Issues in. Merged PRs out.

1. 01

   Triage

   Deterministic lint rejects bodies under 80 chars or without acceptance criteria; the model then picks one of four labels. wontfix is only ever proposed.

2. 02

   Claim

   The dispatcher re-reads the issue, assigns itself, takes a per-ticket flock, and creates .factory/wt-<n> on agent/<n>.

3. 03

   Work

   The worker gets the issue, comments, standing instructions, a per-ticket brief, and on retries the previous gate report or reviewer findings.

4. 04

   Gate

   Conflict markers, your [[gate.check]] list in order, then a leak-scan of added lines. Exclusive checks serialise on a host-wide lock.

5. 05

   Review

   The reviewer sees the diff, issue, and gate report. REVISE returns to the worker; APPROVE adds factory-approved.

6. 06

   Merge stage

   One PR per pass, requiring gate PASS, factory-approved, green GitHub checks, and a head containing the configured integration branch tip (main by default).

7. 07

   Escalation

   Unresolved work gets ready-for-human with a reason, logs, and handoff evidence. The worktree stays for forensics. Human intervention takes precedence over automatic recovery.

8. 08

   Manage · opt-in

   When configured, the manager revisits untouched escalations within a round limit. Factory applies a closed decision menu and records the result; unresolved cases stay with a human. The configured executable is trusted, not sandboxed.

1. Latest release v0.4.6 October 23, 2026

   v0.4.6 release notes.

   • Gate checks replace invalid UTF-8 bytes in captured output instead of crashing before the gate report is written (#361).
   • Structured schema-1 decision records now capture triage, initiative binding, scheduler admission, worker selection, and claim alternatives, inputs, rationale, causes, actor, and model cost in the rotating journal. `factory explain N [--json] [--since ISO]` projects those rows plus legacy audit events into a bounded decision graph with cross-ticket edges and coverage gaps (#356).
   • Automatic release PRs now keep public installation guidance current (#351): `factory release` updates and stages the version file, `CHANGELOG.md`, the fixed-version install pins in `README.md` and `docs/index.html`, the project page's latest-release date, and a new page history entry built from that version's changelog notes. Every expected documentation spot must match exactly once or the release is refused.
2. Release v0.4.4 October 22, 2026

   v0.4.4 release notes.

   • Worker and reviewer capabilities are explicit instead of tied to an `omp` executable name (#334): `[workers.<label>].transcript = "omp-session"` captures compatible JSONL for token usage, scoring and learning, while `[review].json_mode = true` enables OMP-compatible JSON review parsing. Commands named `omp` retain both defaults, undeclared workers still report usage as `n/a (not reported)`, and `factory stats --by-worker` now includes reported worker token totals.
   • `factory watch` no longer reports repeated `review_revise` escalations as engine defects (#324): every `review_revise` is a ticket outcome, whether bounded review rounds are exhausted, review after a manager `FIX` still requests changes, or fresh review after a merge-stage refresh/rebase requests changes.
   • A refresh conflict onto moved `main` no longer escalates (#295): the merge stage withdraws `factory-approved`, journals a `conflict-requeue` row (PR, head, conflicting files, round) and releases the ticket, removing `ready-for-human` and adding `ready-for-agent` (including after a manager `FIX`), so the next claim runs a worker in the same worktree on the same `agent/<n>` branch with a brief to rebase onto current `origin/main`, resolve the conflicts, keep the approved intent and not widen scope; the normal gate, lease push, independent review and round limits follow, and only a fresh approval restores the label. It still escalates as `refresh_conflict`, with the conflicting files in the packet, for a protected path without `factory-protected-override`, after `review_rounds` conflict requeues, or when the ticket is not open. A manager `FIX` whose kept worktree is ahead of the PR head with a head a journaled worker attempt produced (#264) pushes it with a lease on the recorded PR head and goes on; any other head escalates as `refresh_head`, a refused lease as `lease_refused`.
   • Automatic releases publish the GitHub Release (#320): after pushing the `v*` tag and `stable`, `factory release` runs `gh release create <tag> --verify-tag --latest` titled with the tag, with that version's `CHANGELOG.md` section as notes. A Release that already exists is left alone; a `gh` failure is journaled as a `release` row with outcome `publish_failed` and retried at the start of the next pass without rolling back the tag. v0.4.2 had stayed behind v0.4.1 as GitHub's latest until it was published by hand.
3. Release v0.4.1 October 16, 2026

   The factory watches itself.

   • Post-ship monitor: factory watch turns the factory's own runtime failures into bug issues that enter ordinary triage; nothing waits for a human to notice.
   • Session evidence: every OMP worker transcript is kept per ticket and factory score rates it for efficiency and compliance, the raw material for self-improvement.
   • Escalations as data: every escalation carries a mechanical code; factory stats counts failure classes per week so a repeated code becomes a factory defect, not another human decision.
   • Release tier: when = "release" gate checks run only on tags, so main and a release get different scrutiny.

   The decision: every routine escalation that lands on a human is a bug. The loop now records enough to find and fix them.

4. Release v0.4.0 October 11, 2026

   One agent, many factories.

   • MCP server: factory mcp exposes repository discovery and the four evidence reads to any stdio MCP client across every registered repository. Read-only; no mutation tools, resources, or network transport.
   • Outcome-control workspace: one packaged operator application with live Flow, attention and outcome evidence, persistent FM chat, Atlas ticket-flow reference, and revision-checked Settings. Direct /api/act is removed; human actions use expiring previews and restart-safe receipts.
   • Meta loop: automatic factory learn, report-only mutation testing as a pr-tier gate check, candidate escaped-defect attribution, and human-correction metrics.
   • Dispatcher identity: a separate machine account for the dispatcher; workers and reviewers never hold a GitHub credential.

   The decision: the factory's state is GitHub plus its journal, so the agent interface is a read surface, not a control plane. Writes stay on the reviewed ticket intake.

5. Release v0.3.8 October 3, 2026

   Run lean. Keep the record.

   • Bounded journal: idle passes no longer journal parked tickets, and events.jsonl rotates into gzip segments while every outcome row stays readable.
   • Lighter dashboard: the Ops snapshot bounds its executions list, and a docked ticket drawer replaces the overlay on desktop.
   • Steadier sync: upstream sync lands the conflict-free prefix instead of parking the whole backlog behind one conflict.
   • Resume and wrap: workers resume from retained accepted results; a host-owned argv prefix wraps every worker launch.
   • Experiments: bounded observation, experiment-informed roadmap next actions, and a split-aware eval promotion contract.

   The decision: cap what the factory writes, never what it can prove. Rotation drops only closed-execution lifecycle noise; audit rows survive.

6. Release v0.3.5 September 18, 2026

   Collaborate with evidence. Operate from one console.

   • Retained results: independently accepted worker handoffs survive worktree cleanup with exact-head provenance, integrity, bounded lifetimes, and explicit incomplete states; readable through the evidence interface, console, and briefings.
   • Shared planning: initiative routing, execution guards, immutable bindings, durable routed handoffs, and shared roadmap/owner-attention reads are implemented. The real two-human acceptance checkpoint remains open; the collaboration programme is not yet proven delivered.
   • PR recovery: the manager frontier handles stalled Factory-owned PRs and late feedback with bounded recovery and exact-head approval.
   • External PR review: opt-in discovery, cited SHA-bound reviews, changed-head re-review, CI readiness, and the human queue are available. This lane reviews only; it never edits contributor branches or merges their PRs.
   • Operator UI: nine themes, title-first navigation, and Ops as the default; five local Settings controls; read-only dashboard Chat; and optional Codebase history.

   The decision: keep one canonical GitHub plan and one review-only contributor lane. Initiative execution is guarded; reading, filtering, or discussing evidence grants no mutation authority. Track the held two-human acceptance checkpoint.

7. Release v0.3.0 September 8, 2026

   Recover deliberately. Review with evidence.

   • Manager recovery: opt-in, bounded decisions for untouched escalations, journaled before GitHub changes. Failed mutations are not replayed.
   • Grounded review: blockers need a criterion, rule, or concrete defect; optional suggestions no longer masquerade as required fixes.
   • Safer merges: refresh from the remote PR branch when no local branch exists, reject empty refreshes, and preserve upstream-sync ancestry.
   • Authoritative runtime evidence: execution, child-process, wait, and handoff records; a read-only runtime projection; human-touch and per-label pass metrics.

   The decision: let the model recommend, let code constrain the action, and keep the evidence. A passing gate is not permission to ignore a real defect. Read the review-contract decision.

8. Foundation The original pipeline Before tagged releases

   Start with a ticket. End with evidence.

   Label-driven triage, one worktree per ticket, repo-defined gates, second-model review, and fail-closed merging established the core loop. The dashboard, handoff notes, and learning loop made that work inspectable.

   • Use the agent CLIs and deterministic checks the repository already has.
   • Keep operational state in GitHub and local event files, with a re-runnable dispatcher rather than a hosted control plane.

   The decision: automate the implementation, not the definition of success. Your acceptance criteria and checks remain the contract.

Full changelog · Tagged releases · Changes since v0.3.8

Also exploring: a tested local-model configuration for triage and review. Still an unscoped idea, not a zero-cloud guarantee; the pipeline continues to use GitHub.

Have evidence, a use case, or a different tradeoff? Join the linked issue discussions. Plans can change as implementation teaches us more.