v1.37.0

Added

• ASHLR_MCP_HOST now recognises gemini-cli and cursor, so the startup banner and host gates report the real client instead of generic. docs/multi-host-mcp.md gains a Gemini CLI (~/.gemini/settings.json) section. Behavior is unchanged for every existing host.

Changed

• ashlr__edit returns the header only for any exact, unique match, not just edits under 120 characters, because the removed/added echo is the caller's own text played back. Fuzzy matches and replace-all edits keep the detail lines. On the synthetic in-repo benchmark at the same commit (bun run scripts/run-benchmark.ts): medium edit 52.0% to 88.2% saved, large edit 96.5% to 99.0%, edit mean 49.5% to 62.4%, overall mean 80.7% to 82.3% (the overall change is the edit scenarios only; read samples are identical). The benchmark gain is by construction, because the benchmark's edit scenarios mirror the server's summary format; the real effect is about 37 estimated tokens on a roughly 400-character edit (51 down to 14, synthetic estimate).
• SessionStart and PostToolUse hooks run from one node launcher per event (hook-bootstrap.mjs --group) instead of one per hook, and the launcher no longer starts an extra bun --version probe per hook. Each hook still runs as its own bun process, in parallel, with outputs merged. Measured in a Linux container: CPU per native Read 579 → 320 ms, per Edit 712 → 319 ms, per SessionStart 634 → 362 ms; wall time per Edit 212 → 154 ms.
• Benchmark grep figures are now measured against plain rg -n output, which is what an agent would otherwise read, instead of ripgrep's much larger JSON stream. Greps save about −27% (was about −50%); the cross-language headline is TS −60% · Python −63% · Rust −55%. README, docs and site figures are updated from the regenerated docs/benchmarks-v2.json.

Fixed

• ashlr__edit strict mode now counts overlapping occurrences, so a search such as }\n} in }\n}\n} is refused as ambiguous instead of being applied to the first one under a header-only 1 of 1 summary.
• The tier-2 fuzzy match in ashlr__edit and ashlr__multi_edit no longer swallows the line ending after the matched lines (which joined the next line onto the replacement, and dropped \r\n in CRLF files). Fuzzy edits always show the removed/added lines, and "removed" is the file text actually replaced rather than the caller's search.
• The first-start Bun install no longer prints the installer's output on stdout, which for ashlr-mcp and the plugin MCP servers is the protocol channel; it goes to stderr. ashlr-plugin, ashlr-mcp and the MCP bootstrap now die by the signal that stopped Bun (exit 130/143 in a shell) instead of exiting 0.
• npm i -g ashlr-plugin without Bun no longer leaves ashlr-plugin and ashlr-mcp failing with /usr/bin/env: 'bun': No such file or directory. Both package bins now start under node: ashlr-plugin finds Bun on PATH or in ~/.bun/bin and otherwise says how to install it, and ashlr-mcp goes through the same bootstrap as the plugin manifests (it installs Bun on first start unless ASHLR_NO_AUTO_INSTALL=1). A Bun installed to ~/.bun/bin but not yet on PATH now works without opening a new terminal.
• The server no longer fails with SQLITE_BUSY or "duplicate column name" when two processes open a fresh database at once (for example the server and the issue-token CLI). It waits on the write lock and applies its schema inside an IMMEDIATE transaction. This was the intermittent cloud-sync-flow integration failure.
• Background ashlr__bash sessions, test watchers and hook-timing locks no longer treat an exited but unreaped (zombie) process as alive. Containers whose PID 1 does not reap orphans kept such entries forever.
• ashlr__grep without a genome index (and ashlr__grep in the efficiency server) returns ripgrep's plain path:line:text matches, cut only at whole lines with a count of what was left out. It used to return ripgrep's --json stream cut mid-object, which agents could not read.
• MCP tools now work on the project the host launched the server in. The launchers moved the router to the plugin's install folder, so a relative path such as ashlr__read {"path":"src/a.ts"}, and tools called with no path (ashlr__tree, ashlr__ls, ashlr__test), resolved against the plugin itself unless a session-start hint pointed elsewhere. The launchers now record the launch folder in ASHLR_LAUNCH_CWD, and the router moves back into it.
• Hooks running under Claude Code are no longer treated as Codex hooks. Detection keyed on CLAUDE_SESSION_ID, which Claude Code does not set, so hooks ran in nudge mode and suggested bare ashlr__* tool names. CLAUDECODE=1 now marks a Claude Code run.
• The SessionStart hook exits when its work is done, in about 0.6 s instead of a fixed 2.1 s. Three timeout timers were never cleared and kept the process alive.
• ashlr-plugin stats --json prints an empty ledger as JSON before any savings exist, instead of a plain-text error, and --since reads the ledger's per-day totals (lifetime.byDay); it used to return an empty result every time.
• ashlr-plugin --help describes each command, and ashlr-plugin doctor runs codex-doctor instead of failing as an unknown subcommand.
• When ripgrep is stopped early (15 s timeout, or output past the 64 MB / 10 MB buffer), the ashlr__grep fallback now says so and gives its counts as lower bounds, and drops the line cut mid-path. It used to state how far ripgrep got as the total number of matching lines.

CI

• All Linux checks run in one job, and macOS and Windows checks in one job per OS. A PR now waits for two runners in sequence (checks, then CI Gate) instead of four, and uses two runners instead of eight. All check steps still run and report even when one fails.