yvgude/lean-ctx v3.8.2

on GitHub

Fixed

• Linux: ctx_* tools broke for projects under /c/… and other
  single-letter roots (#397): the MSYS2/Git-Bash drive mapping
  (/c/Users/… → C:/Users/…) in the MCP path normalizer ran
  unconditionally — on Linux/macOS, where /c/… is a literal directory,
  every file-addressing tool then failed with file not found on a
  nonexistent C:/… path (and absolute arguments were re-joined under the
  already-translated root, doubling it). The mapping is now gated on Windows
  hosts (cfg!(windows)) — that is the only platform where MSYS2/Git-Bash
  clients hand POSIX drive paths to a native Windows binary. On other hosts,
  /c/… passes through untouched; regression tests cover both sides.
• lean-ctx doctor reported "no rules file found" right after lean-ctx setup
  (#396): the 3.8 layout (GL #555) intentionally replaced the always-loaded
  ~/.claude/rules/lean-ctx.md with a CLAUDE.md block + on-demand skill — setup
  even removes the legacy file — but the doctor check still demanded it, so a
  clean install could never reach a full pass and the suggested fix
  (init --agent claude) couldn't recreate the file either. Both doctor views
  (doctor and doctor integrations) now share one layout detector
  (claude_instructions_state) that accepts every state setup can produce:
  CLAUDE.md block (+ skill), dedicated injection (SessionStart hook + skill),
  legacy rules file, project scope, and rules_injection=off. Docs that still
  described the retired rules file were updated as well.
• macOS still prompted "lean-ctx would like to access files in your Documents
  folder" on every upgrade (#356, reopened): the first fix (3.8.0) removed the
  scan-heuristic probes, but the prompt actually came from the launchd
  daemon's boot path — a process that is its own TCC identity, and whose
  grant is invalidated by every update (binary swap → new cdhash → re-prompt).
  Traced empirically with a deny-sandbox + crash-stack bisection; two
  independent boot-time offenders fixed:
  1. serve booted with cwd / and walked every stored session, stat-ing
     each session's project_root/shell_cwd (project-marker probes +
     canonicalize) — paths that usually live under ~/Documents. Broad
     roots ("/", HOME, agent sandboxes) now bail out before the scan — they
     can never own a session (this also stops shell_cwd.starts_with("/")
     from leaking an arbitrary project's session into the daemon default).
  2. ContextLedger::load → prune ran realpath over every persisted ledger
     entry at boot for its dedupe key; the key is now lexical-only.
     Defense in depth: launchd-owned processes (ppid 1) are detected as
     TCC-standalone and never stat/canonicalize paths under
     ~/Documents/Desktop/Downloads in heuristics (has_project_marker,
     session-root matching, normalize_tool_path); editor/CLI children inherit
     their host's TCC grant and keep full behavior. Verified with a
     SIGKILL-on-Documents-access sandbox: daemon boot (30 s soak), proxy boot,
     and the full lean-ctx update rewire now run clean against a real data dir
     with 600+ sessions rooted under ~/Documents.
• Pi: ctx_grep/ctx_find/ctx_ls silently searched the wrong directory
  (#395): path was optional and fell back to the extension's cwd, so an
  agent working elsewhere got results from the wrong tree and was derailed;
  the calls also rendered without their arguments. path is now required
  (schema + description make the scope explicit), and the three tools reuse
  Pi's native call renderers so every invocation shows its pattern and
  directory in the transcript.
• OpenCode × ChatGPT-OAuth broke behind the proxy (#366): proxy enable
  exported OPENAI_BASE_URL without the /v1 suffix the OpenAI SDK convention
  expects (default is https://api.openai.com/v1). OpenCode therefore sent
  Responses-API calls to …:4444/responses — a path its ChatGPT-OAuth plugin
  does not recognize (it matches /v1/responses), so subscription traffic
  leaked through the proxy to the platform API with the wrong credential:
  "Missing scopes: api.responses.write". The shell exports and the Codex CLI
  config now advertise http://127.0.0.1:<port>/v1; with that base, OpenCode's
  OAuth plugin correctly routes ChatGPT-subscription requests directly to
  chatgpt.com (analogous to the Claude Pro/Max guard), while API-key traffic
  keeps flowing through the proxy. Stale /v1-less entries in Codex
  config.toml are migrated on the next proxy enable; the proxy also
  collapses accidental /v1/v1/… double prefixes from clients that append
  /v1 themselves. Verified end-to-end against OpenCode 1.2.15.
• Dashboard token race: lean-ctx dashboard persisted its fresh auth token
  before binding the port. Two racing starts both wrote dashboard.token;
  the bind loser exited, leaving a token on disk the surviving server never
  accepted — every "already running" browser open then hit silent 401s. The
  token is now saved only after a successful bind.
• Live Activity feed masked errors as "No events recorded yet": a failed
  /api/events poll (daemon restart, expired token, timeout) was rendered as
  an empty feed. The dashboard now keeps the last known events and shows the
  actual error with a recovery hint instead.
• Status bar showed "No session" while agents were active: /api/session
  matched sessions against the dashboard process's own cwd (usually HOME — a
  broad root that rightly matches nothing). It now falls back to the most
  recently updated session rooted in a real project.

Performance

• /api/events no longer re-parses the event log on every poll: the
  file-backed event load is cached on (path, mtime, length) — the 3-second
  dashboard poll now costs a stat() instead of reading and parsing up to
  10k JSONL lines.

Upgrade

lean-ctx update                 # recommended (auto-downloads + refreshes shell hooks)
cargo install lean-ctx          # or
npm update -g lean-ctx-bin      # or
brew upgrade lean-ctx

Note: After upgrading via cargo/npm/brew, run lean-ctx setup to refresh shell aliases. lean-ctx update does this automatically.

Full Changelog: v3.8.2...v3.8.2