feat: add codex hook integration

CHANGELOG.md

@@ -0,0 +1,190 @@
+# Changelog
+
+All notable changes to OpenWolf are documented here. The format follows
+[Keep a Changelog](https://keepachangelog.com/), and OpenWolf uses
+[Semantic Versioning](https://semver.org/).
+
+> This fork (`@alptech/openwolf`) keeps its own package version. Upstream
+> feature history below is retained for reference; package version remains
+> under the fork's numbering (currently 1.0.6).
+
+## [1.0.6] - 2026-08-12
+
+### Changed
+
+- Synced business code from upstream [cytostack/openwolf](https://github.com/cytostack/openwolf)
+  (through 2.0.2 feature set): multi-agent adapters, PreCompact, durable
+  anatomy store, measured token usage, redesigned dashboard, bundled skills.
+- Retired Design QC path in favor of upstream architecture.
+- README rewritten for the synced feature set; added Chinese README and
+  upstream acknowledgments. Package name/author/version remain this fork's.
+
+## [2.0.2] - 2026-07-15
+
+### Added
+
+- Antigravity agent adapter (beta, context-level via `AGENTS.md`).
+  `openwolf init --agent antigravity` and `--agent all` now include it, and
+  auto-detection picks it up when Antigravity is installed.
+
+### Changed
+
+- Documentation and website refreshed to reflect v2 throughout: the seven
+  lifecycle hooks including PreCompact, the durable anatomy store with
+  symbol-level reads, measured token usage, the redesigned dashboard, the
+  `/reframe` skill, and per-project dashboard ports. Retired Design QC
+  content removed. Positioning generalized across supported agents.
+
+## [2.0.1] - 2026-07-15
+
+### Fixed
+
+- Dashboard no longer white-screens when the server rejects the token. A 401
+  now renders a clear "token rejected" message with guidance instead of
+  crashing the page. Root cause: `StatusBadge` threw on an undefined status,
+  and failed API responses were being fed into component state.
+- Multi-project port collisions resolved. Projects upgraded from 1.x all kept
+  the shared default dashboard and daemon ports, so only the first project's
+  dashboard would ever open. Three fixes work together: `openwolf update`
+  reassigns a free port pair when a project's ports collide with another
+  registered project, `openwolf dashboard` starts this project's server on a
+  free port when the configured one is held by another project's daemon
+  (instead of opening a URL that gets a 401), and the daemon accepts an
+  `OPENWOLF_DASHBOARD_PORT` override. Fresh installs already received unique
+  ports; this brings upgraded projects to parity.
+
+## [2.0.0] - 2026-07-15
+
+OpenWolf 2.0 turns the second brain for Claude Code into a context layer for
+every AI coding assistant, with verifiable token measurement, a hardened
+security posture, and a re-architected project index.
+
+### Added
+
+Multi-agent support:
+
+- Agent adapter architecture: `openwolf init` now auto-detects the coding
+  agents installed on your machine and wires each of them to the same `.wolf/`
+  brain. Explicit control via `--agent codex opencode gemini cursor`, `--agent all`,
+  or `--agent claude` to opt out.
+- Codex CLI integration: project-level lifecycle hooks via `.codex/hooks.json`
+  plus an `AGENTS.md` protocol block.
+- OpenCode integration: a native plugin installed to `.opencode/plugin/` that
+  maps OpenCode tool events onto the `.wolf/` state.
+- Gemini CLI integration: `GEMINI.md` protocol block.
+- Cursor integration: an always-applied rule at `.cursor/rules/openwolf.mdc`.
+- Protocol blocks are marker-fenced and idempotent: your own content in
+  `AGENTS.md` or `GEMINI.md` is never modified, and re-running init never
+  duplicates anything.
+
+Measured token usage:
+
+- The Stop hook reads real API usage from the harness transcript (input,
+  output, cache read, cache write tokens, and API call count) into the token
+  ledger. Estimates and measurements are reported side by side.
+- New `openwolf report` command: estimated vs measured usage in the terminal.
+- Per-agent session attribution: every ledger session records which agent ran it.
+
+Context management:
+
+- Session digest: at session start, a token-budget-capped digest of the most
+  valuable state (STATUS.md next phase, Do-Not-Repeat list, recent bug fixes,
+  anatomy pointer) is injected directly into the model's context.
+  Budgets are configurable per agent in `config.json`.
+- Compaction survival: a new PreCompact lifecycle hook
+  snapshots in-flight session state, and session start after compaction
+  re-injects a digest of the files already modified. Session state is no
+  longer wiped on resume or compaction.
+- Anatomy staleness detection: scans pin the git HEAD; if the HEAD moves or
+  the scan ages past the configured interval, the agent is told to rescan
+  before trusting the index.
+- End-of-turn reminders now reach the model through the `additionalContext`
+  channel instead of invisible stderr.
+- `STATUS.md` session handoff document: resume any session in one small read.
+
+Anatomy re-architecture:
+
+- Durable store: the source of truth for the project index moved from
+  `anatomy.md` itself to `.wolf/anatomy-index.json`, with `anatomy.md`
+  rendered from it. Concurrent writers now coordinate through a
+  cross-platform lock; simultaneous edits no longer lose entries.
+- Version-skew safe: markdown written by older hooks or edited by hand is
+  detected by content hash and absorbed additively into the store.
+- Symbol-level entries: files above 500 estimated tokens index their
+  top-level functions and classes with line ranges and per-slice token
+  estimates (TypeScript, JavaScript, Python, Go, Rust). The pre-read hint
+  points agents at exact line ranges so they can read one function with
+  offset/limit instead of the whole file. Hints are suppressed automatically
+  if the file on disk has changed since indexing.
+
+Skills and tooling:
+
+- Bundled skills installed on init for Claude Code, Codex, and OpenCode:
+  `/security-audit` (layered audit: dependencies, secrets, injection
+  surfaces, authorization, ranked report) and `/reframe` (framework
+  selection and migration plus a design audit/fix mode).
+- `scripts/openwolf-check.mjs`: a standalone, read-only inspector that
+  reports whether OpenWolf is installed in a project, which agents are
+  wired, recency, and lifetime plus recent-session statistics.
+- `openwolf update` now has parity with init: it creates missing files,
+  re-runs the recorded agent adapters, refreshes bundled skills, and
+  performs one-time data migrations, all after taking a timestamped backup.
+
+Dashboard 2.0:
+
+- Complete redesign: monochrome dot-matrix design system with a single
+  signal-red accent, top navigation, bento stat tiles, and hash-based deep
+  links to panels.
+- Surfaces the 2.0 data: measured vs estimated tokens, cache economics,
+  per-agent breakdown table, wired-agents widget, context health (scan
+  freshness, pinned git HEAD, digest budget), and the STATUS.md handoff.
+- Reliable Run Now for cron tasks over authenticated HTTP with visible
+  running/queued/failed feedback.
+
+### Changed
+
+- Reframe now leads with an anti-generic design mandate: a blocklist of the
+  recognizable AI-generated aesthetic plus positive principles, applied to
+  every framework migration prompt. Distinctiveness is an acceptance criterion.
+- Astryx added as the 13th framework in the Reframe knowledge base.
+- Contributors are credited in the README; detailed attribution lives in
+  commit trailers.
+- STATUS.md template localized to English.
+
+### Fixed
+
+- CRLF line endings no longer wipe `anatomy.md` on Windows (#50, #24).
+- Concurrent post-write hooks no longer lose anatomy entries.
+- Old `config.json` files without newer sections no longer crash commands (#26, #27).
+- `openwolf init` and `openwolf update` no longer reset per-project ports;
+  fresh projects get a free port pair automatically (#37, #38).
+- `bug search` is null-safe across buglog schema drift (#44).
+- `EPERM` on WSL2 with EFS-encrypted directories fixed via a copy shim (#33).
+- Files outside the project root no longer pollute the index (#56).
+- Documentation and config edits are no longer mislogged as bug fixes, and
+  auto-detection can be disabled (#28, #57).
+- Dart language support in the scanner (#10).
+
+### Security
+
+- Dashboard binds to 127.0.0.1 by default and requires a per-project token
+  (timing-safe comparison) for all API and WebSocket access (#30, #34).
+- Command injection eliminated: every dynamic process invocation uses
+  argument arrays; a shell-mode spawn was removed from the cron engine.
+- Path traversal guards (realpath-based, symlink-safe) on cron AI task file access.
+- File-watcher broadcasts capped at 1 MB to prevent memory abuse.
+- Secret-bearing files (keys, keystores, credential files, `.npmrc`, and
+  more, not just `.env`) are excluded from all index and memory capture (#54).
+- A security regression test suite runs with `pnpm test`, including a guard
+  test that fails the build if injectable process calls ever return.
+
+### Removed
+
+- Design QC screenshot capture (agents capture and read their own
+  screenshots now); the `puppeteer-core` dependency is gone.
+- The unverifiable token comparison chart in the dashboard; only measured
+  numbers or clearly labeled estimates are shown.
+
+## [1.0.4] - 2026-03-20
+
+Final 1.x release. Claude Code only.