Skip to content

v0.5 — Skill improvements, sensitive-file-guard, installer auto-merge

Choose a tag to compare

View all tags
@kimon1230 kimon1230 released this 28 May 11:43
· 5 commits to main since this release
Immutable release. Only release title and notes can be modified.
v0.5
45cd5d2

Added

  • sensitive-file-guard.py (PreToolUse hook) — blocks reads of .env, SSH keys, credentials, AWS configs, and package tokens. Intercepts both Read tool and Bash commands via shlex.split() tokenization. Case-insensitive basename matching with explicit allowlists (.env.example, .envrc). *.key requires keyword guard to avoid false positives. 83 tests.
  • Output Contracts for /code-review, /critical-review, /security-audit — each skill now defines the required structure of its final synthesized report (findings table, aggregate counts, verdict, remediation offer).
  • Rationalizations to Reject — 9 security dismissals injected into /security-audit subagent prompts ("It's behind a VPN", "Only admins can reach this", etc.) and 6 code quality dismissals into /code-review.
  • Red Flags — thought-pattern watchlists injected into /security-audit (5 items) and /code-review (5 items) targeting the agent's own corner-cutting impulses.
  • CTF-sourced detection patterns for /security-audit — PHP attack surface (gated behind php_detected flag), encoding/parsing mismatches (Unicode normalization, Shift-JIS, U+00A0), Go len() byte/rune confusion, prototype pollution, JWE token handling, auth race conditions, deserialization depth (XMLDecoder, Castor XML, pickle, PHP serialization), SSRF-to-Docker/gopher, CSP bypass taxonomy, DOM clobbering, behavioral framework XSS narrowing.
  • Implementer status protocol for /implement-batch — subagents report STATUS: DONE, DONE_WITH_CONCERNS, NEEDS_CONTEXT, or BLOCKED. Main session handles each status with defined escalation paths.
  • Contradiction detection for /implement-batch — scans subagent output for phrases like "requires manual", "should work", "TODO" alongside success claims. Catches the "claims done but admits failure" pattern.
  • Cascading-fix escalation for /implement-batch — 3-strike rule. If fixes keep breaking previously-passing modules, stop and flag as architectural issue.
  • Complexity gate for /implement-batch — flags tasks meeting 2-of-3 conditions (>5 files, >3 acceptance criteria, cross-module dependencies) before spawning subagents. Soft gate — user can override.
  • Decision logging for /critical-review — appends accept/reject/defer decisions to decision-log.md with rationale. Triggers on all user response paths.
  • Task Delivery States in CLAUDE.md — mental model for task progression (intake → planning → executing → validating → reviewing → done / blocked).
  • Subagent context isolation rule in CLAUDE.md — always paste full task text into subagent prompts, never make subagents read plan files.
  • Blast radius risk labeling for /code-review — findings include Risk: HIGH|MED|LOW based on structural heuristics. Secondary sort by risk within same severity.
  • Mock quality detection for /code-review — Agent 5 now flags mock-heavy tests (3:1 ratio), missing real module imports, behavioral-only assertions, and over-mocked integration tests.
  • BEFORE/AFTER remediation verification for /code-review and /security-audit — records failing state before fix, verifies passing state after. Uses function/symbol anchors for non-testable findings.

Changed

  • Skill descriptions — all 6 skills rewritten for CSO compliance (trigger conditions only, no workflow summaries). Prevents Claude from shortcutting skill bodies.
  • Skill frontmatter — all 6 skills now include risk: safe|critical field and optional risk-note.
  • Rules frontmatter — python.md, javascript.md, shell.md now include paths frontmatter for file-type-specific loading.
  • /implement-batch step 4 — split into 4 (triage: status + contradiction), 4b (consistency review), 4c (validation).
  • /implement-batch step 7 — phase gate: recommends new session for batch ≥ 3.
  • /critical-review — renumbered to 8 steps (decision logging is step 5; non-approval path explicitly logs rejections).
  • CLAUDE.md Code Quality — overloaded paragraph broken into sub-bullets.
  • CLAUDE.md Planning — batch numbering convention: always start at 1, never 0.
  • install.sh — now fully supports upgrades in addition to fresh installs:
    • install_settings() function: jq-based merge of hooks and statusLine from settings.json.reference into ~/.claude/settings.json. Backs up before modifying.
    • Already-installed symlinks detected and skipped without prompting (compares raw and resolved paths).
    • Stale symlink cleanup: broken symlinks pointing into the repo (from removed/renamed components) are detected and offered for removal.
    • Settings.json upgrade: repo-managed hook entries are stripped and re-added from the current reference on every run, handling renamed hooks, changed command formats, and removed hooks. User hooks are preserved.
    • Duplicate hook entries from older installer versions are deduplicated.
    • Repo-managed statusLine is updated to match the current reference; custom statusLine configurations are preserved.
    • CLAUDE.md, settings.json, and component prompts all skip when already current — a fully up-to-date install produces no prompts.
    • Extracted merge_settings_json() and install_settings_needed() helpers to eliminate logic duplication.
  • uninstall.sh — new uninstall_settings() function: removes hook entries pointing to /.claude/hooks/ and statusLine from ~/.claude/settings.json. Replaces old manual reminder.
  • settings.json.reference — added PreToolUse section for sensitive-file-guard (Read + Bash matchers).
  • /security-audit synthesis — now reports php_detected status alongside web_app and iac_detected.
Assets 3
Loading