Claude Code (Opus 4.6) repeatedly ignores documented governance rules within same session

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

Follow mandatory process loops documented in CLAUDE.md and governance docs: Builder→Reviewer→Skeptic→Fixer sequential loop for every code change. Reviewer must read governance documents before reviewing. Never fix without understanding context. Tests must never work around bugs. Fix all issues regardless of severity.

What Claude Actually Did

1. Shipped 8+ changes without running the Reviewer→Skeptic loop (tooltips across 10 forms, alignment fixes, navigation fix, skipText forwarding, hideSubmit, education layout, stream case fix, grade tooltip)
2. Launched Reviewer and Skeptic in parallel despite explicit sequential rule — corrected 5+ times in the same session, kept reverting
3. Never instructed Reviewer agents to read governance documents (ENTERPRISE_CODING_STANDARDS.md, CODE_REVIEWER_STANDARDS.md) — the docs were never checked
4. Made blind grep-replace fixes (linkedinUrl→linkedinProfile across 5 files, col-span-6→col-span-1) without reading file context first
5. Test builder worked around bugs instead of asserting expected behavior (skipped proficiencyLevel field, accepted 0% completion from wrong field names)
6. Deferred LOW severity issues as "tech debt" instead of fixing them
7. After each correction, acknowledged the rule, promised to follow it, then broke it again within hours
8. Rules documented in 4 locations (CLAUDE.md with 17 rules, 35+ memory entries, 8 governance docs, testing standards) — none consistently followed

Expected Behavior

AI follows documented governance process consistently without requiring human oversight for every step. Rules documented in CLAUDE.md and governance docs are treated as mandatory, not advisory.

Files Affected

- CLAUDE.md — 17 project rules, all applicable
  - docs/governance/core/ENTERPRISE_CODING_STANDARDS.md — never read during reviews
  - docs/governance/core/CODE_REVIEWER_STANDARDS.md — 25-point checklist never applied
  - docs/governance/core/TESTING_STANDARDS.md — rules #8, #9, #10 violated
  - .claude/projects/*/memory/ — 35+ feedback memories documented but not followed

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

1. Document governance rules in CLAUDE.md (Builder→Reviewer→Skeptic sequential, governance doc verification mandatory)
2. Store reinforcement in memory files (feedback entries correcting the same issues)
3. Add rules to governance docs (CODE_REVIEWER_STANDARDS.md, TESTING_STANDARDS.md)
4. Ask Claude to build a feature
5. Observe: Claude ships code without running the review loop
6. Correct Claude — it acknowledges and follows for 1-2 cycles
7. Ask Claude to build another feature
8. Observe: Claude breaks the same rules again
9. Correct again — Claude adds ANOTHER rule to prevent it
10. The new rule is also not followed

Claude Model

Opus

Relevant Conversation

Extended session (2 days) with conversation compaction. Multiple corrections documented in memory files feedback_reviewer_skeptic_sequential.md, feedback_no_workarounds_in_tests.md, feedback_context_before_fixing.md, feedback_code_review_standard.md.

Impact

Medium - Extra work to undo changes

Claude Code Version

2.1.86

Platform

Anthropic API

Additional Context

• Model: Claude Opus 4.6 (claude-opus-4-6, 1M context)
• OS: Windows 11
• Session type: Extended with context compaction (conversation exceeded context limits and was compressed multiple times)
• The pattern persists even after context compaction — rules survive in CLAUDE.md and memory files but behavioral compliance does not
• Adding more rules makes the problem worse — there are now 17 rules in CLAUDE.md, 35+ memory entries, and 8 governance docs, creating a large surface area that the AI acknowledges but doesn't consistently enforce
• The issue is most acute when the AI is under "pressure" — multiple parallel tasks, user saying "don't stop", or changes that seem "small enough to skip review"