Skip to main content

Gitmoot v0.5.2

A large orchestration + runtime-robustness release: a smarter multi-agent "Orchestra" (goal-anchored synthesis, dependency dataflow, independent verification, human-in-the-loop pauses), more resilient agent/runtime management, and automated cross-platform releases. All changes are additive / opt-in or bug fixes — default behavior is unchanged.

What's new since v0.5.1​

Orchestra — multi-agent delegation​

  • feat(workflow): goal-anchored finale continuations (#418) — when delegated jobs finish, the coordinator's continuation now restates the original goal (resolved depth-stably from the root job) and synthesizes an answer to that goal — reconciling child conflicts and flagging gaps — instead of a generic "decide the next step." Termination semantics are unchanged.
  • feat(workflow): deps[] as dataflow (#419) — a dependent leg can now receive its succeeded upstream dependencies' results (decision, summary, PR, changes, artifact body) injected into its prompt, so legs build on each other instead of only knowing a dep finished. Opt-in via [orchestrate].inject_upstream_dep_context (default off; flag-off output is byte-identical).
  • docs(orchestra): verifier delegation pattern + built-in template (#421) — a new built-in verifier coordinator recipe for independent (different-model, read-only) checking of delegated work against the goal, generalizing the decompose-and-verify verify leg. Cross-evaluation beats self-grading; documented in both doc trees.
  • feat(workflow): result-aware non-progress loop detection (#339) — delegation trees that perturb the work set each round but produce no new verifiable side effects now trip the loop ladder after a configurable streak ([orchestrate].max_delegation_non_progress_streak, default 2), not just the blunt depth cap.
  • feat(workflow): escalate_human failure_policy (#340) — a delegation can pause the whole tree durably for a human (awaiting_human), @-tag them on the PR/issue + surface it in the dashboard Attention list, and resume via /gitmoot resume <jobID> retry|continue|abort. Zero tokens while waiting; a TTL auto-finalizes a never-answered escalation; paused time is excluded from the per-root wall-clock budget.
  • fix(daemon): pre-flight failures advance the parent DAG (#409, #413) — a delegation child that fails before its agent runs (checkout / branch-lock / permission / etc.) now still advances the parent so its failure_policy fires (including escalate_human), instead of stranding the tree silently. The checkout guard was also relaxed so a worktree-less delegation child isn't false-failed.
  • perf(db): index the delegation tree by root_id (#420) — root-scoped engine lookups now use an indexed root_id column instead of full-table scans that re-parsed every job, removing an O(n) cost on every child completion under large trees. Behavior is identical.
  • refactor(prompts): single-source the gitmoot_result contract (#422) — the agent JSON contract is generated from the Go structs at build time with a drift-guard test, so the prompt can't silently diverge from the code; the delegation field docs are now complete (15/15 fields).

Agents & runtimes​

  • feat(cli): gitmoot agent restart (#414) — rebind a registered agent to a fresh runtime session in place (preserving role / repo / template / capabilities), instead of remove-and-recreate; refuses while a live session or queued/running jobs are in flight.
  • fix(doctor): accurate Claude auth health (#414, #427) — agent doctor, gitmoot doctor, and plugin doctor now run a live Claude probe instead of false-failing when only an env token is missing (foreground works via cached credentials), with daemon-aware guidance that background jobs still need a token in the daemon env. The status is warn (not a hard fail) when the box is actually authenticated.
  • feat(cli): record owner PID+host on runtime-session locks (#425) — so gitmoot agent restart (and recovery tooling) can distinguish a live foreground session from a stranded lock.
  • feat(skillopt): recover --generation (#303) — reclaim a stranded SkillOpt option-generation lock (left by a killed train continue) and salvage the already-persisted options; --advance-state completes the iteration.

CLI, setup & comments​

  • feat(cli): tagging-ready gitmoot setup (#428) — setup now defaults to --watch-issues, so @<agent> ask … mentions on issues route to jobs out of the box.
  • feat(workflow): readable findings in posted comments (#429) — findings[] objects render as readable markdown in PR/issue comments instead of raw JSON.

Release & install​

  • ci(release): automated cross-platform releases + linux/arm64 (#412) — release binaries (linux/darwin × amd64/arm64, static, CGO-off) are now built, checksummed (sha256sums.txt), and attached by a GitHub Actions workflow on release publish; the one-line installer supports linux/arm64 (e.g. Raspberry Pi).

Install / upgrade​

  • Existing install: gitmoot update
  • Fresh install: curl -fsSL https://gitmoot.io/install.sh | sh