| name | forge-fix |
| description | Fix rejected Forge issues based on review or QA feedback. Use this skill when an issue has been reopened with rejection comments — reads the feedback, applies a scoped fix, builds, re-tests, and pushes. Triggers on: /forge-fix, fixing rejected issues, addressing review feedback, fixing QA failures, resolving reopen comments, fixing CI build failures. Also use when the pipeline needs to move an issue from reopen back to developed (or testing for xs/s). |
| user_invocable | true |
| arguments | documentId |
Forge Fix
Applies scoped fixes based on review or QA rejection feedback. This is NOT a full reimplementation — it reads what failed, fixes only that, and re-submits.
The key discipline: fix what the feedback says, nothing more. Expanding scope during a fix cycle leads to new bugs and infinite review loops.
Usage
/forge-fix <documentId>
Tools
- forge_issues — get issue data, update status
- forge_comments — read rejection feedback, post fix summary
- forge_coolify_deploy — trigger deployment after push
- Codebase tools — Read, Edit, Write, Glob, Grep, Bash
Read references/fix-workflow.md for parsing rejection formats, branch handling, and fix strategy details.
Workflow
Pull-model note: On large issues forge_step_start returns a lean manifest (bodyTruncated:true). Fetch plan/description/acceptanceCriteria as needed via forge_issues.get { documentId, fields: ['plan'] } rather than assuming full body is present.
Step 1: Fetch Issue & Feedback
forge_issues → get → { documentId: "<id>" }
forge_comments → list → { filters: { issue: "<documentId>" } }
Verify status is reopen. Find the latest rejection comment — either:
- Code Review (starts with
## Code Review) — from forge-review, has severity table
- QA Test Report (starts with
**QA Test Report**) — from forge-test, has pass/fail table
If feedback is unclear or missing → post clarifying comment, set needs_info, stop.
Step 2: Understand the Feedback
Parse the rejection:
- From review: extract Bug and Minor severity findings. Ignore Low items.
- From QA: extract FAIL test cases and failure descriptions.
Each finding = one fix task. Don't invent additional fixes.
Step 3: Confirm Branch & Check Out
First confirm current git state — in pipeline-resumed sessions the working tree may be on an unexpected branch:
git branch --show-current
git status
If dirty, stash or clean before switching. Then check out the ISS-* branch. The feature branch is always kept alive through the pipeline (forge-release merges it to production at the end):
git checkout ISS-XX-*
git pull origin ISS-XX-*
If not found locally, fetch and check out:
git fetch origin
git branch --list 'ISS-*' | grep <issue-number>
git checkout ISS-XX-short-title
Step 4: Apply Scoped Fixes
Work one finding at a time. Stop-the-line: finish the current finding before touching the next — a half-fixed earlier finding makes the later ones unreliable.
For a bug / QA-FAIL finding, fix the root cause, not the symptom:
- Reproduce — make the failure actually happen (run the failing case / hit the endpoint / load the screen). If you cannot reproduce it, say so in the fix comment instead of guessing — roughly a third of "I already know what it is" guesses are wrong.
- Localize — find which layer truly fails (UI / API / data / build), reading the file at the cited line.
- Fix the root cause — address the underlying issue, not just the visible symptom.
- Guard against recurrence (Beyonce Rule) — if the project has a test surface for that layer, add or adjust a test that fails without your fix and passes with it, so the bug can't silently return. A bug fix with no guarding test is an incomplete fix; if there is genuinely no harness for this layer, note that in the comment.
For a Minor / pattern finding, just correct the specific pattern at the cited line.
Then simplify only the lines you just changed (preserve behavior, stay inside your diff) — do not expand into adjacent code.
Do not: refactor adjacent code, add new features, "improve" things the feedback didn't mention, or change the overall approach. The plan was already approved — the fix stays within its boundaries.
Step 5: Test
Re-run the Step 4 reproduction first — confirm the failure is gone (regression test green / endpoint returns expected / screen behaves). Then: if the fix touches API endpoints, run real API tests (curl affected endpoints, verify responses); frontend-only fixes need only a successful build — QA handles frontend testing.
Step 6: Commit & Push
git add <specific files>
git commit -m "fix: address review feedback — <summary>"
git push
Separate fix commit — don't amend or squash into the original.
Push the ISS-* branch.
Deploy mode detection: Call forge_config → get and forge_coolify_deploy → list. If previewDeploy is null/missing AND Coolify list is empty → local-only mode. Otherwise → deploy mode.
Local-only mode (no Coolify, no preview URL):
git push origin ISS-XX-short-title
No baseBranch merge. No Coolify deploy. Stop here for push.
Deploy mode — all complexities (push the fix, merge into baseBranch = staging, then deploy in Step 7 — same as the code step, so the re-verify environment has the fix):
git push origin ISS-XX-short-title
git checkout <baseBranch> && git merge ISS-XX-short-title && git push origin <baseBranch>
git checkout ISS-XX-short-title
(Decompose child/parent target the integration branch instead of baseBranch — see .claude/skills/forge-plan/references/decompose-execution.md.)
Step 7: Deploy
Local-only mode — skip this step entirely. No environment to deploy to.
Deploy mode — trigger Coolify deployment after push so the environment is updated before the pipeline advances:
forge_coolify_deploy → deploy → { issueId: <current issue documentId> }
If no Coolify resources are configured within deploy mode, skip.
Step 8: Post Comment & Set Status
Status update must be the LAST action. It triggers downstream pipeline steps, so all work (push, deploy, comment) must complete first.
forge_comments → create → {
data: {
body: "**Fix** — <what was fixed>\n\nAddressed N findings from <review/QA>.",
issue: "<documentId>",
author: "Blastoise"
}
}
Set status — never deploying (retired; from reopen the valid exits are developed, testing, in_progress). Mirror the code step so the fix is re-verified the same way the original change was:
Deploy mode:
xs / s → testing (set staging previewUrl/previewApiUrl) — re-QA the fix directly on the deployed staging build; the inline self-review covered it.m / l / xl → developed — the independent forge-review re-checks the fix, then it advances to testing.
Local-only mode — developed for all complexities (human re-checks at developed, closes manually).
Fix-specific output reminder
Don't restate the review/CI feedback you parsed — just fix. (See pipeline preamble for general output rules.)
