3 Strikes Protocol

When the same bug is reported 3+ times with different fixes, STOP.

When to Use

• Same test failing across 3+ consecutive commits (mechanical STRIKES_LOG trigger)
• Same component fixed twice with different root causes
• bug-tag: repeating in commit body for the same issue
• Escalated from debugging-and-error-recovery after repeated fix attempts

When NOT to Use

• First occurrence of a bug (use debugging-and-error-recovery instead)
• One-off test flakiness or network-dependent failure
• Different components failing independently (not the same root cause)

Output Contract

Strike log entry + escalation diagnosis — .git/STRIKES_LOG recording each strike with test name and timestamp, root cause documented per strike, route-to-skill assigned at strike 3 (test gap → TDD, spec gap → review, process gap → doubt, architecture gap → analysis), user notified before any fix code at strike 3.

Trigger: Test Failure Recurrence

A mechanical pre-commit check reads .git/STRIKES_LOG. Same test name failing 3 consecutive commits → BLOCK and escalate.

Commit 1: test-X FAIL → strike 1 (logged)
Commit 2: test-X FAIL → strike 2 (logged)
Commit 3: test-X FAIL → BLOCK. Escalating.

Other triggers: session-level agent memory (same component fixed twice), bug-tag: in commit body repeating.

Protocol

Strike 1 — Diagnose then fix. Every attempt starts with evidence, not speculation. Inspect DOM, network, console, logs. Document root cause. Fix. Verify tests pass.

Strike 2 — Your diagnosis was wrong. Re-diagnose with fresh context. Fix + add regression guard (test coverage, instrumentation, or lint rule).

Strike 3 — Systemic failure. The diagnosis-fix-verify cycle works but the system reproduces the bug. BLOCK. No more fix code.

Agent must on strike 3:

1. Stop all commits
2. Inspect real state (computed styles, network, console, DOM)
3. Document root cause vs expected behavior
4. Report to user before any fix code

Escalation Matrix

Anti-Patterns

1. Same fix, harder — The definition of insanity. Strike 3 means your approach is wrong.
2. Skipping diagnosis — "I know what's wrong" without inspecting. You didn't last 2 times.
3. Not escalating — Fixing strike 3 locally when the problem is systemic.
4. Ignoring strike log — Strike 3 didn't come from nowhere. The log told you test X was failing.

Verification

• Strike 1: diagnose before fix, evidence documented
• Strike 2: re-diagnose, add regression guard
• Strike 3: BLOCK, escalate to appropriate skill
• bash tests/run-all.sh passes after fix