v0.8.0 - Agent-log scorecard, write-path guards, security fixes

Added

Agent tooling

• restore_file gains an optional from_base mode that restores a file
  from origin's default branch, giving an agent a first-class way back when
  its own commit captured a bad change (the destructive-git block refuses a
  base-scoped checkout). The default path is unchanged, and the per-path cap
  applies to both modes.
• scripts/agent_log_scorecard.py <worktree>/agent.log [--json] reports a
  per-attempt scorecard from an agent log: step totals, per-tool counts, the
  first mutating step, the productive-step ratio, guard trips by normalized
  label, and the outcome with its done summary or parked reason. It slices
  at the final [boot] line so a resumed run is scored on its last attempt,
  and reads the same grammar the log writers emit.

PR review

• A warn-only pr_title_mismatch notification fires after a PR opens when
  its title shares no 4+ letter word with the story summary or the changed
  files. It fails open and never alters the PR, verdict, or merge path.

Scheduler

• The scheduler notifies when the running checkout trails upstream by
  commits that touch pipeline/ or app/, since merged orchestration
  changes are otherwise invisible until the daemon restarts. The notice
  dedups on the upstream SHA and points at scripts/reload_pipeline_daemon.sh.

Changed

• Python 3.11 is now the floor everywhere — install check, ruff
target-version, and a new 3.11 CI leg (3.11-3.14) — with a consistency
  test so the three cannot drift. Python 3.10 is no longer supported.
• Dependencies: fastapi>=0.143.0, ruff==0.17.0.
• The on-device dispatch tier is documented as not cleared for story
  dispatch, including its measured first-pass-clean rates and the preflight
  an operator must run to route it back.

Fixed

Security

• Acceptance fixture paths that are absolute or contain .. are rejected at
  ingest and skipped with an operator notice at dispatch, so plan-authored
  data can no longer write outside the worktree (security audit 2026-10-19).
• Pipeline secrets (PIPELINE_NOTIFY_EMAIL_PASSWORD, PLANE_API_KEY,
  GH_TOKEN, GITHUB_TOKEN) are stripped from every agent subprocess
  environment, including locally spawned agents that inherit the parent
  environment.

Dispatch and merge

• A story re-dispatched while still todo no longer preserves a stale or
  half-written acceptance fixture from a dead earlier attempt: the skip is
  gated on the story's status, so a fixture-only oracle cannot go green with
  no implementation (Mode 45).
• A merge-gate CI-fail rework keeps any earlier review feedback instead of
  replacing it with only the CI error, so the implementer cannot silently
  re-break what the reviewer asked for.
• An escalation whose resolved target equals the story's current
  backend/model is skipped rather than resetting the rework budget and
  inflating the first-pass-clean metric.

Local agent

• Every write path rejects a file whose last non-blank line is a trailing
  omission marker (e.g. # ... (rest of file remains same)), so a
  context-exhausted model cannot silently land a truncated file.
• replace_lines computes dropped top-level symbols unconditionally, so a
  renamed definition can no longer slip past the symbol-loss check.
• create_file overwrites of non-.py files reject content that drops more
  than ~10% (floor of two) of the original lines unless confirm_removals
  is set, covering markdown and other text files the AST guard cannot see.

Review, chat and classification

• The review loop makes one final no-tools call for a verdict when it
  exhausts without one, failing closed to the existing no-verdict path.
• At the turn cap, chat strips [TOOL_CALL] markup and answers in prose; a
  near-cap nudge asks the model to do so before the budget runs out.
• Legacy event-less sidecar notices map to the same reason names structured
  events use, so escalations are no longer double-counted and benign topic
  words no longer mark a clean story non-clean.

Scheduler, usage and notifications

• The staleness check receives the config fingerprint it compares, so the
  trailing-checkout advisory actually fires in production.
• Subscription usage percentages parse when the resets ... suffix is
  absent (e.g. Current session: 0% used), instead of falling back to the
  request-count approximation and potentially pausing an available
  subscription.
• The review-gate deferral notice fires once per gated episode instead of
  once per scheduler tick (one plan accumulated 159 identical records in
  3.2 hours).

v0.7.0 - Detached test-author step, scheduler and review-gate fixes, local-agent guards, bench tooling

Added

Dispatch

• The test-author step now runs detached from dispatch_story through
  non-blocking start/collect primitives, with a decision step for when to
  use it; the planner block moved into dispatch_phases (no behavior
  change for that extraction).
• A PR title is derived from the branch's Conventional Commits subject
  .

Local-agent harness

• view_file returns a repeat advisory instead of re-emitting identical
  bytes, and the one-sided-read latch guard keys reads by their own bound
  so orientation in a large file is not a false-positive park.
• FULL_SUITE_DONE_BAR is armed on the oracle harness, the oracle gets
  the story summary in its env, and its auto-commit has a story-derived
  subject.

Scheduler

• Health is written at startup before the first reconcile and refreshed
  during long watchdog joins; SchedulerDaemon.start()'s reconcile is
  bounded by the watchdog.

Benchmark tooling

• Every role is pinned per arm through role_config, the grid refuses to
  run when its pins cannot resolve, and a one-command published-matrix
  driver records pinned provenance.
• A Tier 3 inventory_pagination cross-module task, Claude dollars per
  success in the scorecard, preserved dispatch cost logs, and Claude
  cells skipped when credits or rate limits run out.

Changed

• The maintainer's home path is scrubbed from archived plans, specs and
  retros, and the committed launchd files use a placeholder home
  directory.
• The exported overlord policy spec matches the live parked-story policy.
• A bare pytest run is clean on a fresh clone.

Fixed

Scheduler and merge

• A dispatch lease whose owner pid is dead is treated as not live.
• CI status queries run in the scoped plan repo, and an unreadable CI
  status parks the story instead of being treated as no CI.
• A trailing DONE line in agent.log counts as agent_done in the wedge
  scan.

Review gate

• ClaudeCliDriver.resource_status accepts model_tag, so the review
  gate no longer crashes the scheduler.
• Review is gated on the model it will actually run, the Claude usage
  pause is re-derived from cached percentages, and consecutive
  rate-limited review deferrals are bounded and then park.

Local agent

• A successful create_file resets its own repetition count.

v0.6.0 - Language-agnostic gates, global rules installer, security-audit tracking, file splits

Added

Language-agnostic gates (LAG)

• A repo can declare its test, lint and build commands, and its
  test_globs, in a root .fagan.json; the scope gate treats declared
  test globs as test files (#1037, #1039).
• Build, test and lint command detection moved out of build_detect.py
  into ecosystem_detect.py, which now also detects Go and .NET and fails
  visibly on any other unsupported ecosystem (#1036, #1038).
• Maven and Gradle projects run verify / check through the project
  wrapper when one is present, and their test failures are parsed so
  acceptance runs scope to the failing JVM test classes (#1041, #1044).
• One test-path classifier is shared by the scope gate, ingest and
  plan_conflict_ruling, and the ingest sizing check recognises any
  backticked relative source path (#1035, #1040).

Global rules bundle

• An opt-in installer (scripts/install_global_rules.py, hinted by
  install.sh) writes the engineering standards and the pipeline workflow
  rules into each supported agent CLI's global instructions file inside a
  managed block, honouring XDG_CONFIG_HOME; documented in the README
  Quickstart and REFERENCE (#992-#1000, #1019, #1021).

Security audit tracking

• The record_security_audit MCP tool and a per-repo audit record, with
  due thresholds and a scheduler notice when a repo is due (#1009-#1012).
• The security-engineer verdict is posted on the PR, and ingest warns when
  a brief needs a security review the pipeline will not run (#1007, #1008).

Scheduler and harness diagnostics

• The scheduler warns at startup when the launchd plist and the MCP
  server config disagree, by diffing the two environments rather than a
  hand-maintained catalog; secret-like and non-PIPELINE_ values are
  masked (#1027, #1032, #1059).
• Each dispatch attempt's log starts with a [boot] line, and a watchdog
  kill snapshots the end of agent.log into the plan dir (#1023, #1024).
• A rejected story is re-reviewed when the reviewing logic has changed,
  and a skipped re-review is notified once per SHA (#1049, #1052).
• A two-tier CI gate enforces the 1000-line rule: blocking on production
  Python, warn-only elsewhere (#1048).
• The one-line installer finishes Claude Code setup, and the README leads
  with a demo and the one-line install (#982, #987).

Changed

• A triage mark_done ruling requires a merged PR before it marks a story
  done, and a PR merged outside the pipeline is recorded as story_merged
  when mark_story_done closes it (#1033, #984).
• Each automated rebrief cause has its own event name, and the rebrief
  header detectors match an occurrence rather than a mention (#1050,
  #1031).
• effective_env_config's effective field is labelled as this-process
  rather than resolved against code defaults (#1058).
• scripts/local_agent.py, scripts/local_agent_oracle.py,
  app/dashboard.py, pipeline/dispatch.py, pipeline/story_status.py
  and pipeline/worktree_patch.py were each brought under 1000 lines.
  No behavior change (#1051, #1053-#1057).
• The retro backlog is capped so the pending-marker list stays drainable
  (#1047).

Fixed

Dispatch & merge

• Dispatch is refused when a plan's repo_root no longer exists, and a
  story that is already done or has an open PR is not re-dispatched
  (#1016).

• A story's registry model-name pin resolves to its tag, and the scope
  gate treats nested and Rust test paths as test files (#1004, #1003).

• Every pipeline runtime marker is self-excluded in worktrees, and the
  acceptance oracle is kept out of story commits (#1014, #1015).

• A dispatch attempt clears the previous attempt's done markers, and the
  watchdog skips its SIGTERM when the stored pid belongs to a process
  started after dispatch (#1025, #1026).

• Triage defers a transient overlord failure instead of parking the story,
  and reports git evidence against the worktree-resolved base branch
  (#981, #1042, #1043).

• Triage's sweep no longer writes back a stale whole-manifest snapshot: it
  re-reads the manifest and applies only the sweep's own changes, so
  writes made while a ruling was in flight survive (#1061).

Tests

• The smoke precondition test no longer depends on the operator's local
  model registry (#1062).

Reporting accuracy

• The local-success report is filtered to a single repo and names its
  cohort; the plan summary flags a reviewer-bounced story beside its
  baseline verdict; ruling and sizing notices carry their story key
  (#1030, #989, #983, #979).

Dashboard

• The toast message gets its own full-width line (#990).

v0.5.0 - Local-dispatch convergence, module splits, cold imports, reporting accuracy

Added

Local-dispatch convergence (LDC)

• A local agent that reaches its step cap now finishes as done when its
  branch carries production changes against the default branch and the
  full suite and lint pass, instead of burning a rework cycle on work that
  already landed. A rework redispatch that produced no new commit starts
  the next attempt from a fresh context seeded with a root-cause
  diagnosis, and a third restore_file on the same path in one run is
  refused (#932, #933, #937, #938).
• The test-author phase is skipped for stories whose declared files are
  all docs or config, and its prompt now carries a size budget and asks
  for structural assertions rather than prose wording (#928, #934).
• Each completed plan's summary, including a per-story first-pass-clean
  verdict, is written to <plan>.report.md in PLAN_DIR (#931, #936).

Changed

• Re-ingesting an already-ingested plan is rejected when any story lacks
  an explicit key. Such a re-ingest used to mint duplicate stories
  that dispatched concurrently; give every story a key to re-ingest
  (#929).
• triage.py, service.py and server.py were each brought under 1000
  lines, extracting triage_patch_acceptance, triage_story_actions,
  service_decisions, service_role_config and
  server_tools_lifecycle as their own modules. No behavior change
  (#941-#944, #946).
• Every pipeline module now imports cold, in any order - the remaining
  import-order dependencies on pipeline.server were removed and the
  stale comments that described them corrected (#966-#969, #971-#974).
• Improved the descriptions of the lowest-scoring MCP tools (#927).

Fixed

Dispatch & merge

• A story's bare registry model name now resolves to its tag instead of
  silently falling back to the default local model (#930).
• A worktree whose HEAD is detached is re-attached to the story branch
  before the rebase, or the story is parked when re-attaching would
  orphan commits (#965).
• A rescoped story's PR title is refreshed when the PR already exists, so
  the squash commit subject describes the final work (#945).

Local harness

• A capped view_file result is cut on a line boundary and says the file
  on disk is intact, so a weak executor no longer "repairs" a file that
  only looked truncated; the scope gate's revert advice is now runnable on
  the local harness (#958, #961).

Reporting accuracy

• The plan rollup no longer counts the synthetic uncorrelated bucket as a
  story, collapses metric groups that share a story key, and counts
  watchdog kills and reviewer send-backs as rework cycles (and a reviewer
  send-back as a dirty first pass). Per-story metric, sizing-auto-route
  and preflight-override notices carry their story key; step-cap and
  give-up rebriefs emit brief_patched; retry-cap notices render the cap
  instead of an object address (#948-#951, #954-#957, #959, #960).

Dashboard

• A long multi-line notification toast now renders cleanly (#953).

Test determinism

• Two tests that raced git's background auto-maintenance or the shared
  advisory fetch lock across xdist workers are now deterministic (#935,
  #975).

v0.4.0 - MCP registry listings, systemd deployment, overlord autonomy, MCP tool surface hygiene

What's new in 0.4.0

Added

MCP registry listings

• Added server.json for the official MCP registry
  (io.github.motock/fagan) and a Dockerfile for Glama's build/
  introspection check, plus a Glama badge and dashboard screenshots in the
  README (#811, #812).

Linux systemd deployment

• Added systemd user-unit templates for the advance-scheduler daemon and
  the usage-poller, a service+timer pair, a logrotate template, a
  scripts/generate_systemd_units.sh generator, and README documentation
  for the Linux setup path (#822, #824, #825, #831, #838).

Dashboard plan-ingest picker & Comms polish

• A new backend endpoint lists plans available to ingest, backing a
  dynamic select in the ingest picker in place of free-text plan-name
  entry; the Comms panel shows a typing indicator while a reply is
  pending (#820, #821, #823, #828, #830).

Local-dispatch observability (LD90-W0)

• Every escalation, fallback, park and rework notice across
  story_status, triage, escalation, review_orchestrator,
  dispatch and advance now carries event, story_key and
  correlation_id. A first_pass_clean/disqualifying_events rollup and
  a new pure local_success classifier feed a local_success_report CLI:
  rolling first-pass-clean rate by dispatch tier, with a reason breakdown,
  read from the plan notification sidecars (#850-#858, #865).

Reviewer & scheduler hardening (LD90-W5, MFR, SWH)

• The reviewer now recognizes and classifies transport failures (502/503/
  504, timeouts) and defers the review like a rate limit instead of
  charging them against review_inconclusive_count; a diff file-inventory
  criterion was added to the review checklist. The reviewer's bash tool
  and the notification-outbox drain phase are now bounded by a hard
  timeout so a runaway command or a wedged drain can no longer hold a
  scheduler tick open. The scheduler now publishes the checkout revision
  it's actually running, and preflight reports it too (#861, #863,
  #866-#873, #906, #907, #909).

Rework & done-bar correctness (FSU, TGE, SST, SU, HRG, ASB, RFW)

• A red full suite gets one retry before the done-gate rejects it.
  Pre-existing baseline failures are now exempted from the tick-side
  detached grade and both agent-side full-suite gates, with the baseline's
  failing node ids recorded in the worktree for the agent to read. A
  rework round that carries an acceptance block is graded on the same
  done-bar the agent was originally held to; a rework transcript that
  already claims done is never resumed; a reviewer-feedback rework round
  can no longer be marked done on oracle-green alone. The agent scratchpad
  is untracked before grading, and the planner/dispatch prompts now state
  it must never be committed (#874-#885).

Overlord autonomy round 2 (OA2)

• Triage now recovers born-broken-oracle stories (blocked_oracle),
  resets its own attempt counter after a successful recovery, and skips
  stories with an unresolved infra-transport failure instead of ruling on
  them. Watchdog kills converge via a streak counter so repeated hangs
  escalate. request_decision fails open - parking the story for a human
  • when the overlord backend errors. Agent give_up now routes through
    the existing rebrief diagnosis path instead of notify-only. Merge-park
    re-adjudication was widened to the risk-threshold reasons the gate
    actually emits, and PIPELINE_AUTO_TRIAGE now ships enabled in the
    shipped launchd template (#834, #886-#893).

Resume-rebase remote sync (SRB)

• A resumed agent's story branch is force-pushed after the resume rebase
  so the resumed agent's own push is a fast-forward, and the seed prompt
  is now forbidden from pulling, merging, or rebasing the branch from
  origin itself (#894, #895).

Post-LD90 priorities: file-size discipline & plan-conflict handling
(PLD90)

• dispatch.py, story_status.py and advance.py were each split to
  bring their line counts under 1000, extracting dispatch_routing,
  dispatch_attempt, dispatch_baseline, dispatch_worktree,
  detached_grade and advance_merge as their own modules. A new
  plan-conflict classifier and overlord-ruling module distinguish a red
  suite caused entirely by pre-existing, branch-untouched tests from a
  real regression, routing the former to an overlord ruling instead of
  burning a rework cycle. A scope gate sends a branch with out-of-scope
  production changes back before review, naming the exact offending
  paths. ingest_plan gained an optional files field to scope and size
  stories, with an ingest-time warning when a local/cloud-open-source
  story's file scope exceeds the documented dispatch-sizing caps. The
  reviewer prompt now receives the story's brief, sibling status and
  Deferred lines (#833, #841, #896-#920).

MCP tool surface hygiene (MCPHYG)

• Prompted by an external tool-quality review: check_story_status is now
  actually registered as an MCP tool - it was fully implemented and
  documented but never decorated, making it unreachable via MCP.
  mark_story_in_progress and mark_story_done now serialize their
  manifest writes under the same plan lock every other write path uses,
  closing a race with the scheduler's tick. The duplicated
  _VALID_STORY_STATUSES allowlist (each copy listing "done" twice) is
  now a single source of truth. A new patch_plan tool edits plan-level
  fields (currently role_config) under the same locked, allowlisted
  pattern as patch_story. checkpoint is renamed to checkpoint_story
  for naming consistency with the rest of the tool surface, with
  checkpoint kept as a working alias (#922-#926).

Changed

• get_role_config removed. It was a strict subset of
  get_effective_config (5 roles vs. 9, fewer fields, same underlying
  resolver) - use get_effective_config(plan_name=None) instead. This is
  a breaking change for any external caller of the MCP tool surface
  (#924).
• resource_status now uses a provider-aware default free-memory floor
  (2048MB for MLX, 512MB generic default for Ollama/LM Studio), and local
  model context sizing (num_ctx) is now driven by an empirically-tuned
  per-model table and sized per parallel slot, in place of a single static
  override (#818, #844-#846, #849).

Fixed

Merge & dispatch correctness

• Merge now pushes from the plan's own worktree instead of the pipeline's
  REPO_ROOT; dirty worktree state is committed before the merge-gate
  rebase instead of retrying the same unstaged-changes failure verbatim;
  the merge gate persists its own fresh test result onto the story instead
  of leaving last_test_check stale; the squash-merge commit gets an
  explicit, stable title instead of trusting GitHub's single-commit
  default, and no longer repeats the story key in the PR title; two
  dispatch/decision-log bugs found while live-diagnosing a parked plan
  were fixed (#819, #829, #835, #839, #840, #879).

Circular imports

• Broke a pipeline.ci <-> pipeline.build_detect <-> pipeline.server
  circular import that failed whenever pipeline.ci was imported first
  (#843).

Scratchpad tracking

• A checkpoint commit no longer tracks .agent_scratchpad.md, even if an
  earlier commit already had (#836).

Rebrief & triage prompts

• The rework diagnosis prompt now receives the story's original brief so
  it can't invent a fix that contradicts an already pre-authorized edit;
  triage skips stories with an unresolved infra-transport failure instead
  of misclassifying them, and keeps the real exception once the
  escalation ladder is exhausted (#834, #837).

[no-new-tests] anchoring

• The opt-out token is now anchored to the start of the brief, so a
  sentence elsewhere in the brief that merely mentions the token can no
  longer accidentally trigger it (#833).

Dashboard/UI polish

• The Comms typing indicator now anchors after the thread with tighter
  panel spacing, and the stale plan-board progress bar is hidden once the
  executor signals completion (#830, #905).

Docs accuracy

• Fixed a stale gpt-oss:20b 32768 num_ctx claim and documented the
  gpt-oss-20b-high override; documented the launchd install+reload
  requirement and added a reload helper so a plist edit actually reaches
  the running daemon; fixed the "Both routes" wording in the patch review
  & apply section (#827, #846, #847).

v0.3.0 - Chat-driven patch apply, overlord autonomy, registry-authoritative routing

What's new in 0.3.0

Added

Chat-driven worktree patch apply (WAP)

• Chat can now propose and apply a patch directly against a stuck story's
  worktree: origin-gated propose/apply routes distinguishing chat from UI
  callers, unified-diff parsing with size/path/whole-file-replacement limits,
  an in-memory patch-record store with a TTL and an HMAC token bound to the
  patch id and diff hash, a plan-locked atomic apply engine restricted to
  stuck stories, and every propose/apply journaled and notified with a
  correlation id (never diff content). A dashboard UI renders the stored
  diff and requests apply with a UI-origin header, and an end-to-end test
  suite covers the full propose/review/apply path (#750, #752, #753,
  #756-#761, #766, #768-#770, #772, #773).

Overlord parked-story autonomy

• The overlord can now rule on parked stories under a documented decision
  matrix and autonomy-mode ladder, informed by a live git-state probe and a
  fail-closed SPLIT payload parser. It executes its own rulings directly -
  split_story creates child stories in the manifest, mark_done corrects
  the story record from live evidence, patch_acceptance rewrites a
  diagnosed-broken fixture through validation - and adjudicates high-risk
  merges itself in full-autonomy mode, with every executed action recorded
  in the decisions log alongside its prior state (#736-#738, #740,
  #742-#745, #747).

Registry-authoritative provider/model routing

• An interactive picker (scripts/choose_providers.py) lets an operator
  choose each role's provider from the command line, and the
  getting-started smoke now announces and runs on the actually-resolved
  dispatch provider instead of refusing everything but claude (#795-#800).
• resolve_role is now the single source of truth for dispatch and
  escalation backend resolution: registry roles outrank
  PIPELINE_BACKEND_<ROLE>, which becomes an empty-state fallback rather
  than an override, preflight reports the registry-aware resolution, and
  the launchd templates stopped shipping routing env vars now that the
  registry is authoritative (#801-#807).

Dashboard & Comms chat

• The Comms chat panel renders markdown (bold, lists, code) while keeping
  the raw transcript export (#731, #734, #735), and dashboard plan lists
  can be filtered to the active repository with an all-repos toggle,
  backed by repo_root on plan summaries (#732, #733). A plan-ingest panel
  was added to Comms so a plan can be ingested from the dashboard UI
  without a chat-origin route, since ingest_plan is unreachable from
  chat by design (#782-#784, #787, #788).

One-line install & public-readiness

• scripts/remote-install.sh (curl -fsSL ... | bash) clones or updates a
  fixed local checkout and runs install.sh; README/CLAUDE.md gained the
  community-health files and fixes needed for public visibility; and the
  launchd plist/template files and generator were renamed from
  com.claude.pipeline.* to com.fagan.pipeline.* end to end, including
  runtime label references (#779-#781, #785, #786).

Fixed

Merge-park & CI-evidence correctness

• A story parked by the merge gate is no longer offered to triage as a
  candidate, an unreadable or in-flight CI status is never presented to the
  overlord as "no PR checks," a parked high-risk merge hold is
  re-adjudicated when its recorded evidence changes, CI evidence gathering
  now probes the story's real branch instead of a stale one, and high-risk
  merge adjudication records name the story instead of leaving it implicit
  (#751, #754, #755, #762-#765, #767, #771).

Scheduler reconcile resilience

• The scheduler process's per-call model budget is now clamped inside the
  reconcile join deadline, and the abandon-restart escape hatch is driven
  by whether the abandoned worker is actually still alive rather than a
  resettable streak counter (#739, #741, #746).

Chat streaming

• Fixed a duplicate reply bubble on the server's reply+result frame pair,
  stopped dropping the user's message and prior turns from the prompt
  across tool round-trips, and widened stall detection to catch
  attribute-style [TOOL_CALL name=...] openers (#774, #777, #778).

Getting-started smoke hardening

• The smoke now drives through the real ingest_plan path against a
  scratch repo with a real origin and a success bar scoped to
  tests_passed, rejects a malformed acceptance entry at ingest instead of
  crashing, resolves its announced dispatch backend through resolve_role,
  and its subprocess-spawning tests no longer leak the developer's real
  .pipeline.env into the child process (#791-#794, #808, #809).

Local dispatch reliability

• An in-flight local-backend story is no longer interrupted on local memory
  pressure (#790).

Test infrastructure

• Node frontend test suites now gate merges instead of being silently
  skipped, a workspace chat-body test was made behavioral instead of
  pinning source text, and a benchmark MockBackend regression was fixed
  (#775, #776, #789).

v0.2.0 - Streaming chat, per-backend usage reporting, release tooling

What's new in 0.2.0

Added

Streaming chat

• Streaming chat over SSE: POST /api/chat/stream backed by a stream_turn
  event generator, with the dashboard consuming the stream for incremental
  tool activity (#685, #686, #689, #695, #707).

Per-backend usage reporting

• Per-backend usage reporting: collect_backend_status surfaces
  resource_status per configured role, served by the dashboard usage
  endpoint and rendered as one row per backend in the usage banner
  (#688, #693, #694).

Plan-completion notifications

• Plan-completion notifications: plan completion is detected exactly once and
  emitted as a plan_completed notification, spooled to a per-plan outbox and
  delivered over SMTP with credential handling and redaction; the scheduler
  tick drains the outbox (#710, #713, #714, #715, #716, #717, #718).

Canonical agent.log grammar

• Canonical agent.log line grammar: a single agent_log_format module defines
  the grammar, local_agent delegates to it, and a Claude stream-json
  translator feeds ClaudeCliDriver dispatch while raw NDJSON is kept in a
  sidecar (#687, #690, #691, #692, #699).

Process

• Started an ADR practice: docs/adr/ with an index, a template and the first
  four decision records (#682).
• Introduced this CHANGELOG (#681).

Fixed

Plan-lock starvation

• Merge adjudication moved to the start of the tick (#704), and the plan lock
  is now released around dispatch_story and review_story, guarded by a
  dispatch lease so a released lock cannot double-dispatch (#697, #698, #709,
  #712); per-plan per-tick dispatch is capped (#708).
• The .agent_done marker is claimed before its event is published and is
  written as soon as finish_if_green decides the run is finished (#703,
  #705).
• reconcile_fn gained the same bounded join watchdog scan_fn already has
  (#700).
• The daemon exits for a launchd restart after repeated abandoned watchdog
  workers, since a leaked plan lock is only released by process death (#706).

Watchdog

• The stale-activity watchdog floors the activity age at the current
  dispatch's elapsed time, so a redispatch into a reused worktree is not
  killed at elapsed 0s (#711), and records the last agent.log line and the
  branch commit count when it kills a dispatch (#702).

Done-bar

• The done-bar no longer narrows to a story's own new test file (cf48767),
  guarded by tests/unit/test_done_bar_not_narrowed.py.

v0.1.0 — first public release

v0.1.0 - first public release

Spend tokens on judgment, not typing.

An autonomous SDLC pipeline that splits software engineering along the line
where model cost actually differs: a frontier model decomposes the work, plans
it, reviews the diff and adjudicates anything risky, while a local model writes
the implementation at no marginal cost. What makes the cheap half trustworthy
is inspection - the same finding Michael Fagan published at IBM in 1976.

In this release:

• MCP server, web dashboard and scheduler daemon
• Multi-provider dispatch (claude, ollama, lmstudio, mlx, litellm) with
  per-role provider/model routing
• TDD enforced before implementation; independent review gate;
  acceptance-oracle grading
• Risk-tiered overlord that stops for a human on anything irreversible
• Merge gate that re-runs the suite against the rebased branch
• CI green on Linux (3.12/3.13/3.14) and macOS

This is a single-maintainer research project, not a maintained product with an
SLA. Its failure modes are real and documented - see the README's
"Reliability & limitations" section before pointing it at anything you care
about.