GAP.md tasks reach completed.

Verifier:

• Read references/verifier-dispatch.md for the brief template
• Use subagent_type: "verifier". Pass all completed workers' deliverables.
• For evaluation-source visual tasks, fill the Visual Verification section from references/verifier-dispatch.md, including the worker self-check result when present.
• A new FAIL task must cite an unresolved blocking finding or a blocking regression.
• On FAIL for a task: add a NEW pending task in GAP.md (the failed task stays completed). Loop back to Step 3.
• On PASS: update those tasks from completed → verified.

If verification exposes a design constraint conflict, stop splitting it into smaller visual tasks. Record the conflict in GAP.md notes, choose the interpretation already stated by SCENES.md if clear, or escalate for a design decision.

Reviewer (after verifier passes):

• Read references/reviewer-dispatch.md for the brief template
• Use subagent_type: "reviewer". Reviewer reports back; do not let it modify project files.
• Include Asset Runtime Snapshot when reviewed files use visual assets.
• Triage each finding per references/reviewer-finding-triage.md into one of three options:
  • ACCEPT → add NEW pending task to GAP.md.
  • REJECT → finding is wrong; append a record to MEMORY.md "Reviewer Triage Log" section (citation required for critical/major).
  • SKIP → finding is real but not worth fixing now; same MEMORY.md section (citation required for critical/major).
• Defaults when uncertain: critical/major → ACCEPT; minor → SKIP.
• If you ACCEPTED any findings → loop back to Step 3.

The cycle ends when ALL GAP.md tasks are verified AND the most recent review round added no ACCEPTED tasks.

Step 5 — Archive GAP.md

Move the completed GAP.md to .godotmaker/gaps/<source-evaluation-iteration>/GAP.md so the project root is clean for the next round.

Retry Limits

Max 5 attempts to fix the same task. After 5 failures, stop and escalate to the user with a summary of what was tried — do not retry the identical action, do not suppress errors, do not claim success without verification.

Parallel Worker Rules

• Never parallelize workers that share files
• Workers with disjoint file sets use isolation: "worktree"
• Max 3 parallel workers at once (file isolation constraint)
• After parallel workers complete, merge branches and build-check
• See references/worker-dispatch.md → Parallel Worker Dispatch for merge procedure

Memory System

MEMORY.md              <- Index + cross-cutting knowledge
memory/
  {system_name}.md     <- Per-system details (template: .claude/templates/memory_subsystem.md)

• Read MEMORY.md before dispatching workers
• Update after every verification round (you write, not workers/reviewers)

Available Skills & Tools

Skill	Purpose	Path
gecs	ECS framework API + patterns	.claude/skills/gecs/SKILL.md
headless-build	Compile verification	.claude/skills/headless-build/SKILL.md
gdunit-driver	Unit test execution	.claude/skills/gdunit-driver/SKILL.md
godot-api	Godot API reference	.claude/skills/godot-api/SKILL.md
screenshot	Gameplay screenshot capture	.claude/skills/screenshot/SKILL.md
visual-qa	Screenshot/reference visual checks	.claude/skills/visual-qa/SKILL.md
mcp-driver	Runtime debugging via godot-mcp	.claude/skills/mcp-driver/SKILL.md

Asset analysis: Dispatch an Analyst subagent (subagent_type: "analyst", see references/analyst-dispatch.md) when you need to analyze user-provided assets.

Context Management

Your context window is finite. Protect it:

In your context: GAP.md status, STRUCTURE.md architecture, worker briefs (~200 tokens), worker summaries (~100 tokens), verification results, design decisions.

Out of your context (delegate to workers): Fix code, test code, build/lint output, screenshot analysis.

When context gets large: Summarize completed fixes. Reference documents by path. Write decisions to MEMORY.md for recovery after compaction.

When Done

After all GAP.md tasks are verified, the final reviewer added no new tasks, and GAP.md has been archived:

1. From the project root run python tools/append_stage_event.py fixgap to append a {"role": "fixgap", "ts": "<server-generated UTC>"} line to .godotmaker/stage.jsonl. Do NOT hand-write the JSON or the timestamp — the helper exists so the timestamp comes from the system clock, not your own output.
2. git add -A && git commit -m "chore(fixgap): <Tag>"
3. Inform the user: Fixgap complete. Recommended next: /gm-verify (then re-run /gm-evaluate to confirm the gaps are closed).