agent-witness
CLAUDE.md1 starsChanged 5 months agoApache-2.0View on GitHub
A session recorder and audit log for AI coding agents. Records what a Claude Code / Codex session actually did (tool calls, files, commands), shows it as a TUI timeline, and persists a replayable JSONL audit trail. Observation only — not a sandbox. The primary merge gate is just verify locally; GitHub Actions CI mirrors the exact same gate (public-OSS trust signal — see ADR-0003 for why this diverges from avatar-core's PoC no-CI stance). If just verify is green, the…
- Installs packages
In this repository12 other files it gives its agents.
- AGENTS.mdAGENTS.md
- auditSkill
- decomposeSkill
- dev-allSkill
- dev-investigateSkill
- devSkill
- digSkill
- investigateSkill
- prSkill
- reviewSkill
- tech-debtSkill
- update-docsSkill
The file
We split this CLAUDE.md into the rules it gives an agent
1 rule. About 0 are common, 0 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈0
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 73
No magic numbers; errors via thiserror in lib crates, anyhow in bins.
≈ “Common” is an estimate: at least 1 in 5 of the 28,029 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
agent-witness
@REVIEW.md
Project Overview
A session recorder and audit log for AI coding agents. Records what a Claude Code / Codex session actually did (tool calls, files, commands), shows it as a TUI timeline, and persists a replayable JSONL audit trail. Observation only — not a sandbox.
Core Values
- Honest observation — never claim more than what was observed. Every event carries
attribution: direct | observed | inferredand a confidence. Overclaiming is the one unforgivable bug. - Zero-friction adoption —
agent-witness initand you're recording. No sudo, no entitlements, no wrapper required in v0.1. - Evidence you can share — every session becomes a replayable record and a markdown report you can paste into a PR or issue.
Won't Do
- Sandboxing / prevention: Anthropic ships an official sandbox; we observe and audit, we do not block (policy layer in v0.2 is hooks-based deny, still not a security boundary)
- Being an agent / agent framework: saturated, out of scope
- v0.1 scope cuts (deliberate, from design review): PTY wrapping, fs watching, process polling → v0.2; socket-level network observation → v0.2 proxy / v0.3 eBPF
- Quantitative "we catch everything" claims: hooks don't see inside
Bash("script.sh")side effects — document the gap, never paper over it - Windows first-class support in v0.1: best effort only
Build & Run
rustup default stable
cargo install cargo-nextest just
just verify # primary local gate: check (fmt+clippy+build) + nextest. Run before merge.
just check # fmt + clippy + build
just test # nextest
just buildVerification
The primary merge gate is just verify locally; GitHub Actions CI mirrors the exact same gate (public-OSS trust signal — see ADR-0003 for why this diverges from avatar-core's PoC no-CI stance). If just verify is green, the change is mergeable.
Architecture
crates/
witness-core/ # lib: AgentEvent model, JSONL session store, adapters (hooks receiver, transcript)
agent-witness/ # bin: CLI (init/watch/ls/show/report/emit/top) + ratatui TUI
tests/fixtures/ # sanitized hook-event JSONL from real sessions (golden test inputs)Event flow: Claude Code hooks → unix socket (emit bridge) → normalizer → JSONL store (~/.agent-witness/sessions/<id>/) + ring buffer → TUI / report.
Tech Stack
| Layer | Technology |
| async runtime | tokio |
| TUI | ratatui + crossterm (TestBackend for golden tests) |
| CLI | clap (derive) |
| serialization | serde / serde_json (JSONL) |
| errors | thiserror (core) / anyhow (bin) |
| distribution | cargo-dist + Homebrew tap + cargo-binstall |
Key Gotchas
- Claude Code hooks payloads:
PreToolUse= tool_name/tool_input/tool_use_id;PostToolUseadds tool_response/duration_ms; hook input includestranscript_path— use it per-session, never glob~/.claude/projects/*.jsonl - Transcript JSONL internal schema is NOT a stable contract — transcript adapter is best-effort, versioned, feature-flagged; hook stdin JSON is the canonical record
agent-witness initedits~/.claude/settings.jsonAND installs~/.claude/skills/witness/SKILL.md— both must be idempotent, merge-safe, and keep a backup; the skill file is marker-owned and a foreign file at its path is never touched- Fixtures must be sanitized (no real paths/secrets from recorded sessions)
- Naming: crates.io
agent-witness/agent-witness-core; note cursor/agent-trace is an export-format interop target, not a competitor
Rust Conventions
just checkmust pass (fmt, clippy-D warnings, build); tests via nextest.- No
unsafe. Determinism in core logic: no wall-clock or RNG inside pure paths — inject them. - No magic numbers; errors via
thiserrorin lib crates,anyhowin bins; nounwrap/expectoutside tests.
Development Harness
Issue-driven development with /dev (single issue) and /dev-all (sequential). Other skills: /audit, /update-docs, /decompose, /investigate, /review, /tech-debt, /pr.
Review accumulation (ADR-0003): valid review findings are promoted into rules (.claude/rules/), lints, and skills. Promotion: a finding that recurs twice becomes a rule. Retirement: a rule unused for 3 months is removed.
Human judgment points: event/report schema changes, CLI surface, launch copy. Every PR includes a short design-decision note (launch article material). No squash merges.
Phasing
Design is decided ahead (docs/adr/, zero-base design docs), but implementation may diverge. v0.1 issues are fully detailed; v0.2+ exist only as ADR notes and are detailed at the v0.1 gate. The v0.1 gate is a human decision after the first vertical spike (calibrate actual pace; shrink scope if estimates double).
Language
- Code comments, variable names: English
- Commits: concise single line, English
- GitHub Issues and PRs (titles, bodies): English — this is a public build-in-public repo
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

