# Execute-Plan Anti-Patterns Common failure modes during plan execution, with recovery moves. ## 1. Marking `[x]` with failing tests **Symptom.** A task is checked off but `{{TEST_CMD}}` reports a regression. **Why it's bad.** The plan's contract is "every `[x]` is green." Breaking that breaks every downstream signal — reviewers, automation, the next agent session. **Recovery.** Revert the checkbox flip with `git revert ` (if pushed) or `git reset HEAD~1` (if not). Fix the code. Re-commit with the checkbox in the same diff. --- ## 2. Batching tasks in one commit **Symptom.** `git log` shows one commit with two `[x]` flips and code for two sub-IDs. **Why it's bad.** Loses traceability. Bisect can't isolate which task broke things. **Recovery.** - Not pushed: `git reset HEAD~1`, stage and commit each task's diff separately. - Pushed: surface to user. Either accept the bundle (note in PR body: "FR-001.2 and FR-001.3 ship together — they share a refactor") or open a fixup with the user's approval. Prevention: **return control to the user between every task.** Phase 2's hand-back is the gate. --- ## 3. Auto-chaining tasks **Symptom.** Agent finishes FR-001.1 and immediately starts FR-001.2 without asking. **Why it's bad.** The user is the human gate. Auto-chaining defeats Phase 2's per-task review point. **Recovery.** None needed if no harm done — just stop and re-engage the user. If a task introduced bad code that the next task built on, surface the chain so the user can decide rollback strategy. Prevention: re-read [`prompts/02-task-loop.md`](../prompts/02-task-loop.md) Step 7 — *"Wait for the user's call before starting the next task."* --- ## 4. Three-strike loop **Symptom.** Same test fails 3 times. Agent keeps retrying with small variations. **Why it's bad.** Loops waste tokens and rarely break out of a misunderstanding. **Recovery.** STOP. Tell the user: > *"FR-001.2 has failed 3 times — pattern: [describe the failure]. Suspected cause: [hypothesis]. Options: A) ask user to clarify AC, B) update the plan, C) skip this task and revisit."* Let the user decide. --- ## 5. Silent scope creep **Symptom.** Diff includes files the plan didn't list. Agent "noticed" a related bug and fixed it. **Why it's bad.** Mixes plan-implementation with surprise refactoring. Reviewer can't tell which change satisfies which AC. **Recovery.** - Not pushed: amend the commit to remove the unrelated change. File a separate issue or note. - Pushed: surface to the user. Common resolution: open a follow-up PR for the surprise fix, link it from the main PR. Prevention: stay inside the file list. Trivial one-liners (an import, a type widening) are OK if noted in the commit body. --- ## 6. `--no-verify` to bypass a hook **Symptom.** Pre-commit hook fails (lint, format, test). Agent commits with `--no-verify` to "keep moving." **Why it's bad.** Hooks exist for reasons the agent doesn't always know. Bypassing them ships broken code. **Recovery.** Revert. Investigate why the hook failed. Fix the cause. Re-commit through the hook. The framework's rule: **never `--no-verify` without explicit user approval.** And even with approval, surface what was bypassed in the commit body. --- ## 7. Mid-task ambiguity → invent **Symptom.** Implementing FR-001.2, agent realizes the AC doesn't say what HTTP status to return on rate-limit. Agent picks `429` and moves on. **Why it's bad.** The user is the source of truth, not the agent's defaults. A hidden decision in code is worse than a flagged decision in the plan. **Recovery.** STOP. Surface the ambiguity: > *"AC-001.5 says 'rate limited' but doesn't specify status code. Standard would be 429 with `Retry-After`. Confirm or change?"* The framework's rule: **ambiguity → ASK.** Even mid-task. --- ## 8. Skipping the design read **Symptom.** UI task implements something that doesn't match `docs/designs/FR-001-login.html`. **Why it's bad.** Design rework after implementation costs more than design read before. **Recovery.** Revert the UI work. Re-open the design HTML. Re-implement to match. Prevention: Phase 1 explicitly requires opening every `docs/designs/-*.html`. Don't skip. --- ## 9. Plan checkbox in a separate commit **Symptom.** Code commit `feat(auth): add lastLoginAt to user model`, then a follow-up `chore: mark FR-001.1 done`. **Why it's bad.** History no longer shows code and plan in agreement at every commit. Bisect and PR review get confusing. **Recovery.** If not pushed: `git reset HEAD~2`, re-stage including the plan, re-commit as one. If pushed: surface to user. The framework's rule: **mark `[x]` in the same commit as the work.** --- ## 10. Closing the loop without finalize **Symptom.** All tasks `[x]`, agent stops there. No PR description, no CLAUDE.md feedback. **Why it's bad.** The user is left to write the PR body and lose the lessons-learned signal. **Recovery.** Run Phase 3 ([`prompts/03-finalize.md`](../prompts/03-finalize.md)). Score the checklist. Produce the PR body. Surface any recurring corrections. Prevention: don't declare the skill done until Phase 3 completes.