Skip to main content

Gitmoot v0.8.6

Token usage becomes trustworthy across every codex session shape, memory's confirmed pool starts filling from ordinary jobs, and the dashboard gains a Learning section that shows both halves of the self-improvement loop. Everything additive; defaults unchanged.

What's new since v0.8.5​

Codex token usage — captured correctly for every session shape (#658, #661, #662, #664, #669)

  • Fresh codex sessions now run with --json and record real token usage on the job. Resumed sessions report session-cumulative usage, so gitmoot tracks a per-session baseline and records only the delta each job actually consumed — no more multi-million-token phantom entries. Ephemeral/temp workers (which resume a Start-created session on their first job) are flagged single-use so their first delivery counts too. Proven by an isolated-home daemon E2E. Kimi 0.19.2 emits no usage data; documented as a known limitation (#659).

Memory: confirmed pool fills from ordinary jobs (#645, #647)

  • Enrolled agents' daemon-observed facts (review outcomes, fix-round counts) are now written on every job, not just orchestrate runs — the confirmed pool grows from normal daily work and feeds the prompt-injection tier. [memory] is now advertised in the gitmoot init config template (#653).

Session jobs — record here-method work as tracked jobs (#660, #673, #676)

  • gitmoot job open/close/record lets interactive ("here-method") sessions register their work as first-class tracked jobs, so dashboards, usage, and memory see it. agent prompt --record accepts bare template ids.

Daemon: reboot recovery via boot_id (#651, #668)

  • On startup the daemon compares the machine boot_id: after a reboot it recovers in-flight jobs and reclaims runtime locks that would otherwise strand for their full TTL.

Runtime: config-driven metadata registry + resilient model resolver (#674, #652, #679)

  • Runtime metadata (binaries, flags, model lists) moves to a config-driven registry, and default_model resolution is per-section and resilient — a bad model entry in one runtime section no longer breaks the others. Model-resolution rules are now documented.

Workflow/orchestrate hardening (#649, #666, #654, #667, #665, #671, #675)

  • Delegation job budgets are projected before a batch enqueues (no more over-budget batches that die halfway) (#666); read-only fan-out children are warned their worktree is the committed tip (#667); session isolation + task-run recovery fixes land contributor work by @plotarmordev (#671); heartbeats gain a policy-gated implement action with per-heartbeat runtime override (#675).

Web dashboard: Learning section + galaxy/mobile polish

  • New Learning section with two tabs: Skills (SkillOpt template evolution — versions, canary state, candidate reviews with scores) and Knowledge (the memory brain graph: agents → categories → facts with evidence, in an Obsidian-style layered layout) (#672, #677, #680). Agent pages show per-agent config and memory pool counts (#670). Galaxy layout no longer collapses into a ring after polls or filter round-trips (#648, #678). The whole UI is mobile-responsive (#644), and the agent panel diffs template versions (#643).

Docs

  • Dedicated Dashboard section with screenshots on gitmoot.io (#646); AGENTS.md operating manual (#656); agent-template publish/pull distribution clarified (#655).