v0.1.1

Fixed

• The wrapper now writes the exit status of a run to <log>.exit, and the
  hook trusts that file rather than the printed marker line. A marker is
  honoured only when its id was minted by the PreToolUse hook in the same
  session, no receipt carries the id yet, and the status file exists; other
  marker lines are dropped before the output is read. Before this, a
  [stalegreen] exit=0 receipt=... line printed by the agent could mint a
  receipt, refresh an old one with the current tree, or change the exit of a
  real run.
• A verification run that finished in the background (Codex wait) is dated
  from its start and carries no fingerprint, so edits made while it ran make
  it stale. Before this, its fingerprint was taken when the output arrived
  and edits made during the run went unnoticed.
• One receipt is written per id; a pending record whose id already has a
  receipt no longer swallows a later run.

Full Changelog: v0.1.0...v0.1.1

stalegreen 0.1.0

stalegreen 0.1.0: hooks that keep a coding agent's green claims honest, for Claude Code, Codex and DeepSeek Harness.

Added

• Runner catalog covering test, typecheck, lint and build commands for the
  JavaScript, Python, Go, Rust, JVM, .NET, Ruby, PHP, Elixir, Swift, Dart and
  C/C++ ecosystems, with pass and fail fixtures for every runner.
• Claim grammar with a labelled corpus of 350 sentences; assertive claims are
  matched to a category and a scope, hedged, future, negated, quoted, relayed
  and question forms are ignored, and scoped claims are reported but never
  blocked.
• Receipts for every verification run, with counts, exit status, masking
  analysis and a working-tree fingerprint that survives git commit.
• Edit events from the file tools and from shell commands that change files.
• Freshness gate for Claude Code (Stop and SubagentStop) that blocks a green
  claim when its evidence is stale, failed or masked, with a loop guard so a
  claim is blocked at most once per turn.
• stalegreen check, stalegreen receipt and stalegreen doctor, with
  doctor --prune removing sessions older than the configured window.
• stalegreen history: replays past Claude Code sessions (main transcript
  and subagents, streamed) through the receipt builder and the gate and lists
  the claims whose evidence was stale, failed or masked, with --explain for
  the receipt, the command and the sentence behind each one.
• stalegreen stats: stale, failed, masked and unbacked claim rates and the
  hidden-exit rate of verification runs over a window, per harness, per
  model and per session kind. Sessions without tool calls are excluded and a
  status line repeated word for word within a session is counted once per
  verdict.
• Codex support: install --codex writes the same four hooks to
  ~/.codex/hooks.json (or the repository's .codex/hooks.json). The Codex
  adapter reads the Bash and apply_patch tools, the Script completed,
  Script failed and Script running with cell ID headers, finishes a
  still-running command from the wait tool's output, returns the rewrite
  with the allow decision Codex requires, and blocks at Stop with
  {"decision":"block","reason":...}. The rollout reader behind history
  and stats merges child threads into their parent and reads the shell
  commands out of Codex's JavaScript exec cells.
• The hook handlers are shared across harnesses through a small adapter
  (turn id, shell output shape, edit tools, allow decision, block format).
• DeepSeek Harness support: a dsh adapter (runDshHook) that reads the
  bash tool's [exit code: N] marker and the edit, write and
  str_replace_editor tools, and dsh-plugin-stalegreen, a Cordis plugin
  package in this repository that mounts the gate on tools/pre-execute,
  tools/post-execute and agent/turn-stopping and steers one correction
  step when a claim is stale, failed or masked.
• stalegreen redact: a shareable copy of a session (receipts, edits,
  verdicts, pending and deferred runs, the tail of each log) with paths
  shortened, secrets masked and the agent's prose replaced by the matched
  claim. The false-block issue template asks for it.
• A release workflow that publishes to npm through trusted publishing on a
  version tag and creates the GitHub release. CI runs the suite on Ubuntu and
  macOS with Node 20 and 22, and the pure-logic tests on Windows, where hook
  paths and report paths are written with forward slashes.
• Claude Code plugin packaging: plugin/ holds the manifest, the hook
  registration (${CLAUDE_PLUGIN_ROOT}/hook.js) and a copy of the compiled
  hook that npm run build refreshes and a test keeps identical to the
  build; .claude-plugin/marketplace.json at the repository root serves it
  through /plugin marketplace add pavangupta352/stalegreen. The plugin
  lives in its own directory so that installing it never triggers a
  dependency install.
• The unmask rewrite: the PreToolUse hook wraps verification commands in
  POSIX sh so the full output is logged, the tail is shown, an explicit
  [stalegreen] exit=<code> receipt=<id> marker is printed and the exit
  status is preserved. Filter pipes such as | tail -5 and | grep are
  dropped, tee targets are kept, and heredocs, backgrounding, process
  substitution, sudo, nested shells, redirects and watch modes are left
  untouched. Tested against 210 commands under sh, bash and zsh.
• Permission-aware rewriting: the hook returns permissionDecision: "allow"
  only when the user's own Claude Code rules already allow the original
  command, so wrapping never widens permissions (permission setting:
  inherit, allow or ask).
• stalegreen install --claude and uninstall --claude, with --project
  and --advisory; doctor reports the registered hooks.

Install: /plugin marketplace add pavangupta352/stalegreen then /plugin install stalegreen@stalegreen, or npx stalegreen install --all. DeepSeek Harness: dsh plugin add dsh-plugin-stalegreen.