feat(ce-debug): add systematic debugging skill

[tmchow]

Summary

New /ce-debug skill that systematically finds root causes and fixes bugs. Replaces /reproduce-bug, which was limited to GitHub issues and stopped at reproduction without fixing.

ce-debug handles any debugging entry point: stack traces, failing tests, issue tracker references (GitHub, Linear, Jira), or vague symptom descriptions. It traces the full causal chain before proposing a fix, then lets the user choose whether to fix now, share the diagnosis via Proof, or rethink the design via brainstorm.

Fixes #533

What changed

New skill: ce-debug/SKILL.md (191 lines)

• 5 self-sizing phases: Triage, Investigate, Root Cause, Fix, Close. No complexity tiers. The same phases handle a missing import (fast) and a race condition (deep).
• Causal chain gate: cannot propose a fix until the full chain from trigger to symptom has no gaps.
• Predictions for uncertain links: when a causal chain has non-obvious steps, requires a falsifiable prediction in a different code path. Wrong prediction + working fix = symptom fix.
• Smart escalation: when 2-3 hypotheses fail, diagnoses why (architecture problem, wrong mental model, environment issue, symptom fix) instead of just stopping.
• Test-first fix phase with workspace safety checks, conditional defense-in-depth, and conditional post-mortem.
• User chooses next step after diagnosis: fix now, view in Proof, or rethink via brainstorm.

New reference: references/anti-patterns.md (91 lines)

• Loaded before hypothesis formation to create friction against common debugging shortcuts.
• Covers: prediction quality (good vs bad examples), shotgun debugging, confirmation bias, "it works now" trap, shortcut signals, smart escalation patterns.

New reference: references/investigation-techniques.md (161 lines)

• Loaded when deeper investigation is needed (intermittent bugs, framework-specific tracing, browser debugging).
• Covers: backward root-cause tracing, git bisect, intermittent bug techniques, framework-specific debugging (Rails/Node/Python), race conditions, browser debugging with agent-browser.

Deleted: reproduce-bug/SKILL.md

• ce-debug covers all reproduce-bug trigger conditions and more. The trigger description includes issue tracker references, stack traces, test paths, and vague descriptions.

Updated: README.md

• Added ce-debug to Core Workflow table, removed reproduce-bug from Workflow Utilities.

Test plan

• bun run release:validate passes (42 skills, 51 agents)
• bun test passes (657 tests, 0 failures)
• Eval: stack trace entry point correctly traces NPE to upstream silent null return
• Eval: failing test entry point finds root cause without just making the test pass
• Eval: vague description + stuck user triggers prior-attempt awareness, identifies root cause upstream from where user suspected
• Eval: browser/API bug traces 404 from frontend redirect to missing API response field
• Eval: visual CSS bug drives agent to attempt agent-browser before falling back to code reading

---

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f907973c08

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".