factory 0.3.8

Compared against 0.3.5 (e3abd2b). Released channel stable points at v0.3.8. Versions 0.3.6 and 0.3.7 were not released.

• Idle dispatcher passes write no lifecycle rows for parked tickets (#130): a manage/landing check that changes nothing no longer journals its enter/lock/handoff/exit trail; passes that act keep the full trail. (PR #133)
• Bound the dashboard snapshot's executions list (#134): Ops no longer loads hundreds of MB on long-lived repositories.
• Roadmap stops reporting runtime_unavailable for closed or runtime-known tickets (#137).
• Dashboard docked ticket drawer: desktop docks the detail drawer beside the queue; narrow screens keep the modal overlay with focus trap, Escape and reduced motion. (PR #143)
• Upstream sync lands the newest conflict-free upstream prefix instead of parking the whole backlog behind one conflicting commit. (PR #117)
• Split-aware eval promotion contract (#126): docs/eval-promotion-contract.md and factory learn --promotion claim.json; keep only when the held-out split improves. (PR #128)
• Experiment evidence in a fresh FM session (#99); reviewer-calibration corpus grown to 11 cases with confirmed oracles and adjudication rows (#34, PRs #144–#148).
• Repository gate runs uv run --frozen --no-managed-python, so declared dependencies are installed and uv.lock is never rewritten (PRs #127, #129); the interpreter-escaping test fixture is portable to relocatable Python builds (PR #135).
• Journal rotation (#131): .factory/events.jsonl rolls into gzip segments events.jsonl.N.gz under the journal flock once it passes [journal] max_mb (default 64; retention segments, default 8). Open-execution lifecycle rows stay live; every outcome row stays readable via lifecycle.read_events, which now includes segments. Only closed-execution lifecycle rows of expired segments are dropped; the oldest kept segment grows with carried outcome rows.
• Host-owned worker argv prefix (#3): [defaults.worker_wrap] / [repo."owner/name".worker_wrap] command is prepended to every worker launch (all labels); a committed [worker_wrap] fails to load. doctor checks its executable on PATH. A launch prefix, not sandboxing.
• Bounded worker boot/resume context (#89): factory dispatch's worker prompt gets a ## Resume context section, built only from the latest retained accepted result (#88) and the local plan-bound journal (#57), when a ticket has prior retained history. It reports the prior accepted head/result and whether the admitted scope has since moved (unchanged/changed/unavailable); never a worker log, prompt, or transcript, never a scope rewrite, and no widened context budgets.
• O1 bounded observation producer (#97): stdlib-only experiments/o1-observation/observer.py (start/status/stop/report) with frozen protocol.json. Opens .factory/events.jsonl O_RDONLY, freezes a chunk-digested byte range (256 MiB cap, suffix + explicit coverage beyond it) before classifying the latest 12 recorded merged/escalate ticket executions by exact structured reason codes only. Runs land in exclusive, never-overwritten directories under the main root's .factory/experiments/o1-observation/, capped at 10 minutes, 64 MiB output and three iterations per lineage; no subprocess, network or model calls. Disposition starts pending.
• Experiment-informed roadmap next actions (#100): each roadmap/initiative plan (CLI evidence investigate, FM fm_investigate, dashboard Roadmap) gains experiments — the latest retained run of every experiment whose protocol.json names that initiative as roadmap, with its decision link, latest result, per-run recorded disposition and accepted/rejected/deferred/pending grouping, marked informed_by and never a blocker — and next: the next action (implement or decide) with a cited explanation from accepted NOW/NEXT/THEN/LATER Plan tiers, eligibility and real Blocked by edges, plus separate current-priority decisions, runnable, blocked, in-flight, later-held and displaced work and unresolved evidence. Read-only.

factory 0.3.5

Compared against 0.3.0 (289cac5). Released channel stable points at v0.3.5.

• Retain independently accepted worker handoffs outside worktrees (#88): exact-head provenance and integrity, 90-day content/365-day metadata, 256 KiB artifact/1 MiB bundle/256 MiB repository limits, explicit incomplete/privacy/expiry states, and fail-closed cleanup when retention fails. Add local paged kind: result evidence reads and console/briefing consumption without widening context budgets, transcript ingestion, or execution authority. (PR #113)
• External pull-request review lane (#5–#9): discover open, non-draft PRs opted in by needs-review or a review request, preserve that opt-in after viability verdicts, publish cited reviews bound to the recorded head, re-review changed heads within bounded rounds, record fail-closed required-CI readiness, and surface the six-state Ops queue plus actionable Inbox rows. The lane never edits or pushes contributor branches and never merges their PRs. Viability replay remains keyed to the label-add event, not the PR head.
• Bundle Cyberpunk, GPUFlo, District, Factory, ROCm, Porcelain, Sandstone, Slate and Forest dashboard themes with a persistent browser-local picker, preserving configured repository CSS as the default. Keep the title/status row above the consistent navigation row in a shared sticky header. Open Ops at the root without automatic FM briefing requests; explicit Inbox and existing deep links remain available. (PR #110)
• Add dashboard Settings for exactly five repository controls: concurrent tickets, time limit per ticket, worker/gate attempts, reviewer revision rounds, and Factory Manager model. Show effective values and sources, require an explicit reviewed save, reject stale writes, preserve unrelated TOML comments, and keep changes local and uncommitted. TOMLKit is the sole required Python dependency on current main; it is loaded only by settings write helpers. (PR #111)
• Shared roadmap and owner attention (#58): dashboard and read-only FM consume the same initiative/routing/drift producers. Owner/team filters select questions without changing canonical plans or granting authority; source failures, unknown ownership, missing baselines and assigned-but-non-runnable human takeover stay explicit. C1 adds roadmap, initiative and retained-drift reads with citations; declared stage and implementation closure never prove owner-confirmed delivery. Complete-content revision identity and historical accepted evidence remain intact. No publication, intake release, deployment or live two-human acceptance is implied.
• Immutable initiative bindings (#57): linked tickets carry a schema-1 baseline of the complete initiative's normalized Outcome, Boundaries, Plan, and Success evidence; admission pins the full human-approved ticket snapshot for every retry and fails closed on malformed, changed, incomplete, or unavailable sources while legacy unlinked tickets remain unchanged. Read-only factory plan baseline and factory plan drift expose proposals and accepted-versus-live drift without mutation or inferred attribution. Manager REWRITE preserves the exact accepted binding, and SPLIT preflights the live source and every child before creating intake-ready work; models cannot rebaseline.
• Correct routed-handoff integration (#56): neutralize incidental mentions in explanations while preserving validated destinations; journal manager CLOSE comment receipts on the correct PR/issue timeline; attempt terminal handoffs even when an earlier manager phase fails without hiding that error or allowing new scheduling. Refresh the architecture atlas with the handoff phase and current manager/dispatch citations.
• Preserve human takeover after a failed escalation comment: without the current escalation's recorded comment, the manager fails closed and leaves routing to the human handoff pass. Later handoff receipts cannot authorize recovery or hide human unassignments. (#56)
• Routed human handoffs (#56): once automatic recovery for an escalation is terminal (manager HUMAN, manager command failure as an explicit unable-to-diagnose, manager.rounds exhausted, no manager, or an escalation the loop can never act on again: unapplied decision, missing packet, unreceipted escalation comment), factory manage publishes one public-safe request per escalation generation with the question, bounded evidence links, proposed next step and the factory plan route owner/candidates and rationale; mentions only on the initial handoff and on an actual owner change, never assigning anyone. Intent and comment id are journaled (handoff, comment rows); a crash or failed gh between them is reconciled from the timeline before any retry, a failed lookup posts nothing, and --dry-run journals nothing. Human-takeover checks now trust only recorded comment ids (escalation, manager and handoff comments), never text prefixes; an edited recorded comment is human activity. Upgrade note: the first pass publishes a request for every open ready-for-human ticket whose escalation is already terminal, including pre-#56 escalations (their comments carry no receipt, so the manager no longer runs on them).
• Manager PR frontier (#15): factory manage observes every factory-owned agent/<n> PR through the schema-1 feedback producer (#79) and escalates once on red CI, undelivered current-head feedback, or manager.stale_days inactivity; late feedback is delivered at most once per (evidence_id, source_revision) and never from partial coverage or unverified ownership. New manager CLOSE decision closes the PR and proposes wontfix-proposal on a human issue (closes a factory-created child); wontfix-proposal is now a provisioned label. manager.review = "all" is implemented: the factory-approved label waits for a manager APPROVE bound to the exact head, never re-bound across a refresh. pr-opened journal rows carry the PR number; SPLIT records issue-created.
• The initiative label is now an unconditional execution guard, read fresh at each owning boundary: triage refuses before any model call, factory dispatch (including --ticket) refuses to claim, the manager refuses viability and escalation handling, and the merge stage refuses an agent/<n> PR whose ticket is an initiative. Refusals are logged, visible in --dry-run, and mutate nothing. Ordinary tickets are unchanged. (#55)
• Initiative issue template (.github/ISSUE_TEMPLATE/initiative.md, label initiative only) installed by factory init; the initiative label is provisioned with the others. Read-only factory plan list / factory plan inspect N emit schema 1 JSON: declared status/owner, sections, implementation links and linked-issue state, with partial/malformed results explicit. Reading an initiative grants no execution authority; #55 now enforces the initiative execution guard at every owning boundary. (#53)
• factory plan route N --reason <requirements|implementation|ci|unknown> [--path P]... --json and the optional [collaboration] section (fallback, reasons.<reason>, components."<exact path prefix>") resolve the human decision owner of a ticket read-only, with status (selected/candidates/unassigned/invalid), source revision and step-by-step provenance. A **Decision owner** ticket section is a human override; @org/team is rejected on user-owned repositories; no section means unchanged behaviour. (#54)
• Distinguish missing optional routing owners from invalid declarations, bound overrides to their own sections, retain rejected mixed-destination provenance without selecting a survivor, and reject dead reason mappings and duplicate normalized component prefixes. Preserve coverage notices on invalid invocations. (#54)
• Add factory chat: a supported, read-only Factory Manager console over the pinned upstream Pi runtime (console/app) and the schema-1 evidence interface. Seven bounded evidence tools, no shell/edit/write/dispatch, isolated console-owned settings and provider auth, and an explicit provider/model disclosure gate before any inference. Every start and --continue resume reobserves fresh evidence rather than trusting stale conversation. Node/Pi stays an optional console dependency (npm ci --ignore-scripts --prefix console/app); ordinary Factory execution never needs it. (#86)
• Add dashboard /chat: the selected three-column Factory Manager layout with live repository-wide/case/run evidence, citation inspection, truthful request activity, Motion-powered reflowing context panels, and browser-local conversation history. The page remains read-only and uses the existing bounded /api/ask transport.
• Add schema-1 source-versioned PR feedback at full dashboard tickets[].pr.feedback, shared by Review, Inbox and briefing. Retain simultaneous native review/thread/check evidence and provenance-backed Factory reviews with deterministic identities, explicit unknown/partial coverage, fixed 100-item/two-page/20 KB-body/32-error/30-second bounds, and head-race handling. Reviews carry the provider's own updatedAt, so an edited review revises its source revision. Detail reads run only for open PRs; closed and merged PRs keep the schema-1 envelope with not_collected sources, which is unknown rather than empty or unsupported. Read-only: no feedback delivery, readiness or merge authority; runtime JSON is unchanged. (#79)
• Add optional Codebase history: stable commit-timeline maps, baseline comparisons, confidence-aware relationships, pinned source citations, and bounded background refresh through factory[atlas]. (#67)
• Add opt-in direction viability to factory manage: needs-review PRs before needs-viability issues, evidence-cited BUILD/DONT_BUILD/DEFER comments, and label-event replay protection. Only issue BUILD enters needs-triage; PRs remain recommendation-only, with no review or handoff mechanics.
• Fix manager prompt transport to use files, including factory learn; validate manage...

factory 0.3.0

Changelog

0.3.0 — 2026-09-08

Compared against 0.2.0 (f9122cd). Installed fleet-wide through District at 9478e9d.

Manager stage (inert until [manager].command is configured)

• factory manage: resolves untouched ready-for-human escalation packets with a closed, code-applied decision menu — RETRY, REWRITE, SPLIT, ROUTE, HUMAN. Malformed output is HUMAN. A manage event is recorded before any GitHub mutation; a failed mutation leaves the ticket with the human, records a lifecycle mechanism_failure, and is never replayed. (#13)
• [manager] config table: command, rounds, review = "escalated" | "all"; legacy string commands are parsed with shlex. factory doctor reports the manager only when configured. (#12)
• Manager notes: .factory/manager/notes.md read into every manager prompt, with write-back and consolidation. (#14)
• factory learn runs through the manager when configured and opens a chore PR carrying only .factory-lessons.md; unchanged without [manager]. (#16)
• CURATE decision, accepted only from factory learn: proposes AGENTS.md / .omp/skills/** / CONTRIBUTING.md changes as a chore PR; verification paths are never touched. (#18)
• [defaults.manager] caps (max_active_cap, budget_min_cap) for District's fleet-level manager pass. (#21)
• Structured escalation packets written by escalate() to .factory/escalations/<n>.md: reason, attempt table, gate/review evidence, kept worktree. (#11)

Review stage

• Reviewer contract grounds every blocking finding in an acceptance criterion, a documented rule with its source, or a concrete correctness/security defect with trigger and impact. Required fixes are separated from optional suggestions; REVISE only while required fixes remain. Net-new abstractions beyond the brief need justification; missing justification alone does not block. A passing gate does not excuse a defect it did not detect. (#36)
• Reviewer defaults to omp with anthropic/claude-fable-5-1.

Merge stage

• Fix: refresh_pr_branch erased a PR whose agent/<n> branch had no local copy — it started the branch from main, gated an empty tree, and force-pushed it; GitHub then auto-closed the PR. Refresh now starts from origin/agent/<n>, refuses to push a head with nothing ahead of main, and pulls such a PR from merge candidacy so it escalates once rather than starving the queue. (#49)
• Fix: upstream-sync PRs were squashed once upstream moved past the PR's tip, dropping the ancestry the sync exists to preserve. Merge method is now judged by the PR's merge-base with upstream, not upstream's current tip; sync PRs merge main in rather than rebase. (#41)

Workers

• [workers.<name>].when rules route tickets by label/state; ci-fix and conflict default profiles. (#20)
• Per-ticket brief at claim time: .factory/brief-<n>.md from git log -S/grep of ticket nouns, recent PRs touching those files, the triage brief, and matching lessons — deterministic, token-capped, appended to the worker prompt. (#17)

Runtime evidence and observability

• F01: authoritative execution lifecycle journal — enter/exit, child processes, handoffs, outcomes and reasons — per stage, in .factory/events.jsonl. (#26)
• F02: known waits and evidenced lock ownership; ticket_lock_contended, merge_lock_contended, ci_pending and similar are recorded, not inferred. (#27)
• F03: factory dashboard --runtime-json — bounded schema 1 runtime projection from local read-only evidence only; partial source failures stay structured. (#28)
• Bounded read-only FM evidence interface with a Pi consumer; grounded full decision briefings and contextual FM questions. (#25, #38)
• Runtime quality graded independently of history-window completeness; journal read once per full snapshot. (#43)

Stats and dashboard

• Human-touch metrics: escalation count, resolver attribution (human/factory/unknown), minutes in ready-for-human, re-queues; escalations_per_week and human_resolved_pct in --json and the dashboard KPI row. (#10)
• Per-label pass rates in factory stats. (#19)
• Dashboard and site adopt the mikeroySoft design system; sage product theme.

Install and onboarding

• factory install runs triage before dispatch in the unit, jitters timers, and serialises triage on the host lock; restarts only units that existed and changed.
• factory init writes .github/workflows/ci.yml when the repo has no workflow; doctor warns on a missing or placeholder workflow.
• Repository adopted under District; agent-factory renamed to factory.

Known limitations

• The manager stage runs an arbitrary configured executable; configure the agent CLI in read-only/no-tools mode. Factory instructs it not to edit files but does not sandbox it.
• A stale local agent/<n> (behind the remote after another host advanced it) can still win in ensure_worktree; --force-with-lease does not protect because the refresh fetches first. Only relevant with a second dispatcher host. Tracked on #49.
• cargo has no minimum-release-age control; crates.io installs are unguarded by District's package-age policy.