self-correction-loop

Self-Correction Loop

When your edit produces a warning/error signal, you're in a correction loop. This skill is the discipline that keeps the loop productive instead of thrashing.

Trigger signals

You should engage this loop when ANY of these happen after your edit:

• PostToolUse hook injected clang-tidy warnings: / lint warnings: / similar
• A Bash command you ran returned non-zero exit
• A test runner reports failures
• A type-checker reports errors
• The user says "that broke X" / "it's failing now"

The Loop (RED → diagnose → minimal-fix → GREEN)

This is TDD's RED-GREEN-REFACTOR adapted for correction:

1. CAPTURE the signal
   - What is the EXACT error message?
   - Which file / line?
   - What was my last change?

2. DIAGNOSE root cause (not symptom)
   - Read the error carefully — the answer is often in it
   - Check if my change broke an invariant elsewhere
   - Do NOT guess. If unclear, narrow with prints / smaller test cases

3. MINIMAL FIX
   - Smallest possible change that resolves THE specific issue
   - No "while I'm here" tangents (that's scope creep — see scope-jump)
   - Do NOT mask the error (e.g. catch-and-ignore) without explicit reason

4. RE-VERIFY
   - Run the same command that produced the signal
   - Observe the new output
   - Capture it as evidence

Loop budget

This loop is for surface signals that a quick minimal fix resolves. If a minimal fix doesn't clear the signal within 1–2 iterations — or the failure spans components / needs backward root-cause tracing — switch to superpowers:systematic-debugging (its 4-phase root-cause process). Don't keep thrashing minimal fixes here.

If you've iterated 3 times on the same signal without resolving it, stop and escalate:

• Switch to superpowers:systematic-debugging for structured root-cause investigation (its budget also caps at 3 fixes → then question the architecture; the two thresholds are intentionally aligned)
• Read more of the surrounding code than you have so far
• Invoke docs-lookup skill for the specific error class (Context7 → WebSearch fallback)
• Tell the user: "I've tried 3 approaches to <issue>. Latest output: <...>. I need more context. Recommend either <option A> or <option B>?"

Looping > 5 times is almost always a sign of:

• Wrong mental model of the bug → step out, re-read
• Missing tool/dependency → install it before continuing
• Genuinely ambiguous case → consult the user

Two budgets: re-run vs. edit

The budget above is the re-run budget — re-running the same failing command. There's a second, sneakier loop: the edit budget.

• Edit budget (doom-loop): if you've made ~3–5 edits to the same file chasing the same goal without the signal clearing, you're micro-adjusting one approach. STOP. Re-read the original spec / done_criteria — not your own diff (the classic failure is re-reading your own code, deciding it "looks ok", and continuing). Then either switch to a structurally different approach or escalate to the user.
• Don't "fix" a failing test by rewriting the test to match your code — that silences the signal. Treat tests as a reserved space: change the implementation, or question the requirement, not the test.

This mirrors LoopDetection middleware ("reconsider the approach after N edits on one file") — break out before the 10th micro-adjustment, not after.

What "fixed" looks like

• The original failing command now succeeds (exit 0 + expected output)
• You can articulate WHY it failed and WHY the fix resolves it
• No new failures introduced (re-run the broader test suite, not just the one test)

If you can't articulate WHY, your fix is suspicious — even if green. State this uncertainty per anti-hallucination-gates.

Before moving on: if the failure you just fixed is a recurrence (second+ time you've seen the class), capture it as a golden rule in this same turn — a rough stub via capturing-golden-rules — not as a mental note for session end. Deferred captures get distorted or forgotten when context compacts.

Anti-patterns

• Silencing instead of fixing: // @ts-ignore, --no-verify, deleting failing tests, broadening exception catches. These hide problems; sometimes they're warranted but require explicit user note.
• Chasing different bug: a lint warning on line 42 doesn't justify refactoring lines 100-200.
• Optimistic re-run: running the same failing test again "in case it was flaky" without changing anything. Flakiness is a separate issue to track, not a fix.

How this interacts with hooks

harness-anchor hooks are warn-only. They inject additionalContext describing the problem; they don't block your action. That means:

• You can ignore a warning, but the warning is recorded in your context — the user can see you saw it. If you ignore, do so deliberately and say why.
• A test failure injected by PostToolUse should generally be addressed before moving on; if you choose to defer it (e.g., feature is mid-build and the test is for the next milestone), note it as a known-failing in session-handoff.md.

When this skill is NOT applicable

• The signal is about a different feature than the one you're working on. Note it as planned in feature_list.json and continue current scope.
• The error is in third-party code you don't own. Document it; route to the right project.

Triaging fresh-context sensor findings

A fresh-context sensor's report — /verify (verification-runner), /test-plan (coverage-analyst), /gc (drift-analyst) — is reviewer feedback from a deliberately context-blind evaluator: it cannot see your intent (that independence is the design). Triage it with superpowers:receiving-code-review rigor — verify each finding against the code, push back on false positives with technical reasoning, act one-at-a-time — rather than blind-applying. This matters most inside /verify --fix: the bounded loop should apply verified findings, not churn on every recommendation. This is not a licence to dismiss the sensor — receiving-code-review is rigorous; you verify, then act.

Related

• anti-hallucination-gates — keep claims honest while looping
• capturing-golden-rules — a recurring failure you just fixed becomes a rule in the same turn
• receiving-code-review (superpowers) — the discipline for triaging a fresh-context sensor's findings (verify each, push back with reasoning, don't blind-apply)
• systematic-debugging (superpowers) — for deep debugging beyond surface signals
• verification-runner (agent) — for fresh-eyes evaluation of "is this really fixed?"