Skip to content

mugiwara v2.0.0

Latest

Choose a tag to compare

@github-actions github-actions released this 02 Nov 08:08

29 changes since v1.0.5

Gate

  • Housekeeping Retarget gate-selftest mutations to the refactored code; call slow suites directly 4fce35e

Build

  • Fixed Rebuild hooks with the bun version CI pins (1.4.2); the stale-build error now names it 6eaafc7

Install

  • Fixed Wire pipeline-guard to Stop, not only SubagentStop; doctor checks it; update repairs older installs 175e802
  • Fixed A Claude or opencode dry run says its file count leaves out the last-step files c35b775

Hooks

  • New Opt-in MUGIWARA_HOOK_DEBUG log of each stop, to diagnose headless closing 9920c31
  • Fixed Run the closing check before the stop_hook_active return so heal cycles and close are not cut to one hold 3f5c28e
  • New Mode=auto heals open findings before closing; guided and semi are never pushed or closed; fix verdict match on '0 blockers' 846a2ec
  • New Close a mode=auto mission when the hold budget is spent so a headless run ends with report.md 61ab2e9
  • Fixed Hold an explicit /mugiwara session only for a full-lane or risky diff, matching the fast path 9d29752
  • New Indonesian and path risk signals, exact next step for review, archive hint 75134c2
    • Risk detection moves to src/risk.ts with Indonesian terms and path signals. The reminder names Robin and Jinbe; the stop hook says which review agents are missing and when closing files exist without report.md. archive records absent review evidence as Not verified and exits 1 for an unknown mission.

Close

  • Fixed Hold the turn once both review files exist; close archives without plan.md and lists it under Not verified 970e22b

Hooks Doctor

  • New Hold the turn so a headless run reaches mugiwara close; doctor warns when no global CLI is on PATH 4629583

Content

  • Fixed Review and security commands point at .mugiwara/missions//, not the pre-0.7 layout 926df97
  • Refactored Merge 21 skills into 14, move eval-runner out of the package f5b954a
    • Orchestration absorbs workflow, resume and lessons; execution absorbs git; quality absorbs gates and testcases; healing absorbs root-cause. Old features= tokens stay as aliases. Agent and skill prose rewritten for plain reading, and a risk-overrides-size rule added to orchestration.

Cli

  • New Record writes flows evidence from real facts; close records then archives 2071908
  • New Reset removes one mission; risk_words= adds project risk words to the prompt hook 75f28e1
  • Fixed Warn about duplicate config keys and mode mismatch, record risky missions as full lane 25b09cb
    • Found by trying the install as a new user. A key set twice kept the first value silently and doctor said all clear; a mission keeps the mode it started in, so editing the config did nothing and nothing said why; status showed lane direct for an auth mission. doctor, status and continue now say so, and the risk reminder asks for --lane full in the mission record.
  • New Add mugiwara doctor and clear errors for corrupt mission state a0a4754
    • doctor checks installs, hooks, config, lane, ledger and the npx pin and says how to fix each failure, without writing files. continue, status, cost and archive now name the bad file and the fix instead of crashing or reading zeros. install warns about a stale command pin.

Ci

  • Housekeeping Install lifecycle smoke on three operating systems, dated host status 7164de0
    • smoke-lifecycle runs install, upgrade and uninstall for claude, opencode, codex and cursor in a temp repo and temp HOME. CI runs it on ubuntu, macos and windows. The host matrix and install page carry dated per-host status, and the troubleshooting runbook explains why the crew may not engage.

Engine

  • New Independent review, enforcement hooks, provider cost, small-request path 0a78422
    • Robin and Jinbe dispatch as real subagents on the full lane with a hook ledger and an archive closure gate; explicit /mugiwara sessions are held when the real diff is risky; a prompt hook reminds the crew about risk areas; small requests skip the pipeline. Stop hook records provider token usage. Config, savepoint, lane budgets, report verdicts, installers for Codex and Cursor, and tests updated to match.

Benchmark

  • New Defect tasks, calibration and two-arm comparison harness 11164e0
    • 22 tasks with hidden checks and reference/naive solutions, --calibrate, --tasks-from, --run-offset, --arm, --prompt-suffix and a Codex host option. Results recorded as anecdotes with their limits.

Docs

  • Remove UPGRADING.md (release notes live in the PR and GitHub release); give install-heavy tests a 30 s timeout 42ecae3
  • Fix inaccuracies found by checking every runbook, concept and reference page against the real CLI 1fe7019
  • Drop the last go/no-go wording, fix the Codex agent count, point README at doctor and archive d0f992c
  • Record the 10-run risky benchmark; say which evidence files archive expects 023946d
    • Three risky tasks, ten fresh runs per arm: hidden checks passed 12 of 30 for the crew against 1 of 30 plain, at about four times the cost, with one task carrying most of the gain. The archive artifact gate now names the flows/ files to write instead of a generic message.
  • Restructure docs and runbooks, rewrite README and add UPGRADING aa84a4f
    • Capabilities page, concept pages, single install page, practical runbooks, and measured-versus-not claims including the benchmark results.

Housekeeping

  • Release v2.0.0 840de4a
  • Merge pull request #65 from ionivetech/feat/audit-cleanup c06f6f5
    • Feat/audit cleanup
  • Cover active-mission and the benchmark scripts so the coverage gate passes 3a6be9e
    • Move the script bodies into exported main(argv) so tests run them in-process, since coverage is not counted from spawned subprocesses.