v0.33.0

The performance & autonomy overhaul — the tool stops feeling slow and stalled. A cold
engineering review (three parallel deep-dives over the work loop, the plan→expert flow and the
architecture) found the stalls were structural: an unbounded CI wait, a 6-minute reviewer poll
paid on every PR, a dead-code run budget, no automatic epic progression, zero visible signal
when a run stopped, ~50–140 uncached GitHub calls per issue, a ~1.3 s interpreter spawn on every
tool call, ~72 KB of always-loaded instructions, and a telemetry subsystem blind to time. Epic
#561 fixed all fifteen findings: #562–#576, PRs #577–#591, every behavioral change with tests
verified by breaking the code, and every PR through the review gate (Codex external review in
rounds while quota lasted — 40+ findings fixed across the batch — then the sanctioned
self-review fallback, stated on each record).

Added

• The product agenda policy is written down (#576, closing epic #561's list). CONTRIBUTING.md
  now states the rebalance rule the architecture review argued for: releases should carry
  user-facing value (the last measured week ran 8:1 self-directed); before building a new
  self-control mechanism, ask what measurement would make it unnecessary; and field evidence
  (/board telemetry, now with durations) outranks introspection when picking the next piece
  of work. A policy, not a gate — deliberately unenforced.

Changed

• Board-Work.ps1 sheds its inline job scheduler and process killer (#575, part of epic
  #561; first slice of the monolith split). The 3,176-line dispatcher — 18.7% of all script
  code — contained a machine-capacity governor and a process-tree supervisor a board tool has
  no business holding inline. Both clusters moved verbatim into part files
  (BoardWork.Capacity.ps1: capacity snapshot, dispatch plan, fleet slot pacing, dispatch
  governor; BoardWork.Processes.ps1: pid→parent map, descendant walk, guarded tree kill,
  orphan detection and reaping), dot-sourced before the test guard so every caller and test
  sees the exact same surface. Behavior-preserving by construction and by evidence: 371 tests
  across the Board-Work, Board-Doctor and lint suites stay green with zero changes to them.
  Further slices (session registry, worktree launch) can now follow the same pattern.

Added

• .agentic-board/ finally has a garbage collector (#574, part of epic #561). The state dir
  had accumulated 48 files across the project's entire life — briefings and launch scripts for
  issues #123→#532, signal markers, compaction snapshots — because nothing ever cleaned it.
  Clear-AbiosState.ps1 reaps only regenerable per-run debris (briefings, launch scripts,
  expert briefs, per-issue logs, signal markers, snapshots) under three pinned rules: a live
  session protects its files at any age, files younger than -MaxAgeDays (14) stay, and durable
  records (sessions history, run-ledger marker, contract, denial log, fleet files) plus anything
  unrecognized are never touched — unknown is somebody's state, not deletable. Plan-only by
  default, -Force executes, every reap entry carries its reason. This is the GC half of the
  "one run, one record" direction; the record half was seeded by #568's sessions-history.jsonl
  archive (started/ended/outcome). Verified by 9 Pester tests.

Changed

• The always-loaded instruction payload shrank ~60% (#573, part of epic #561). Every
  /board invocation loaded ~72 KB of instructions (≈18k tokens, ~9% of a 200k window) before
  reading a single line of the actual issue — commands/board.md described all 22 verbs in full
  and projects-admin/SKILL.md duplicated the three biggest. The ten heavyweight verbs (work,
  plan, fill, field, changelog, handoff, doctor, cerrar-ciclo, telemetry, triage) now live in
  per-verb reference files loaded on demand when that verb runs; the command keeps the menu,
  one-line routing and the safety rules, and the skill keeps identity/anchoring/conventions and
  the routing table. board.md: 34.7→9.7 KB; SKILL.md: 31.6→12.7 KB. No content was deleted —
  every recipe moved verbatim into its reference. Contract intact: command-surface suite green
  and README regions unchanged (docs-freshness gate). This weight was causing the very
  auto-compactions the run-ledger machinery exists to survive — paying the disease and the cure.
• The brake hook stops taxing every tool call in every session (#572, part of epic #561).
  The PreToolUse hook fires on every Bash/Edit/Write in every session and paid a full pwsh
  interpreter spawn (measured ~1.3 s on the reference machine) just to discover that no
  brake marker exists — the case for every session except a launched autonomous run. At 100
  tool calls per session that was ~2 minutes of pure interpreter startup, paid by every user in
  every repo, and the tool had no way to see it (#568's blindness, costing #572's tax).
  Brake-PreCheck.cmd now answers the common case in ~0.25 s (5× cheaper): no marker in
  the working directory or three ancestors → exit 0 without ever starting pwsh. Fail direction
  unchanged: an armed worktree always reaches the real PowerShell guard with its fail-closed
  behavior intact, and stdin passes through untouched. Verified by 4 new tests, including an
  end-to-end deny through the shim over real stdin and a no-pwsh timing bound.
• GitHub reads gain a cache, and the fail-closed wrapper gains a lint that stops the
  bleeding (#571, part of epic #561). Every gh invocation is a fresh process + TLS round trip
  and nothing was cached — a clean work cycle made ~50–65 calls, many re-reading facts that had
  not changed. Invoke-GhCached is a small cross-process file cache (in-memory caches die
  with each script's pwsh process) for read-only queries with an explicit TTL; failures are
  never cached, -Force busts, TTL 0 always fetches. The board's Status-schema read — the
  exact anti-pattern quoted in the wrapper's own header, still live — now goes through it
  (5-min TTL), and four more of the worst raw sites migrated to Invoke-Gh (issue blockers,
  claims, linked PRs, commit search: a 401 no longer reads as "no blockers/no claim/no PRs").
  A new lint test freezes the per-file raw-gh baseline (59 sites): counts may only go down,
  a new raw call names its file and fails, and stale allowances are themselves flagged.
  Verified by 7 new Pester tests (cache semantics + lint).
• Evidence is written once, linked twice (#570, part of epic #561). The same evidence block
  used to be copied to three destinations (PR body, issue comment, evidence/<issue>.md) —
  three copies of one content, drifting independently, and the "INCOMPLETE → record → re-run"
  loop mostly existed to keep them in sync. Now evidence/<issue>.md is the single source of
  truth and the PR body and issue comment carry a link stub
  (Format-EvidenceLinkStub: marker + pass/fail summary + link to the file). The completion
  check tightened accordingly: a surface satisfies its requirement with the marker plus
  substance (the file reference or a full block — pre-#570 runs stay valid); a bare marker is
  a stamp anyone can leave, and now reads as missing. Verified by 7 new Pester tests.

Added

• The definition of done scales to the diff (#569, part of epic #561). The contract's six
  gates all defaulted to true and the brief listed them flat, so a 10-line docs fix owed the
  same verify pass as a model migration. Get-ApplicableDodGates derives the owed gates from
  the changed paths — bpa/tmdlBreaking only when a semantic-model file changed,
  build/lint/tests only when something executable changed, ci always — and the
  work-class CLI prints which gates the current diff owes and which it does not. Fail direction:
  an unreadable diff owes every enabled gate, the derivation can only disable gates the
  contract enabled (never re-enable one it turned off), and an unrecognized custom gate is
  always owed. The autonomous brief now points the run at this instead of the flat list.
  Verified by 6 new Pester tests.
• The tool can finally measure its own time (#568, part of epic #561). The telemetry
  subsystem measured everything except time: the transcript parser discarded the timestamp
  every line carries, the correlation window was "6 events" (4 seconds or 40 minutes,
  unknowable), the ledger's columns were events and bytes, started had minute granularity, and
  completed session rows were deleted — every run's wall-clock cost vanished at the moment it
  finished. Now: events keep their ISO timestamp; episodes carry ts + durationMs (unknown
  stays null, never a fabricated zero); per-session records and the field ledger gain
  durationMin; started is stamped to the second (the supervisor reads both formats); and a
  completed session row is archived to sessions-history.jsonl with ended and its outcome
  (pr-merged / cleaned / close-loop) before removal. Verified by 8 new Pester tests.
• The visual work-class gate scopes to what the owner actually judges by looking (#567, part
  of epic #561). In a web app (a FabricApp) every change touches html/css/assets, so the
  classifier marked the whole project "visual" and the classification stopped carrying
  information — when the end-to-end gate reopens (#541), that would have routed 100% of web work
  to the human. The policy gains workClass.codeExceptions — globs subtracted from
  visualPatterns, declared per project in the contract (e.g. src/components/**/*.css for
  utility styles nobody screenshots), never guessed: the default list is empty and the
  fail-visual direction is unchanged out of the box. And the verdict now carries
  visualGroups — visual paths grouped by top-level section — so the owner approves "the
  pages section" once instead of 47 file rows. Verified by 6 new Pester tests.
• The epic walker: /board expert auto -Epic <n> dispatches sub-issues wave by wave (#566,
  part of epic #561). Nothing advanced an epic before — Expert-Auto took one -Issue, so a plan
  with N sub-issues cost N human launches (Board-Plan even fetched the sub-issue list and threw
  it away). -Epic reads the epic's native sub-issues, classifies them (done / in-flight /
  blocked / ready) and launches one autonomous session per ready sub-issue with the
  contract's brake and budget, each briefed with the epic's enriched plan plus its own issue
  text. Idempotent by design: merge a wave's PRs and re-run the same command for the next
  wave; done and in-flight sub-issues are never re-dispatched, and a sub-issue whose PR state
  could not be read counts as in-flight — dispatching a possible duplicate session is the worse
  error, so the unknown fails closed. Verified by 7 new Pester tests on the wave classifier.
• A stopped run now signals instead of sitting silent (#565, part of epic #561). A braked,
  out-of-budget, or stalled run was invisible from outside — the deny payload reached only the
  model, the supervisor's verdict existed only on a terminal nobody was required to watch, and
  the human learned about a stall by noticing the PR never appeared. Three surfaces fix that:
  every hook denial appends to a local denials.jsonl and posts one [abios-signal] issue
  comment per (kind, issue) — deduped by a marker file so a retrying agent cannot flood the
  issue (the brake marker now records the repo to make posting possible); the fleet supervisor
  gained -Post, publishing an [abios-stall] comment for each session past the threshold
  with no PR; and the session watch runs the supervisor automatically every 10 poll cycles,
  so stalls surface without a separate human command. All signaling is best-effort and can never
  change a verdict. Verified by 14 new Pester tests across the guard, supervisor and watch.
• The expert run's time budget is now enforced, not advisory (#564, part of epic #561).
  Get-BudgetVerdict computed a handoff verdict that nothing ever called — the contract's
  120-minute budget existed only as a sentence in the brief, so a runaway run had no wall-clock
  limit at all. The budget now travels in the brake marker (budgetMinutes + armedAt) and the
  PreToolUse hook enforces it: past the budget, work commands are refused with an instructive
  message while the wrap-up stays open — /board handoff -Save, committing and pushing WIP,
  closing the run ledger, and leaving a PR/issue comment. The brake always wins over the
  exemption (a push to main over budget is still a refused merge), and the budget fails
  open on a corrupt timestamp — it is a liveness limit, not a safety control, and must never
  brick ordinary work. Iteration counts stay advisory: a hook that sees single tool calls cannot
  count verify-loop iterations honestly. Verified by 24 new Pester tests (pure + hook end-to-end
  over real stdin); mutation-checked (enforcement disconnected → 3 tests red → restored).

Fixed

• The review wait learns from silence, and any reviewer's answer counts (#563, part of epic
  #561). Two defects made every PR pay the full review timeout forever: the wait loop broke only
  on a review whose author was Copilot — a human or external review landing mid-poll kept it
  spinning — and the per-account cooldown (#367) armed only on an explicit "cannot review"
  answer, so a Copilot that was simply silent taught the gate nothing. Now the wait's arrival
  test is the same evidence rule as the verdict — any GitHub review or recorded external
  review ([abios-review] comment) bound to the current head commit ends it (stale evidence of
  earlier commits keeps waiting), the PR comments arrive in the same authoritative GraphQL read
  as the reviews (one call fewer per poll), and silence past the timeout arms a
  1-day cooldown (weaker evidence than an explicit refusal, which keeps its
  -CopilotCooldownDays default of 7). Verified by 10 new Pester tests; mutation-checked
  (author-only arrival and inverted silence check reintroduced → 5 tests red → restored).
• The review gate's CI wait now has a ceiling, and the CI and review waits run concurrently
  (#562, part of epic #561). gh pr checks --watch was the only unbounded wait in the codebase:
  a queued or never-scheduled workflow hung the entire session indefinitely, with no signal. The
  gate now polls a structured checks snapshot in the same loop as the review wait — worst case is
  max(CI, review) instead of their sum — and the CI side expires at -CiTimeoutMinutes
  (default 25) into an explicit "checks still pending" BLOCK, never a silent hang and never
  a pass. An unreadable checks snapshot also blocks (fail closed) instead of being trusted from
  display text. Verified by 12 new Pester tests; mutation-checked by reintroducing both defects
  (cancel-counts-as-pass, sequential-OR exit) and watching 4 tests go red.