/implement-fix

name: implement-fix
description: Minimal pipeline for test-fixing tasks
version: 1.0.0
user-invocable: false
user_facing: false
allowed-tools: [Agent, Read, Write, Edit, Bash, Grep, Glob, mcp__searxng__search, mcp__searxng__fetch, WebSearch]

FIX MODE

> The key words "MUST", "MUST NOT", "SHOULD", and "MAY" in this document are to be interpreted as described in [RFC 2119](https://www.rfc-editor.org/rfc/rfc2119).

Minimal pipeline (5 steps, 4 agents minimum) for test-fixing tasks. Invoked via `/implement --fix "description"`. Skips research and planning since the problem is already known (failing tests).

Coordinator Role — HARD GATE

The coordinator is a **dispatcher**, not a substitute for specialist agents. These constraints apply globally — before any step definitions — and MUST NOT be violated regardless of circumstances.

Agent Management

The coordinator dispatches work to specialists; it MUST NOT do the work itself.

**FORBIDDEN — You MUST NOT do any of the following**:

• ❌ You MUST NOT write, edit, or modify any project files directly — ALL file modifications MUST go through specialist agents (implementer, doc-master)
• ❌ You MUST NOT do an agent's work when the agent crashes — RETRY once, then BLOCK with a clear error message; never substitute coordinator judgment for agent execution
• ❌ You MUST NOT skip any STEP in the pipeline
• ❌ You MUST NOT summarize agent output instead of passing full results to the next agent — verbatim output is required

Pipeline Integrity

Step ordering and post-completion behavior are strictly constrained.

**FORBIDDEN — You MUST NOT do any of the following**:

• ❌ You MUST NOT parallelize agents from different pipeline phases (e.g., running F3 and F4 concurrently)
• ❌ You MUST NOT clean up pipeline state before STEP F5 launches
• ❌ You MUST NOT perform any file edits after agents complete — the coordinator's only permitted post-agent actions are: outputting the final summary, git operations (add, commit, push), and launching STEP F5
• ❌ You MUST NOT skip STEP F4.7 when the changeset matches the activation trigger (SQL, ORM models, execution/order/trade/payment paths) — even if the unit tests pass green. The 4-fix chain #1285 → #1288 happened with all-green unit tests. (#1210)

Output Fidelity

The coordinator MUST transmit agent outputs intact.

**FORBIDDEN — You MUST NOT do any of the following**:

• ❌ You MUST NOT paraphrase or condense agent output when passing to the next stage
• ❌ You MUST NOT pass fewer than 50% of the implementer output words to the reviewer — if context is a constraint, include the implementer's file change list and test results in full before trimming any prose sections

Fix Mode Progress Protocol

Output structured progress at each step. Same format conventions as the full pipeline protocol.

**Step Banner**: `STEP FN/5 — Step Name` with agent info where applicable. **Timing**: Capture `FIX_START=$(date +%s)` at pipeline start. Capture per-step timing. **Gate Results**: `GATE: name — PASS/BLOCKED` **Final Summary** (after STEP F6 persists the CIA report, before STEP F6.5 cleanup):

========================================
FIX COMPLETE
========================================
Step  Description         Agent(s)                           Time    Status
----  ------------------  ---------------------------------  ------  ------
F1    Alignment           —                                  2s      PASS
F1.5  Pre-staged check    —                                  1s      PASS
F2    Test context        —                                  5s      done
F3    Implementation      implementer (Opus)                 2:30    done
F3    Test gate           —                                  8s      PASS
F4    Review + docs       reviewer, doc-master               45s     done
F5    CI analysis         continuous-improvement-analyst     30s     done
F6    Persist CIA report  — (coordinator Write, #1209)       1s      done
========================================
Total: 4:02 | Tests: N passed, M failed | Files changed: N | CIA: .claude/local/cia-DATE-issue-NNNN-fix.md
========================================

Implementation

Steps F1-F2: Alignment and Test Context

STEP F1: Validate PROJECT.md Alignment

**Progress**: Output step banner (STEP F1/5 — Alignment). Capture FIX_START. Output gate result.

Read `.claude/PROJECT.md`. If missing: BLOCK ("Run `/setup` or `/align --retrofit`").

Pipeline State Initialization (Before Alignment Verdict)

Initialize the fix-mode pipeline state file BEFORE running the alignment gate protocol below, so that `record_alignment_verdict` (Issue #1467) writes `alignment_passed` and `alignment_verdict` into an existing state file. This also ensures hook enforcement (prompt integrity, pipeline ordering) is active during fix mode:

python3 -c "
import json, os, time
state = {'mode': 'fix', 'explicitly_invoked': True, 'start_time': int(time.time())}
with open(os.environ.get('PIPELINE_STATE_FILE', '/tmp/implement_pipeline_state.json'), 'w') as f:
    json.dump(state, f)
print('Pipeline state initialized for fix mode')
"

This ensures prompt integrity enforcement (Layer 5) can detect an active pipeline and apply baseline shrinkage checks in addition to the minimum word count gate.

Run the STEP 2 alignment gate protocol from implement.md (Stage 0 → alignment-classifier dispatch → record_alignment_verdict → verdict routing) using the fix description as the feature text. Initialize the fix-mode pipeline state BEFORE the verdict step so record_alignment_verdict writes alignment_passed and alignment_verdict into it.

This is the same alignment gate as the full pipeline STEP 1.

Prompt Baseline Reset (Defensive — Issue #1088 F3)

Before initializing pipeline state, clear any stale `prompt_baselines.json` from a prior session. /fix mode by design dispatches shorter, focused prompts; stale baselines from prior runs frequently exceed the 20% shrinkage threshol