This is an early release preview. You may encounter bugs.
OpenWolf logo
Unclaimed

Tool memory monitoring

OpenWolf

OpenWolf treats Claude Code's token burn as a caching problem … six hook scripts intercept every read/write, maintaining

A 80/100 GitHub score ? This grade is derived from GitHub signals, not user votes. Open for the full breakdown.
No votes yet

01 / About

What OpenWolf is.

OpenWolf keeps one project memory shared across several coding agents and reports what each session cost. It installs lifecycle hooks into the agents it detects, writes its state to a .wolf/ directory in the project, and runs as plain local file I/O — no network calls and no telemetry.

The hooks address three recurring costs in agent sessions. Oversized Bash output is condensed by command family before it enters context, with the full text kept on disk behind a pointer and test and build output left alone by default. A durable project index lets openwolf find locate a symbol in about a thousand tokens and gives large files a symbol map with line ranges so an agent reads one function instead of the whole file. Corrections, conventions, and bug fixes are written to files that survive sessions, travel through git, and are re-injected after compaction alongside in-flight state and path-scoped instructions.

Measurement is read back from the harness transcript rather than estimated: API calls, output tokens, cache reads and writes, the governor's per-call delta at the rewrite point, and cache-rebuild events attributed to a trigger such as a model switch, compaction, version change, or expiry. openwolf report prints those figures and openwolf bench runs the same tasks with and without the hooks against a fixture repository, spending real API budget. A local dashboard binds to 127.0.0.1 with per-project token auth and shows hook health, per-agent breakdowns, and the anatomy browser.

Agent Integration
Claude Code Full: 12 hooks, output governor, skills, verified measurement
Codex CLI Core hooks via .codex/hooks.json and AGENTS.md: session, read, write, compaction, stop
OpenCode Native plugin plus AGENTS.md: session and tool before/after
Cursor Rules file (context only)
Gemini CLI GEMINI.md block (context only)
Antigravity AGENTS.md block (context only)

Features

  • Shared project memory: one .wolf/ directory holds conventions, handoff, bug log, and index for every wired agent
  • Bash output governor: output over roughly 2,000 tokens is condensed structurally, with the verbatim original cached on disk
  • Project index: file descriptions, sizes, symbols, and an import graph, rebuilt with openwolf scan
  • Symbol lookup: openwolf find returns ranked results within a token budget; --file returns one file's symbol map
  • Compaction recovery: state, rules, and scoped instructions are re-injected when the session compacts
  • Rule cadence: top rules are repeated in a short note every 25 tool batches
  • Transcript-based accounting: per-session, per-model usage, hook delivery verification, and cache-rebuild attribution
  • A/B benchmark: openwolf bench measures each token dimension, completion rate, and bash re-run rate with and without the hooks
  • Local dashboard: token-authenticated, localhost-bound, showing usage, hook health, activity, and cron control
  • Bundled skills: /handoff, /security-audit, and /reframe are installed for every wired agent
  • Hook health: every hook writes a heartbeat, and openwolf update selfchecks each hook after install
  • Safety defaults: no shell interpolation, no auto-approval of tool calls, and secret-bearing files excluded from the index

02 / Discussion CREDIBILITY-GATED

Discussion

Reading is open to everyone. Posting and voting need a verified identity or a GitHub grade of B or higher.

  • No discussions yet.

03 / Build

Build with OpenWolf.

Browse the catalogue for frameworks, tools, and harnesses, each scored on real GitHub credibility.

Get OpenWolf →

Browse the catalogue