fix: eliminate security hook trigger patterns in /codex and /autoplan (#1329)

[NikhileshNanduri]

Summary

Fixes the remaining 3 of 4 security hook trigger patterns from issue #1329. PR #1496 already shipped Pattern 2 (eval with tilde path via gstack-paths --get). This PR handles Patterns 1, 3, and 4.

Pattern 1 — source with tilde path (both templates):

• Before: source ~/.claude/skills/gstack/bin/gstack-codex-probe → triggers PreToolUse hook on tilde-path source
• After: Direct calls to ~/.claude/skills/gstack/bin/gstack-codex-* standalone binaries

Pattern 3 — bare cd "$_REPO_ROOT" (codex template only):

• Before: Standalone cd "$_REPO_ROOT" line before codex commands triggers cd + command hook
• After: -C "$_REPO_ROOT" flag on codex commands; git -C "$_REPO_ROOT" diff for git commands; codex exec resume drops the cd entirely (session context preserves directory; -C is not a supported flag for resume)

Pattern 4 — inline python3 with #-comments (codex template only):

• Before: Three inline python3 -u -c "..." blocks with Python-style comments trigger multi-line comment hook
• After: All three blocks pipe to ~/.claude/skills/gstack/bin/gstack-codex-jsonl-parser --mode challenge|consult

New files

File	Purpose
bin/gstack-codex-auth-probe	Standalone multi-signal auth check
bin/gstack-codex-version-check	Warns on known-bad Codex CLI versions
bin/gstack-codex-log-event	Telemetry emitter (reads config internally)
bin/gstack-codex-log-hang	Operational learning writer on timeout
bin/gstack-codex-timeout-wrapper	gtimeout/timeout/unwrapped fallback chain
bin/gstack-codex-jsonl-parser	Python JSONL parser for codex --json output

Testing scenarios

Static validation (free, bun test)

1. Pattern guard tests — test/codex-hardening.test.ts adds 38 new tests:

   • Pattern 1: asserts no source ~/.*gstack-codex-probe in codex/autoplan templates and generated SKILL.md files
   • Pattern 3: asserts no bare cd "$_REPO_ROOT" as a top-level line in any of the 4 files
   • Pattern 4: asserts no $PYTHON_CMD.*-u\s+-c\s+" pattern (inline python) in any of the 4 files

2. Standalone binary tests — test/codex-hardening.test.ts:

   • All 6 binaries exist and are executable
   • Bash binaries are syntax-valid (bash -n)
   • Python parser is syntax-valid (python3 -c "ast.parse(...)")
   • gstack-codex-auth-probe: AUTH_OK on CODEX_API_KEY, AUTH_FAILED with no auth, AUTH_OK on auth.json
   • gstack-codex-timeout-wrapper: executes directly without timeout binary, prefers gtimeout
   • gstack-codex-jsonl-parser: extracts agent_message, SESSION_ID in consult mode, no SESSION_ID in challenge mode, disconnect warning in challenge mode, tokens from turn.completed, [codex thinking] for reasoning, [codex ran] for command_execution, tolerates malformed JSON

3. Updated skill-validation test — Python discovery test updated to check for gstack-codex-jsonl-parser binary invocation instead of old $PYTHON_CMD inline pattern

4. Full free suite: bun test passes with only the 2 pre-existing AGENTS.md/docs/skills.md doc-inventory failures (confirmed on clean main)

Verification command

bun test test/codex-hardening.test.ts test/skill-validation.test.ts test/gen-skill-docs.test.ts
# Expected: 779 pass, 2 fail (pre-existing doc inventory)

🤖 Generated with Claude Code