gza-task-fix

name: gza-task-fix description: Rescue a stuck gza task inline — diagnoses review/improve churn, verifies each blocker against current code, addresses the open ones, runs verify, and commits. Runs entirely in Claude Code, no background worker. allowed-tools: Read, Edit, Write, Glob, Grep, Bash(uv run:), Bash(git:), Bash(mkdir:), Bash(ls:), Bash(cd:*), AskUserQuestion version: 2.0.1 public: true

Fix Stuck Gza Task Inline

Use this skill when an implementation task is stuck in review/improve churn — the same blockers keep reappearing, or a previous improve/fix pass failed to close them. fix is escalation: it diagnoses why the loop is happening before making any edits, then applies a bounded repair scoped strictly to blocker closure.

Unlike /gza-task-improve, this skill requires you to verify each blocker against the current code before deciding whether a change is needed. A stuck task often already has the fix on disk — in which case the answer is "no change, this was hallucinated-closure churn," not another edit pass.

This skill runs entirely inline in the current Claude Code session. Do not invoke gza fix or any background worker — that defeats the purpose of running here.

Process

Step 0: Capture the starting checkout

git symbolic-ref --quiet --short HEAD || git rev-parse --short HEAD

Save as <START_CHECKOUT>. You must return the user here before finishing. If detached, restore with git checkout --detach <START_CHECKOUT>.

Step 1: Resolve the target task and recent review history

The user provides a full prefixed task ID (for example, gza-1234) — implementation, review, improve, or prior fix. Resolve to the implementation task and fetch the last three reviews so you can detect churn:

uv run python -c "
import json, sys
from pathlib import Path
from gza.config import Config
from gza.db import SqliteTaskStore

config = Config.load(Path.cwd())
store = SqliteTaskStore.from_config(config)
task = store.get('<TASK_ID>')
if not task:
    print('ERROR: task not found', file=sys.stderr); sys.exit(1)

impl = task
if task.task_type in ('review', 'fix') and task.depends_on:
    impl = store.get(task.depends_on)
    if impl and impl.task_type == 'review' and impl.depends_on:
        impl = store.get(impl.depends_on)
elif task.task_type == 'improve' and task.based_on:
    impl = store.get(task.based_on)

reviews = store.get_reviews_for_task(impl.id)[:3] if impl else []
print(json.dumps({
    'impl_id': impl.id if impl else None,
    'impl_branch': impl.branch if impl else None,
    'impl_prompt': impl.prompt if impl else None,
    'impl_tags': list(impl.tags) if impl else [],
    'verify_command': config.verify_command,
    'inner_verify_command': config.inner_verify_command,
    'reviews': [
        {'id': r.id, 'report_file': r.report_file, 'output_content': r.output_content, 'completed_at': str(r.completed_at)}
        for r in reviews
    ],
}, default=str))
"

The latest review is reviews[0]; prior reviews are reviews[1:]. If there are no reviews, there is no blocker set — stop and ask the user what they want rescued.

Step 2: Diagnose the loop

Read the latest review's report file (fall back to output_content if the file is missing). Extract each Must-Fix item's file:line evidence and a short blocker key summarizing the topic (e.g., improve-template-review-only-language).

Then read the prior 1–2 reviews and look for repeated blocker keys or overlapping evidence. Classify the overall situation:

• stale_review — the latest review's claims don't match the current code (review is out of date).
• prior_fix_missed — same blocker key appears across 2+ reviews and earlier improve/fix commits didn't close it.
• scope_creep — review flags things not in the original task prompt.
• genuine_open — the blockers are valid and uncorrected.

Present a one-paragraph loop-diagnosis summary to the user. If the situation is stale_review or scope_creep, confirm with the user before making any changes — the right outcome may be zero code changes plus an explicit diagnosed_no_change handoff.

Step 3: Check out the implementation branch

git checkout <impl_branch>

If the branch is checked out in another worktree, default to working in that existing worktree path and continue there without asking. Treat this as an interactive task rescue: prefer the active task worktree over switching the user's current checkout. Ask only if the existing worktree path is unavailable or there are multiple plausible task worktrees and the right target is ambiguous. Do not create duplicate worktrees.

If <START_CHECKOUT> already equals <impl_branch>, do not switch away and back.

Step 4: Verify each blocker against the current code before editing

For each Must-Fix item in the latest review:

1. Read the cited file:line location with the Read tool. The review may be wrong about the current state.
2. Classify the blocker:
   • already_addressed — code already satisfies the review's "Required fix." Record a file:line citation as evidence. No change.
   • open — code does not match. Plan the minimum change needed.
   • ambiguous — state unclear or genuinely contested. Stop and ask the user.
3. For open items only, make the change strictly as described in the review's "Required fix." Do not rename, refactor, or reorganize anything outside the blocker.
4. Add or update the targeted regression test the review names. Do not add extra tests.

Known recipe — installed skill out of sync after src/gza/skills/ edit. If the blocker is a test failure where .claude/skills/<name>/SKILL.md differs from src/gza/skills/<name>/SKILL.md (e.g. test_merge_first_docs_and_fix_skill_schema_stay_in_sync or any *sync* test comparing the two paths), treat it as a source-vs-install-state diagnosis, not an automatic code edit. The bundled source under src/gza/skills/ is the only committable source of truth; .claude/skills/ is a gitignored install artifact. First verify whether the bundled source already satisfies the requested behavior. If it does, refresh the local installed copy only as verification/setup:

uv run gza skills-install --update --project .

Then re-run verify. Do not edit or stage .claude/skills/ by hand — it is local install state, not source. If the only remaining mismatch is the installed copy, classify it as an environment/refresh issue rather than a review blocker. Only change code when the bundled source, installer behavior, or a test asserting the supported refresh path is actually wrong.

If every blocker classifies as already_addressed, do not make changes. Skip to Step 7 with fix_result: diagnosed_no_change.

Step 5: Use fast checks while editing, then run the final gate

If inner_verify_command is configured, you may use it during Step 4 after a blocker-specific fix batch. Otherwise use targeted tests for touched files. Do not keep relaunching the full suite after each individual blocker edit.

After the last planned change, if verify_command is configured, run it once and time the full wall-clock duration with a monotonic clock. Record verify.duration_seconds whenever the full command actually runs. Fix verification fallout only in files you already touched in Step 4 — do not broaden scope. Retry the full gate up to 3 iterations. If still failing, stop and report to the user with fix_result: needs_user.

Step 6: Commit and push (only if changes were made)

git add <changed_files>
git commit -m "Rescue task #<IMPL_ID> (review #<LATEST_REVIEW_ID>)

- <blocker_key_1>: <brief fix>
- <blocker_key_2>: <brief fix>
"
git push -u origin <impl_branch>

If the push fails, stop and report. Do not treat the rescue as complete when changes are only local.

Step 7: Emit the blocker ledger and persist a fix task row

Write the YAML ledger below, then persist a completed fix task row whose summary carries the ledger. Clear review state only if changes were committed.

Ledger format (preserve this schema — downstream tooling reads it):

fix_result: repaired_pending_review | diagnosed_no_change | needs_user | blocked_external
verify:
  command: <configured verify command or "">
  passed: true | false | null
  duration_seconds: <float or null>
  autonomous_verify_timeout_seconds: <int>
  review_verify_timeout_grace_seconds: <number >= 1>
blockers:
  - source_review_id: <review_task_id>
    blocker_key: <short_key>
    summary: <one-line description>
    status: addressed | already_addressed | deferred_to_user | out_of_scope
    closure_evidence: <file:line or commit sha proving the blocker is closed>
    verify_evidence: <verify command + brief outcome, or "n/a" for no-change passes>
    follow_up_review_required: true | false

Ledger integrity rules:

• verify.duration_seconds is required whenever verify_command ran.
• status: addressed requires a commit in this pass. Cite the commit sha in closure_evidence.
• status: already_addressed requires a file:line citation to current code that satisfies the review. No commit.
• follow_up_review_required: true is permitted only when fix_result: repaired_pending_review and commits were made. Never request a review for a no-op pass.

Persist the fix task (mirrors how /gza-task-improve persists its row):

uv run python -c "
from datetime import datetime, timezone
from pathlib import Path
from gza.config import Config
from gza.db import SqliteTaskStore
from gza.git import Git
from gza.runner import _compute_slug_override, generate_slug, get_task_output_paths

config = Config.load(Path.cwd())
store = SqliteTaskStore.from_config(config)

ledger_body = '''<LEDGER_YAML_FROM_ABOVE>'''
origin = datetime.now(timezone.utc).strftime('%Y-%m-%d')
summary_with_origin = f'<!-- origin: /gza-task-fix (manual, {origin}) -->\n' + ledger_body

created = store.add(
    prompt='Manual rescue via /gza-task-fix',
    task_type='fix',
    depends_on='<LATEST_REVIEW_ID_OR_NONE>',
    based_on='<IMPL_ID>',
    same_branch=True,
    tags=<IMPL_TAGS>,
)
assert created.id is not None

try:
    if created.slug is None:
        slug_override = _compute_slug_override(created, store)
        created.slug = generate_slug(
            created.prompt, existing_id=None,
            log_path=config.log_path,
            git=Git(config.project_dir),
            store=store,
            exclude_task_id=created.id,
            project_name=config.project_name,
            project_prefix=config.project_prefix,
            slug_override=slug_override,
            branch_strategy=config.branch_strategy,
            explicit_type=created.task_type_hint,
        )
        store.update(created)

    _report_path, summary_path = get_task_output_paths(created, config.project_dir)
    assert summary_path is not None
    summary_path.parent.mkdir(parents=True, exist_ok=True)
    summary_path.write_text(summary_with_origin)

    created.report_file = str(summary_path.relative_to(config.project_dir))
    created.status = 'completed'
    created.completed_at = datetime.now(timezone.utc)
    created.output_content = ledger_body
    store.update(created)
except Exception:
    created.status = 'dropped'
    store.update(created)
    raise

if <CHANGES_WERE_COMMITTED>:
    store.clear_review_state('<IMPL_ID>')
print(f'Fix saved as task #{created.id} ({created.report_file})')
"

Step 8: Restore the starting checkout

git checkout <START_CHECKOUT>

If detached, use git checkout --detach <START_CHECKOUT>.

If the restore fails, stop and tell the user exactly what checkout you left them on. Do not silently finish on the implementation branch.

Important notes

• Diagnose before editing. fix is escalation. Skipping to edits the way improve does reproduces the same loop.
• Verify every blocker against the current code. The failure mode we rescue from is hallucinated closure on one side — and stale-review false blockers on the other. Both are common; a Read is cheap.
• Bounded scope. Only touch files and tests named by the current review's Must-Fix items. No opportunistic cleanup, no drive-by refactors, no renames.
• Ask the user before broadening scope, deciding an ambiguous blocker, or committing to a stale_review / scope_creep diagnosis.
• No review request without a commit. follow_up_review_required: true is only valid when code changed in this pass. A no-op pass is an operational finding, not work for a reviewer.
• Restore the user's checkout before exit. Final state returns to <START_CHECKOUT>; closing message should name both checkouts explicitly.
• Role separation. review is the independent approval boundary. improve is the normal response to one review. fix is escalation for churn.