openwolf

OpenWolf 2.5.2 adds saved task handover, more reliable project memory, recorded usage by model and quiet session notices.

Changes

• Transfer selected saved context between Claude Code and Codex with explicit handover packets and project checks.
• Recover task checkpoints after supported session restarts and compaction.
• Archive eligible old notes with restore pointers. Preserve active, latest and pinned notes.
• Preserve project-map entries during partial scans and improve updates after file changes.
• Reconcile available Claude, Codex and OpenCode token counters. Estimate API costs using the actual provider and model, with missing data shown.
• Prepare compatible stable hook and plugin updates for new sessions. Major upgrades remain notification-only by default.
• Show short notices for completed OpenWolf work and synchronize handover, usage, memory and activity in the dashboard.
• Fix concurrent writes, missing hook dependencies, malformed settings handling and Windows runtime and daemon path issues.

Upgrade

npm install -g [email protected]
openwolf update --dry-run
openwolf update

Restart project daemons and start new agent sessions. Automatic runtime updates do not replace the global CLI or a running daemon.

Checks and limits

All 322 regression tests passed on Linux, macOS and Windows. Package installation, 2.5.1 upgrade and daemon checks passed, including Node 20. Main validation.

Native Codex recovery through resume and compaction, OpenCode notifications and Grok startup were checked. A full live Claude to Codex to Claude coding handover and broader token-efficiency comparisons remain unverified. Protected durable instructions require independent administrator setup. Handover imports remain explicit. API cost estimates are not subscription charges.

Credits

Thank you to the reporters, PR submitters, code authors and reviewers. Contributor credits and the issue and PR audit record each role and distinguish adapted work from merged PRs.

Full changelog · Validation details

[2.5.0] - 2026-08-22

Positioning, and the removal of the last thing that made "no API calls" untrue.

Added

• Cost. The dashboard prices measured usage at Anthropic's published rates, per model, split into cache reads, output, cache writes and fresh input. On one real 277-call project that is $84, of which 67% is cache reads: a number that was previously visible only as "102.3M" with no way to size it. Labelled as list price throughout, because a subscription user is not billed per token.
• Overhead as a ratio. "1.4K injected" now reads "1.6% of what it kept out", which is the claim the footnote on openwolf.com already makes.
• Governor breakdown by command family. The classifier always knew the family at the rewrite point, but the rollup threw it away and kept only scalars. Sessions now carry bash_governed_by_family, folded into a lifetime map that survives session trimming, so the dashboard can answer which families are paying off and which are set to suggest.
• A plain-language panel. Not everyone using OpenWolf reads ledgers. One paragraph states how many tokens were kept out of context, what that would have cost at the ceiling, what OpenWolf spent doing it, and the net.

Changed

• Reads and writes were two numbers glued with a dot; they are now "edits per file read", with the raw counts underneath.
• A sub-30% anatomy hit rate is called out in the accent colour instead of sitting in a neutral tile: it is the most actionable number on the page.
• The "saved · denied re-reads" tile no longer renders a dead "n/a" in warn mode. Warn mode says what it caught and how to switch to deny.

Removed

• AI suggestions and every other model call. The ai_task cron action shelled out to claude -p for two weekly tasks (cerebrum reflection and project suggestions). It was the only path in OpenWolf that reached a model, and it contradicted the one claim the whole project rests on. Gone, along with the Insights dashboard panel, suggestions.json, and the cron.use_claude_p / cron.api_key_env settings. openwolf update deletes the leftover file; a stale manifest entry now fails loudly instead of silently reaching the network.

Fixed

• Windows hooks, properly this time. 2.0.4 swapped $CLAUDE_PROJECT_DIR for %CLAUDE_PROJECT_DIR% on Windows, which fixed nothing: PowerShell is Claude Code's shell there and has no %VAR% expansion, and the variable is not in the hook environment on any platform to begin with (project context arrives through the stdin JSON payload). Every hook died with MODULE_NOT_FOUND on every tool call. Hook commands now carry the resolved absolute path, so no shell expansion is involved at all. Hooks also derive the project root from their own location when no environment variable is set. Reported in detail by @aevnar.
• A malformed buglog no longer takes anything down. A .wolf/buglog.json written as a bare array of entries (the natural shape when an agent writes the file by hand) reached bugs.length unguarded. In one real project the pre-write hook failed on 465 consecutive invocations with nothing but the heartbeat noticing, openwolf bug search threw, the SessionStart file index was silently dropped, and the dashboard rendered a blank page. Both shapes are now normalised at the boundary, in the hooks, the CLI, and the dashboard loader.
• Dashboard panels render inside an error boundary: one malformed state file degrades a single panel instead of blanking the page.
• detectAgent() recognises Claude Code by CLAUDECODE, which is actually set, instead of CLAUDE_PROJECT_DIR, which is not. Sessions were being attributed to "default" and losing their per-agent ledger rows.
• openwolf update now also removes hand-patched hook entries written with native Windows path separators, which previously survived the filter and ran alongside the new on

[2.5.1] - 2026-08-30

Sixteen reported defects, fixed. Fifteen were found and reported by @davdittrich, one by @krsfer, each with a reproduction. Every fix carries a regression test that was first run against the unfixed code and observed to fail.

Fixed

• daemon stop could terminate an unrelated process (#78). The fallback path sent SIGTERM to every PID listening on the configured dashboard port. Port occupancy is not ownership. The daemon now records .wolf/daemon.pid after a successful bind, and stop signals a process only when the record was written on this host, for this project root, for a live PID that still looks like the daemon. Anything else holding the port is reported and left alone.
• An existing dashboard token kept its permissions (#79). New tokens were written 0600; a token file that already existed at 0644 was returned as-is and stayed readable by other local users. It is now repaired in place, never rotated, since a readable token is still the token a live browser tab holds.
• Read tracking escaped the project root (#80). Containment was a startsWith check, so ../project-private/secret.ts shares a prefix with ../project and was recorded in that project's session state and token metrics. The Bash read channel had no root check at all. All three hooks now share one lexical path.relative check, with a realpath fallback so a project reached through a symlink is not silently dropped.
• Codex install destroyed a malformed hooks.json (#81). A parse failure was caught and OpenWolf-only defaults were written over the user's file, reporting success. The file is now left byte-identical, hooks are not reported as registered, and one actionable warning explains how to recover.
• Parallel hooks lost session state (#83). Atomic writes prevent torn files, not lost updates: two hook processes that each read, modify and write both produce valid JSON, and the second erases the first. Measured at 60 concurrent post-read hooks, 6 of 60 reads survived. Every shared-state read-modify-write now runs through one bounded lock-backed transaction that reads inside the lock. 21 call sites across 13 files.
• Concurrent SessionStart lost ledger increments (#84). 60 concurrent startups created 60 session files and counted 31.
• Cron state writes bypassed the lock after timeout (#86). The fallback wrote anyway once the budget expired, which defeats the lock for every writer that respects it. Contention now fails closed and reports itself.
• Concurrent registry updates lost projects (#88). 60 concurrent registrations kept 31. openwolf init in several projects at once is exactly this shape.
• Concurrent OpenCode sessions shared one state file (#89). Every handler received a session id but all of them persisted to hooks/_session.json, so a second session overwrote the first and later reads from either mutated the survivor. State is now per session id, with the legacy file kept only when no usable id exists, and session files are garbage collected after 7 days.
• The Bash governor pointed at a deleted log (#82). Full output was written, the cache was pruned, and the model was then handed a pointer to a file the prune had just removed, described as "preserved verbatim". Output is now verified present before anything claims it was preserved; an output larger than the whole cache budget says so plainly instead.
• A skipped anatomy write still advanced freshness (#85). On lock contention the scan wrote nothing but recorded the current Git HEAD in _scan-state.json, so the index was reported current when it was empty and the rescan that would have fixed it was suppressed.
• Unknown cron task IDs reported success (#87). runTask logged a warning and resolved, so the API returned 200 and the CLI printed "triggered" for a task that does not

Fix from the pre-publish end-to-end test: the anatomy scanner no longer indexes agent-config directories (.claude, .codex, .opencode, .gemini, .cursor). Harness config is noise in a code index and was even topping importance rankings in real projects.