@thierrynakoa/fire-flow 12.2.1 → 13.0.1

package/TROUBLESHOOTING.md

@@ -1,264 +1,264 @@
-# Dominion Flow Troubleshooting Guide
-
-Common failure modes and their fixes.
-
----
-
-## 1. Autonomous mode won't stop
-
-**Symptoms:**
-- `/fire-autonomous` or `/fire-loop` continues iterating past the point of usefulness
-- Agent ignores Sabbath Rest warnings and keeps working
-- Iteration count climbing with no meaningful progress
-
-**Likely Cause:**
-- `--max-iterations` set too high (default 50)
-- `--no-circuit-breaker` flag was passed, disabling the safety net
-- Completion promise text doesn't match actual output (e.g., "DONE" vs "Done")
-
-**Fix Steps:**
-1. Press `Ctrl+C` to interrupt the current iteration
-2. Run `/fire-loop-stop` to formally end the loop and save state
-3. Check `.planning/loops/fire-loop-*.md` for the loop state file
-4. Review the "Approaches Tried" section to see what was attempted
-
-**Prevention:**
-- Use `--max-iterations 20` for most tasks (30 max for complex work)
-- Never use `--no-circuit-breaker` unless you're monitoring closely
-- Write precise completion promises: `--completion-promise "ALL TESTS PASS"`
-- Use `--aggressive` for tighter stall/spin detection thresholds
-
----
-
-## 2. Wrong files modified
-
-**Symptoms:**
-- Files in a different project were edited by a subagent
-- Changes landed in the wrong directory
-- Cross-project contamination (e.g., editing `my-other-project` instead of `my-project`)
-
-**Likely Cause:**
-- Working directory didn't match the project in CONSCIENCE.md
-- Subagent resolved a relative path to the wrong project
-- Multiple project directories open in the same session
-
-**Fix Steps:**
-1. Run `git diff` in the affected project to see what changed
-2. Run `git checkout -- .` to revert unwanted changes (or `git stash` to save them)
-3. Check CONSCIENCE.md -- does it reference the correct project path?
-4. Verify `pwd` matches the expected project root before re-running
-
-**Prevention:**
-- The Path Verification Gate (Step 3.5 in `/fire-3-execute`) catches this automatically
-- Always use absolute paths when working across multiple projects
-- Run `/fire-0-orient` at session start to confirm you're in the right place
-- When spawning subagents, the `<path_constraint>` block is injected automatically -- verify it matches
-
----
-
-## 3. Memory features not working
-
-**Symptoms:**
-- Episodic memory injection (Step 7.1 in `/fire-loop`) returns no results
-- `/fire-reflect` can't search past reflections
-- "Connection refused" errors mentioning port 6335
-
-**Likely Cause:**
-- Qdrant is not running (native binary at `C:\path\to\qdrant\qdrant.exe`)
-- Qdrant is running but on the wrong port (6333 = Docker backup, 6335 = native primary)
-- Ollama embedding service is down (needed for `nomic-embed-text` 768d vectors)
-
-**Fix Steps:**
-1. Check Qdrant health: `curl http://localhost:6335/healthz`
-2. If down, start it: `C:\path\to\qdrant\qdrant.exe` (runs on port 6335)
-3. Check Ollama: `curl http://localhost:11434/api/tags`
-4. If Ollama is down: `ollama serve` in a separate terminal
-5. Verify the collection exists: `curl http://localhost:6335/collections/power_flow_memory`
-
-**Prevention:**
-- Dominion Flow degrades gracefully: if Qdrant is unreachable, it falls back to file-based search across `~/.claude/warrior-handoffs/` and `~/.claude/reflections/`
-- The file-based fallback is slower but functional -- memory retrieval is never silently skipped
-- Add Qdrant to your system startup if you use memory features regularly
-
----
-
-## 4. Verification always fails
-
-**Symptoms:**
-- `/fire-4-verify` consistently returns REJECTED or CONDITIONAL
-- Score stuck below 49 even for seemingly complete work
-- E2E (Playwright) category drags the score down on non-UI projects
-
-**Likely Cause:**
-- The 70-point checklist includes 10 points for E2E/Playwright testing -- irrelevant for CLI, API-only, or library projects
-- Documentation category (10 points) penalizes when README/JSDoc isn't written yet
-- Validation thresholds: 63-70 = APPROVED, 56-62 = APPROVED*, 49-55 = CONDITIONAL, <42 = REJECTED
-
-**Fix Steps:**
-1. Check `.planning/phases/{N}-{name}/{N}-VERIFICATION.md` for the breakdown
-2. Identify which categories are scoring zero
-3. For non-UI projects: the E2E category should be redistributed to other categories by the verifier -- if it isn't, note this as a configuration issue
-4. Focus on the must-haves first: if all must-haves pass, CONDITIONAL is often acceptable for early phases
-
-**Prevention:**
-- Run `/fire-4-verify` with awareness of what the 7 categories are: Code Quality /10, Testing /10, Security /10, Performance /10, Documentation /10, Infrastructure /10, E2E /10
-- For API-only projects, tell the verifier in the plan: "No E2E -- redistribute points to Testing and Security"
-- Use `/fire-double-check` for a quick sanity check before the full 70-point verification
-
----
-
-## 5. Plans executed in wrong order
-
-**Symptoms:**
-- Breath 2 tasks fail because Breath 1 outputs don't exist yet
-- Tasks with dependencies execute before their dependencies complete
-- File conflicts between parallel executors
-
-**Likely Cause:**
-- `breath:` frontmatter in BLUEPRINT.md files is incorrectly assigned
-- `depends_on:` field missing or wrong in the plan files
-- Execution mode selected PARALLEL when it should have been SEQUENTIAL (file overlap)
-
-**Fix Steps:**
-1. Open `.planning/phases/{N}-{name}/` and check each BLUEPRINT.md frontmatter
-2. Verify `breath:` numbers -- all tasks in Breath 1 should have no external dependencies
-3. Verify `depends_on:` lists reference only tasks in earlier breaths
-4. Re-run `/fire-2-plan {N}` to regenerate plans with correct breath assignment
-
-**Prevention:**
-- The planner agent validates breath assignments during `/fire-2-plan`
-- `/fire-3-execute` auto-detects file overlap and downgrades from PARALLEL to SEQUENTIAL
-- Review the execution manifest at Step 2 before proceeding
-
----
-
-## 6. Context window filling up fast
-
-**Symptoms:**
-- Claude starts forgetting earlier instructions mid-session
-- Output quality degrades after 10-15 iterations
-- Sabbath Rest triggers repeatedly at low iteration counts
-
-**Likely Cause:**
-- Large files loaded into context (reading entire source files instead of relevant sections)
-- Skills library injecting too many skills per iteration (max should be 3)
-- No `.powerignore` file to exclude noise (node_modules, build artifacts, etc.)
-
-**Fix Steps:**
-1. Run `/compact` with a focus topic: `/compact Focus on phase 3 execution`
-2. Delegate research tasks to subagents instead of doing them in the main context
-3. Create a `.powerignore` file in the project root listing directories to exclude
-4. For `/fire-loop`: if iteration > 15, accept the Sabbath Rest and `/fire-loop-resume` in a fresh context
-
-**Prevention:**
-- Keep iteration output concise -- the recitation block is capped at 30 lines for a reason
-- Use `/fire-search` to find relevant skills BEFORE starting work (don't search mid-loop)
-- Read only the relevant sections of large files (use line offsets)
-- After 2 failed approaches, `/clear` and restart with the lessons learned
-
----
-
-## 7. Agent spawns fail
-
-**Symptoms:**
-- "Agent definition not found" error when running `/fire-3-execute` or `/fire-7-review`
-- Subagent starts but immediately errors out
-- SWARM mode fails to create team members
-
-**Likely Cause:**
-- Agent definition file missing from `agents/` directory
-- Agent file has incorrect frontmatter (missing `name:` or `description:`)
-- Agent's `tools:` list references unavailable tools
-
-**Fix Steps:**
-1. Check that all agent files exist: `ls ~/.claude/plugins/fire-flow/agents/`
-2. Expected files: `fire-executor.md`, `fire-planner.md`, `fire-researcher.md`, `fire-verifier.md`, `fire-reviewer.md`
-3. Verify each file has valid YAML frontmatter with `name:` and `description:`
-4. If a file is missing, run `/fire-update` to pull the latest plugin version
-
-**Prevention:**
-- Don't manually edit agent files unless you know the frontmatter schema
-- After updating the plugin, verify agents: `ls ~/.claude/plugins/fire-flow/agents/`
-- SWARM mode requires the experimental teams flag: `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`
-
----
-
-## 8. Skills not found
-
-**Symptoms:**
-- `/fire-search` returns no results for queries that should match
-- Planner agent says "no relevant skills found" during `/fire-2-plan`
-- Skills referenced in plans don't exist at the expected path
-
-**Likely Cause:**
-- Skills library path misconfigured or missing
-- Project-local skills out of sync with global library
-- Skill was deleted or moved to a different category
-
-**Fix Steps:**
-1. Check the skills library exists: `ls ~/.claude/plugins/fire-flow/skills-library/`
-2. Run `/fire-skills-sync --pull` to pull latest skills from global to project
-3. Run `/fire-skills-sync --dry-run` to see what's out of sync
-4. If a specific skill is missing, search by keyword: `/fire-search "{topic}"`
-
-**Prevention:**
-- Run `/fire-skills-sync --pull` at the start of each major session
-- After contributing a new skill, verify it appears: `/fire-search "{skill name}"`
-- Keep skills in their correct category directory -- don't move them manually
-
----
-
-## 9. Circuit breaker tripped
-
-**Symptoms:**
-- Loop stops with "CIRCUIT BREAK" banner
-- State shows SPINNING or STALLED or DEGRADED
-- Message: "Error seen N times. Previous approaches: {list}"
-
-**Likely Cause:**
-- **STALLED:** No file changes for 3+ iterations -- agent is reading/thinking but not acting
-- **SPINNING:** Same error hash repeated 4+ times -- agent retrying the same failed approach
-- **DEGRADED:** Output volume declining 50%+ from baseline -- context rot setting in
-
-**Fix Steps:**
-1. Read the loop state file: `.planning/loops/fire-loop-{ID}.md`
-2. Check the "Approaches Tried" section -- these are what NOT to repeat
-3. For STALLED: the fix is usually to try a fundamentally different approach
-4. For SPINNING: read the error carefully -- the repeated error hash tells you exactly what's failing
-5. For DEGRADED: accept the Sabbath Rest, `/clear`, then `/fire-loop-resume {ID}`
-
-**Prevention:**
-- Use `--aggressive` for tasks where you expect quick resolution (tighter thresholds)
-- Write clear completion criteria so the loop knows when it's done
-- Trust the circuit breaker -- forcing through a tripped breaker wastes iterations
-- After a trip, always read the loop file before resuming to understand what was tried
-
----
-
-## 10. Handoff incomplete
-
-**Symptoms:**
-- `/fire-6-resume` can't find a handoff file or shows stale data
-- Handoff missing key sections (blockers, in-progress work, skills applied)
-- Next session agent doesn't have enough context to continue
-
-**Likely Cause:**
-- `/fire-5-handoff` wasn't run before ending the session
-- Handoff was created but CONSCIENCE.md was out of date at the time
-- Session ended abruptly (crash, Ctrl+C) before handoff could complete
-
-**Fix Steps:**
-1. Run `/fire-5-handoff` manually to create a fresh handoff from current state
-2. Check `.planning/` for any `POWER-HANDOFF-*.md` files -- there may be a partial one
-3. If CONSCIENCE.md is stale, update it manually with current phase status
-4. Verify the handoff follows WARRIOR 7-step format: What / Accomplished / Remaining / Resources / Issues / Observations / Recommendations
-
-**Prevention:**
-- Always run `/fire-5-handoff` before ending a session -- make it a habit
-- The `stop-verify.js` hook warns if tasks are incomplete when you exit
-- For long sessions, create intermediate handoffs every 2-3 hours
-- If using `/fire-loop`, the Sabbath Rest snapshot serves as an automatic partial handoff
-
----
-
-*If your issue isn't listed here, check the references directory for detailed protocol documentation, or run `/fire-debug` to systematically investigate.*
+# Dominion Flow Troubleshooting Guide
+
+Common failure modes and their fixes.
+
+---
+
+## 1. Autonomous mode won't stop
+
+**Symptoms:**
+- `/fire-autonomous` or `/fire-loop` continues iterating past the point of usefulness
+- Agent ignores Sabbath Rest warnings and keeps working
+- Iteration count climbing with no meaningful progress
+
+**Likely Cause:**
+- `--max-iterations` set too high (default 50)
+- `--no-circuit-breaker` flag was passed, disabling the safety net
+- Completion promise text doesn't match actual output (e.g., "DONE" vs "Done")
+
+**Fix Steps:**
+1. Press `Ctrl+C` to interrupt the current iteration
+2. Run `/fire-loop-stop` to formally end the loop and save state
+3. Check `.planning/loops/fire-loop-*.md` for the loop state file
+4. Review the "Approaches Tried" section to see what was attempted
+
+**Prevention:**
+- Use `--max-iterations 20` for most tasks (30 max for complex work)
+- Never use `--no-circuit-breaker` unless you're monitoring closely
+- Write precise completion promises: `--completion-promise "ALL TESTS PASS"`
+- Use `--aggressive` for tighter stall/spin detection thresholds
+
+---
+
+## 2. Wrong files modified
+
+**Symptoms:**
+- Files in a different project were edited by a subagent
+- Changes landed in the wrong directory
+- Cross-project contamination (e.g., editing `my-other-project` instead of `my-project`)
+
+**Likely Cause:**
+- Working directory didn't match the project in CONSCIENCE.md
+- Subagent resolved a relative path to the wrong project
+- Multiple project directories open in the same session
+
+**Fix Steps:**
+1. Run `git diff` in the affected project to see what changed
+2. Run `git checkout -- .` to revert unwanted changes (or `git stash` to save them)
+3. Check CONSCIENCE.md -- does it reference the correct project path?
+4. Verify `pwd` matches the expected project root before re-running
+
+**Prevention:**
+- The Path Verification Gate (Step 3.5 in `/fire-3-execute`) catches this automatically
+- Always use absolute paths when working across multiple projects
+- Run `/fire-0-orient` at session start to confirm you're in the right place
+- When spawning subagents, the `<path_constraint>` block is injected automatically -- verify it matches
+
+---
+
+## 3. Memory features not working
+
+**Symptoms:**
+- Episodic memory injection (Step 7.1 in `/fire-loop`) returns no results
+- `/fire-reflect` can't search past reflections
+- "Connection refused" errors mentioning port 6335
+
+**Likely Cause:**
+- Qdrant is not running (native binary at `C:\path\to\qdrant\qdrant.exe`)
+- Qdrant is running but on the wrong port (6333 = Docker backup, 6335 = native primary)
+- Ollama embedding service is down (needed for `nomic-embed-text` 768d vectors)
+
+**Fix Steps:**
+1. Check Qdrant health: `curl http://localhost:6335/healthz`
+2. If down, start it: `C:\path\to\qdrant\qdrant.exe` (runs on port 6335)
+3. Check Ollama: `curl http://localhost:11434/api/tags`
+4. If Ollama is down: `ollama serve` in a separate terminal
+5. Verify the collection exists: `curl http://localhost:6335/collections/power_flow_memory`
+
+**Prevention:**
+- Dominion Flow degrades gracefully: if Qdrant is unreachable, it falls back to file-based search across `~/.claude/warrior-handoffs/` and `~/.claude/reflections/`
+- The file-based fallback is slower but functional -- memory retrieval is never silently skipped
+- Add Qdrant to your system startup if you use memory features regularly
+
+---
+
+## 4. Verification always fails
+
+**Symptoms:**
+- `/fire-4-verify` consistently returns REJECTED or CONDITIONAL
+- Score stuck below 49 even for seemingly complete work
+- E2E (Playwright) category drags the score down on non-UI projects
+
+**Likely Cause:**
+- The 70-point checklist includes 10 points for E2E/Playwright testing -- irrelevant for CLI, API-only, or library projects
+- Documentation category (10 points) penalizes when README/JSDoc isn't written yet
+- Validation thresholds: 63-70 = APPROVED, 56-62 = APPROVED*, 49-55 = CONDITIONAL, <42 = REJECTED
+
+**Fix Steps:**
+1. Check `.planning/phases/{N}-{name}/{N}-VERIFICATION.md` for the breakdown
+2. Identify which categories are scoring zero
+3. For non-UI projects: the E2E category should be redistributed to other categories by the verifier -- if it isn't, note this as a configuration issue
+4. Focus on the must-haves first: if all must-haves pass, CONDITIONAL is often acceptable for early phases
+
+**Prevention:**
+- Run `/fire-4-verify` with awareness of what the 7 categories are: Code Quality /10, Testing /10, Security /10, Performance /10, Documentation /10, Infrastructure /10, E2E /10
+- For API-only projects, tell the verifier in the plan: "No E2E -- redistribute points to Testing and Security"
+- Use `/fire-double-check` for a quick sanity check before the full 70-point verification
+
+---
+
+## 5. Plans executed in wrong order
+
+**Symptoms:**
+- Breath 2 tasks fail because Breath 1 outputs don't exist yet
+- Tasks with dependencies execute before their dependencies complete
+- File conflicts between parallel executors
+
+**Likely Cause:**
+- `breath:` frontmatter in BLUEPRINT.md files is incorrectly assigned
+- `depends_on:` field missing or wrong in the plan files
+- Execution mode selected PARALLEL when it should have been SEQUENTIAL (file overlap)
+
+**Fix Steps:**
+1. Open `.planning/phases/{N}-{name}/` and check each BLUEPRINT.md frontmatter
+2. Verify `breath:` numbers -- all tasks in Breath 1 should have no external dependencies
+3. Verify `depends_on:` lists reference only tasks in earlier breaths
+4. Re-run `/fire-2-plan {N}` to regenerate plans with correct breath assignment
+
+**Prevention:**
+- The planner agent validates breath assignments during `/fire-2-plan`
+- `/fire-3-execute` auto-detects file overlap and downgrades from PARALLEL to SEQUENTIAL
+- Review the execution manifest at Step 2 before proceeding
+
+---
+
+## 6. Context window filling up fast
+
+**Symptoms:**
+- Claude starts forgetting earlier instructions mid-session
+- Output quality degrades after 10-15 iterations
+- Sabbath Rest triggers repeatedly at low iteration counts
+
+**Likely Cause:**
+- Large files loaded into context (reading entire source files instead of relevant sections)
+- Skills library injecting too many skills per iteration (max should be 3)
+- No `.powerignore` file to exclude noise (node_modules, build artifacts, etc.)
+
+**Fix Steps:**
+1. Run `/compact` with a focus topic: `/compact Focus on phase 3 execution`
+2. Delegate research tasks to subagents instead of doing them in the main context
+3. Create a `.powerignore` file in the project root listing directories to exclude
+4. For `/fire-loop`: if iteration > 15, accept the Sabbath Rest and `/fire-loop-resume` in a fresh context
+
+**Prevention:**
+- Keep iteration output concise -- the recitation block is capped at 30 lines for a reason
+- Use `/fire-search` to find relevant skills BEFORE starting work (don't search mid-loop)
+- Read only the relevant sections of large files (use line offsets)
+- After 2 failed approaches, `/clear` and restart with the lessons learned
+
+---
+
+## 7. Agent spawns fail
+
+**Symptoms:**
+- "Agent definition not found" error when running `/fire-3-execute` or `/fire-7-review`
+- Subagent starts but immediately errors out
+- SWARM mode fails to create team members
+
+**Likely Cause:**
+- Agent definition file missing from `agents/` directory
+- Agent file has incorrect frontmatter (missing `name:` or `description:`)
+- Agent's `tools:` list references unavailable tools
+
+**Fix Steps:**
+1. Check that all agent files exist: `ls ~/.claude/plugins/dominion-flow/agents/`
+2. Expected files: `fire-executor.md`, `fire-planner.md`, `fire-researcher.md`, `fire-verifier.md`, `fire-reviewer.md`
+3. Verify each file has valid YAML frontmatter with `name:` and `description:`
+4. If a file is missing, run `/fire-update` to pull the latest plugin version
+
+**Prevention:**
+- Don't manually edit agent files unless you know the frontmatter schema
+- After updating the plugin, verify agents: `ls ~/.claude/plugins/dominion-flow/agents/`
+- SWARM mode requires the experimental teams flag: `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`
+
+---
+
+## 8. Skills not found
+
+**Symptoms:**
+- `/fire-search` returns no results for queries that should match
+- Planner agent says "no relevant skills found" during `/fire-2-plan`
+- Skills referenced in plans don't exist at the expected path
+
+**Likely Cause:**
+- Skills library path misconfigured or missing
+- Project-local skills out of sync with global library
+- Skill was deleted or moved to a different category
+
+**Fix Steps:**
+1. Check the skills library exists: `ls ~/.claude/plugins/dominion-flow/skills-library/`
+2. Run `/fire-skills-sync --pull` to pull latest skills from global to project
+3. Run `/fire-skills-sync --dry-run` to see what's out of sync
+4. If a specific skill is missing, search by keyword: `/fire-search "{topic}"`
+
+**Prevention:**
+- Run `/fire-skills-sync --pull` at the start of each major session
+- After contributing a new skill, verify it appears: `/fire-search "{skill name}"`
+- Keep skills in their correct category directory -- don't move them manually
+
+---
+
+## 9. Circuit breaker tripped
+
+**Symptoms:**
+- Loop stops with "CIRCUIT BREAK" banner
+- State shows SPINNING or STALLED or DEGRADED
+- Message: "Error seen N times. Previous approaches: {list}"
+
+**Likely Cause:**
+- **STALLED:** No file changes for 3+ iterations -- agent is reading/thinking but not acting
+- **SPINNING:** Same error hash repeated 4+ times -- agent retrying the same failed approach
+- **DEGRADED:** Output volume declining 50%+ from baseline -- context rot setting in
+
+**Fix Steps:**
+1. Read the loop state file: `.planning/loops/fire-loop-{ID}.md`
+2. Check the "Approaches Tried" section -- these are what NOT to repeat
+3. For STALLED: the fix is usually to try a fundamentally different approach
+4. For SPINNING: read the error carefully -- the repeated error hash tells you exactly what's failing
+5. For DEGRADED: accept the Sabbath Rest, `/clear`, then `/fire-loop-resume {ID}`
+
+**Prevention:**
+- Use `--aggressive` for tasks where you expect quick resolution (tighter thresholds)
+- Write clear completion criteria so the loop knows when it's done
+- Trust the circuit breaker -- forcing through a tripped breaker wastes iterations
+- After a trip, always read the loop file before resuming to understand what was tried
+
+---
+
+## 10. Handoff incomplete
+
+**Symptoms:**
+- `/fire-6-resume` can't find a handoff file or shows stale data
+- Handoff missing key sections (blockers, in-progress work, skills applied)
+- Next session agent doesn't have enough context to continue
+
+**Likely Cause:**
+- `/fire-5-handoff` wasn't run before ending the session
+- Handoff was created but CONSCIENCE.md was out of date at the time
+- Session ended abruptly (crash, Ctrl+C) before handoff could complete
+
+**Fix Steps:**
+1. Run `/fire-5-handoff` manually to create a fresh handoff from current state
+2. Check `.planning/` for any `POWER-HANDOFF-*.md` files -- there may be a partial one
+3. If CONSCIENCE.md is stale, update it manually with current phase status
+4. Verify the handoff follows WARRIOR 7-step format: What / Accomplished / Remaining / Resources / Issues / Observations / Recommendations
+
+**Prevention:**
+- Always run `/fire-5-handoff` before ending a session -- make it a habit
+- The `stop-verify.js` hook warns if tasks are incomplete when you exit
+- For long sessions, create intermediate handoffs every 2-3 hours
+- If using `/fire-loop`, the Sabbath Rest snapshot serves as an automatic partial handoff
+
+---
+
+*If your issue isn't listed here, check the references directory for detailed protocol documentation, or run `/fire-debug` to systematically investigate.*

package/agents/fire-debugger.md

@@ -293,6 +293,50 @@ For each hypothesis, starting with the most likely:
 - Consider combining partial hypotheses
 - If still stuck after 5 hypotheses, escalate with full evidence trail
 
+### Step 4.5: Red-Green Bugfix Gate
+
+Before applying the fix, decide whether this bug warrants a regression test.
+
+**Write a test if ANY of:**
+- 2+ hypotheses were eliminated (non-obvious bug)
+- The bug is in business logic, not config
+- The area has no existing test coverage
+- The bug could recur from future changes
+
+**Skip the test if ALL of:**
+- Fix is a one-line typo/config change
+- Existing tests already cover this path
+- Bug is environmental (ports, env vars, OS-specific)
+
+```markdown
+## Red-Green Gate Decision
+
+**Hypotheses eliminated:** [N]
+**Existing test coverage:** [yes/no — check with `grep -r "test.*[function/route]"` or `glob **/[area]*.test.*`]
+**Could recur:** [yes/no]
+
+**Decision:** WRITE TEST | SKIP (reason: [reason])
+```
+
+**If writing a test:**
+1. Write the failing test FIRST (before any code changes)
+2. Run it — MUST FAIL (RED). If it passes, the test doesn't capture the bug.
+3. Apply the fix
+4. Run the same test — MUST PASS (GREEN)
+5. Continue to Step 6 (verify) and Step 7 (regression check)
+
+**Test type selection:**
+| Bug Location | Test Type | Tool |
+|-------------|-----------|------|
+| Pure function / service | Unit test | Vitest |
+| API endpoint | Integration test | Vitest + supertest |
+| UI interaction / render | E2E test | Playwright |
+| Database / ORM | Integration test | Vitest + test DB |
+
+**Test naming:** `{bug-slug}.regression.test.ts` or add a test case to the nearest existing spec.
+
+> Skill reference: @skills-library/testing/RED_GREEN_BUGFIX_GATE.md
+
 ### Step 5: Apply Fix
 
 ```markdown
@@ -414,6 +458,13 @@ files_modified:
 **Cause:** [Detailed explanation of what went wrong]
 **Why it happened:** [Deeper analysis — what allowed this to occur]
 
+## Red-Green Gate
+**Decision:** [WRITE TEST | SKIP]
+**Reason:** [Why test was written or skipped]
+**Test file:** [path/to/test or N/A]
+**RED confirmed:** [yes — test fails before fix | N/A]
+**GREEN confirmed:** [yes — test passes after fix | N/A]
+
 ## Fix Applied
 **Strategy:** [What the fix does]
 **Files:** [List of modified files with line references]
@@ -466,7 +517,10 @@ files_modified:
 - [ ] Each hypothesis tested with targeted evidence
 - [ ] Hypotheses eliminated with documented reasoning
 - [ ] Root cause confirmed (not assumed)
+- [ ] Red-Green Gate evaluated (write test or document skip reason)
+- [ ] If test written: RED confirmed (test fails before fix)
 - [ ] Fix applied addressing root cause (not symptom)
+- [ ] If test written: GREEN confirmed (test passes after fix)
 - [ ] Fix verified against original reproduction
 - [ ] Regression check completed (test suite passes)
 - [ ] Debug session file created in .planning/debug/