chipping-orchestrator release timeline

Source material for a public timeline of chipping-orchestrator: all 15 published releases, from v0.13.0 back to v0.1.0, with links to the GitHub release history and tagged changes. Releases run newest first. The table gives each milestone a date, version, and suggested title; the lists below supply its functionality, upgrade guidance, and supporting sources.

Dates are GitHub release publication dates in UTC, rather than commit or tag dates. Each list describes what shipped at that tag. Changes after v0.13.0 are outside this timeline. Historical names and controls are retained where they explain a release; use the documentation index for current operation and configuration.

Timeline milestones

Published date Release Suggested timeline title
2026-10-06 v0.13.0 Developer reports, verification evidence, and Poetry packaging
2026-09-15 v0.12.0 Human size approval, lifetime budgets, and artifact cleanup
2026-08-31 v0.11.1 Cumulative PR limits and the chipping-orchestrator name
2026-08-26 v0.11.0 Size-gated publication and reusable implementation splits
2026-08-20 v0.10.1 Visible automatic recovery and stranded-fix publication
2026-08-19 v0.10.0 Architecture discussions and skill provenance
2026-08-10 v0.9.0 Installable CLI and canonical package entrypoints
2026-07-23 v0.8.0 Session adoption analytics and a modular runtime
2026-07-09 v0.7.0 Trusted conversations, pause controls, and issue usage
2026-06-30 v0.6.0 Agent trajectories and skill telemetry
2026-06-23 v0.5.0 Repository context and resilient agent sessions
2026-06-09 v0.4.0 Faster analytics and guarded workflow transitions
2026-06-03 v0.3.0 Analytics dashboard and scheduled issue processing
2026-05-26 v0.2.0 Documentation, Q&A, verification, and parallel work
2026-05-19 v0.1.0 GitHub issues become reviewed pull requests

v0.13.0

Published 2026-10-06. Sources: release notes and tagged changes.

Added functionality

  • Revisioned developer reports published as dedicated pull-request comments before implementation or fix work reaches review. Reports describe the final change, its rationale, and unresolved risks; a report-only update needs no commit. Durable publication receipts recover interrupted deliveries without duplicate comments or lost review bookkeeping.
  • Automated reviews and approvals bound to the exact pull request, head, requirements revision, and developer report. Changed reports or requirements invalidate an earlier approval even when the head is unchanged. Orchestrator-owned rebases and conflict rewrites trigger a report refresh before another review.
  • Append-only workflow verification artifacts for local checks and reviewer declarations. A concise visible summary identifies the outcome and tested commit; exact commands, exit statuses, and output remain in a hidden payload. Readers also accept the earlier visible-evidence format.
  • Persisted reviewer verdicts and evidence, reconciled before approval or a change-request handoff. Approval requires current passing evidence covering every configured VERIFY_COMMANDS command; with none configured, it still needs at least one passing check. Equivalent-tree approval squashes preserve proved evidence.
  • Antigravity CLI (agy) support for every agent role, including resumed sessions, usage, skill telemetry, and trajectories. Tracked asynchronous commands must finish before success; premature developer exits receive bounded command recovery, and exhausted recovery parks as a retryable execution failure.
  • Host-local issue writer claims coordinate dispatch, base refresh, child writes, and close recovery across pollers sharing one WORKTREES_DIR. Guarded report, evidence, and reviewer-state commits preserve unrelated pinned fields and refuse decisions whose recorded subject moved.
  • Ordinary re-decomposition inside a late-split lineage preserves ancestry and depth limits, records new snapshot consumers before seeding them, and holds children until their seeds and snapshot protection are proved. Replacements inherit no size authorization.
  • Optional TERMINAL_ARTIFACT_CLEANUP_WINDOW and TERMINAL_ARTIFACT_CLEANUP_TIMEZONE schedule automatic cleanup in a local window, taking precedence over the interval. On-demand cleanup retains its host-quiet requirements.
  • park_awaiting_human analytics and correlations back to the agent run or evaluation that caused the wait. Repeated polls of an existing park emit no new transition; the records use existing analytics storage.
  • Poetry packaging and a lockfile for source deployments, with package installation, upgrade, rollback, and manual release instructions. Installed runs load user configuration, target explicit clones, and answer --help without repository configuration. Runtime dependencies remain PyGithub and psycopg.

Improvements and changes

  • ALLOWED_ISSUE_AUTHORS must name at least one GitHub login. Installed packages require REPOS, and every target must be the root of an existing git checkout; invalid setup fails before connecting to GitHub or launching an agent.
  • DEPENDENCY_POLL_EVERY_N_TICKS, default 5, reduces blocked/umbrella dependency reads. Lazy worker repositories and pull-request filtering avoid unnecessary GitHub detail requests.
  • PR_REF_IN_SUBJECT, default on, normalizes approval and documentation commit subjects to the pull request's reference, removing the tracked issue's reference. Approval can rewrite a one-commit branch for its subject alone; generated PR titles omit the tracked issue's numeric suffix.
  • Trusted feedback settles only the exact issue and pull-request comment batches delivered to the developer. Separate surface watermarks preserve later comments, and recovered fix rounds avoid spending review rounds twice.
  • Published MkDocs documentation, grouped navigation, contributor guidance, issue forms, and strict source/build link checks. Poetry replaces uv in CI and launch tooling; locked runtime major versions and dependency cooldowns bound updates without adding runtime dependencies.

Upgrade notes from v0.12.0

  • Set ALLOWED_ISSUE_AUTHORS before restarting. An empty allowlist prevents every launch mode, including --once and on-demand cleanup. Workflow labels are unchanged and pinned-state additions require no bulk issue-state migration.
  • For source deployments, install Poetry 2.5.1, stop the service, and recreate .venv/ from poetry.lock with env -u VIRTUAL_ENV -u CONDA_PREFIX poetry sync. Select --with docs and/or --with dashboard when needed. Replace uv run in service units, crontabs, and applicable VERIFY_COMMANDS; see the operator migration procedure.
  • Installed packages read ~/.config/chipping-orchestrator/.env and require REPOS entries naming existing local clones; REPO and TARGET_REPO_ROOT remain source-checkout developer settings. Source checkouts keep reading their own .env. Confirm each target's token at ~/.config/<owner>/<name>/token, or use an env-only GITHUB_TOKEN.
  • Preserve the current absolute WORKTREES_DIR when upgrading an active deployment: its default follows the first configured target clone. Every poller on one host must use that same root and this version's claim protocol, with a separate checkout and environment per daemon. Stop all older pollers before starting coordinated ones.
  • Review the developer report and reviewer verification contracts. Reports and verification are published separately; absent, stale, failed, or incomplete approval evidence parks the issue. Existing work without a recorded report may pause at report_undeliverable; follow its reply instructions. The local verification gate still runs after a valid reviewer approval. Verification payloads remain readable in comment source: keep secrets out of commands and recorded output.
  • Set DEPENDENCY_POLL_EVERY_N_TICKS=1 to retain every-tick dependency dispatch. Set PR_REF_IN_SUBJECT=off where GitHub's default squash-merge message would add a second PR reference. A configured cleanup window requires a valid IANA timezone; an unset window preserves the daily interval default.
  • Before returning to v0.12.0, stop every v0.13.0 poller and finish outstanding report, verification, reviewer, and replacement-split recovery. The earlier binary does not reconcile these new records. Follow the Poetry rollback procedure when restoring a source deployment, and the late-state drain when returning to a version that predates late-split state.

v0.12.0

Published 2026-09-15. Sources: release notes and tagged changes.

Added functionality

  • Oversized single verdicts require human approval and an explanation of why no safe split exists. The exact candidate is authorized with /orchestrator authorize-oversized <commit>, preserving the decision through publication retries.
  • Authorized exemptions can follow orchestrator-owned squashes and clean rebases only when contribution fingerprints prove the full contribution unchanged. Added fixes, documentation, descendants, and split children are measured normally.
  • Durable recovery of interrupted squashes/rebases and already-landed pushes, verified against recorded commits, PR, remote base, and lease. Uncertain evidence remains parked for a human decision.
  • Late-split children must each own implementation, tests, and documentation and declare an all-path added-line estimate strictly below the frozen ceiling. Estimates survive recovery and are shown on child issues.
  • MAX_AGENT_RUNS_PER_ISSUE introduces a durable lifetime allowance, defaulting to 50, charged before each launch across roles, fresh runs, resumes, replacements, timeouts, and interruptions. Resets and restarted cycles retain spent runs.
  • /orchestrator add-agent-runs N grants 1 to 50 more runs after a lifetime-budget park; 0 in the global setting disables enforcement while accounting continues.
  • Exhausted daily retry budgets require a trusted /orchestrator continue for one fresh attempt, including across new days, requirement edits, and configuration changes. It does not override the lifetime allowance.
  • Finished-issue artifact cleanup runs daily or on demand with --cleanup-terminal-artifacts, reclaiming eligible local worktrees and local/remote branches only on a quiet host with closed terminal issues and preserved, inactive, clean work.
  • New agent_run_budget and late_transfer observability records, plus analytics results for terminal artifact cleanup.

Improvements and changes

  • Base read/fetch failures retry silently for three consecutive misses; the fourth asks for human intervention. Measurement failures identify the step and scrubbed diagnostic, suppressing repeated notices for the same cause.
  • Publication receipts are bound to their PR, and pushes to merged or closed PRs are refused. Completed splits release their children without stale measurement parks.
  • CI adds Python 3.14; analytics Compose requires existing bind-source directories; lint coverage and locked tooling are refreshed without adding runtime dependencies.

v0.11.1

Published 2026-08-31. Sources: release notes and tagged changes.

Added functionality

  • Cumulative size measurement of the complete prospective PR before every orchestrator-owned update, including fixes, recovery pushes, rebases, conflict resolutions, approval squashes, and final documentation.
  • Oversized updates stay local while the existing implementation PR receives a durable do-not-merge hold. Frozen stage, PR, and prior-head records preserve feedback, review counters, documentation completion, and conflict state.
  • A single decision publishes to the same PR and resumes the originating stage. A split snapshots the work, closes the superseded PR, creates children, and converts the parent to an umbrella, with recovery across each handoff.
  • Late-split telemetry distinguishes pre_publication and post_publication, with originating stage/PR/head context.

Improvements and changes

  • The distribution and installed command are renamed from agent-orchestrator to chipping-orchestrator; repository references and default agent identity are updated. The module launcher and run.sh remain valid.
  • Transient provider overload/refusal is retryable: developer runs use the continue control, while reviewers recover automatically. Unclassified skills resolve against the repository catalog when exactly one source level matches.
  • CI covers Python 3.12 and 3.13, cancels superseded PR runs, applies job timeouts, and smoke-tests the built wheel in isolation.

v0.11.0

Published 2026-08-26. Sources: release notes and tagged changes.

Added functionality

  • A pre-publication size gate through MAX_ADDED_LINES, defaulting to 4000 when decomposition is enabled. It measures additions over a frozen remote base across all paths; exactly the limit is allowed, and binary content adds no lines.
  • Clean committed publication and recovery paths freeze the candidate/base before measurement; failed measurements hold the work without rerunning the developer unnecessarily.
  • Oversized work returns to the decomposer for single, question, or split adjudication. At this release, single permits only the exact accepted commit; questions invite revisions, and recursive splits have a lineage-depth bound.
  • Immutable late-split snapshot refs preserve committed implementation before child creation. Children receive the exact snapshot and lineage so they can reuse the work; refs remain until their recorded direct consumers end.
  • Durable plan-PR holds and recovery across child creation, umbrella conversion, superseded branches/worktrees, notices, and snapshot reclamation.
  • Closed-owner cancellation and cleanup, followed by an explicitly authorized fresh cycle when the issue is reopened and rejected is removed.
  • Correlated audit/analytics events for late measurement, verdicts, failures, snapshots, cleanup, cancellation, and restart.

Improvements and changes

  • Exact approved commit and tree checks before and after pushing protect the final publication handoff.
  • SHA-pinned CI actions, CodeQL, OpenSSF Scorecard, scheduled lockfile vulnerability scans, test coverage reporting, and a private vulnerability-reporting policy expand repository security checks.

v0.10.1

Published 2026-08-20. Sources: release notes and tagged changes.

Added functionality

  • An automatic-recovery follow-up on issues after transient push, developer-timeout, or reviewer failures heal under validating/fixing. It is crash-safe, at most once per park episode, and sends no additional human mention.

Improvements and changes

  • The fixing stage publishes clean commits stranded by interrupted or live-paused runs before returning to validation without new feedback. Review accounting changes only when publication succeeds.
  • Hardened Git status/index/reset operations use absolute worktree paths, supporting relative WORKTREES_DIR settings.
  • No label, configuration, or dependency migration is required for this patch.

v0.10.0

Published 2026-08-19. Sources: release notes and tagged changes.

Added functionality

  • The operator-applied discussion stage supports multi-round architecture conversations through DECOMPOSE_AGENT, including when ordinary decomposition is disabled.
  • A confirmed discussion can produce one validated Markdown plan and a plan PR, with publication and crash recovery handled by the orchestrator. Merging the plan ends the issue as done; closing it unmerged yields rejected.
  • Explicit handoff from discussion into implementation reuses the reviewed plan branch and PR.
  • Codex trajectories cover MCP calls, web search, patches/file changes, todo-list lifecycles, assistant messages, and terminal outcomes. Bounded item accounting and metadata placeholders expose unsupported or excluded items.
  • Skill analytics preserve source provenance (project, user, harness, or unknown), separating same-named skills by level in event records, read models, and dashboard sections.
  • A dashboard 30D date preset and sortable skill-level columns.

Improvements and changes

  • Machine-driven labels gain the workflow: namespace. Startup migrates legacy labels, compatibility reads remain, and human-facing controls retain their bare names. Analytics stage identifiers remain unchanged.
  • Trusted-reply and relabel guards protect question/discussion work from accidental implementation publication.
  • The README becomes an operator guide, and docs/README.md indexes focused reference pages. Documentation-link checks join the repository suite.

v0.9.0

Published 2026-08-10. Sources: release notes and tagged changes.

Added functionality

  • A Hatchling package build and an installed agent-orchestrator console command, the project's name at this release.
  • python -m orchestrator as the canonical module launcher, sharing CLI composition with the installed command.
  • Canonical analytics-dashboard and trajectory-viewer app entrypoints under orchestrator/apps/.

Improvements and changes

  • Responsibility-owned packages replace the legacy flat modules and compatibility facades retained in v0.8.0. Workflow, configuration, pinned-state, and event contracts are preserved through the source-layout migration.
  • Operator commands and downstream imports must use canonical owners; run.sh adopts the new module launcher.
  • The trajectory viewer explicitly shows an empty skills result. Pytest discovery is restricted to tests/, avoiding the operator-owned analytics database volume.
  • Tests and documentation follow the package layout, with repository checks for import boundaries and public APIs.

v0.8.0

Published 2026-07-23. Sources: release notes and tagged changes.

Added functionality

  • Skill observations distinguish confirmed Claude loads, inferred Codex SKILL.md reads, and incidental references.
  • Skill adoption is aggregated by logical agent session, preventing repeated resumes from inflating adoption counts.
  • Session adoption becomes the dashboard's primary skill metric, with sortable availability, adoption, invocation-load, and incidental-reference columns. Invocation trigger rates and the matrix remain available as diagnostics.

Improvements and changes

  • Corrected adoption counts, availability metadata, legacy fallback, reporting windows, and developer-fix attribution.
  • Dirty-worktree checks and squash/reset/rollback operations were hardened against agent-planted core.fsmonitor helpers.
  • Runtime, workflow, GitHub, Git, agents, and observability were split into focused modules behind compatibility facades, preserving imports, patch points, state fields, labels, and the existing module launcher.
  • Wemake Python Styleguide became a required development and CI check, alongside expanded module-boundary tests.

v0.7.0

Published 2026-07-09. Sources: release notes and tagged changes.

Added functionality

  • The paused label freezes an in-flight issue, with checks after agent runs across delivery and conversation stages.
  • /orchestrator continue retries recoverable session failures across stages; session-limit messages become retryable parks.
  • Per-issue accumulated agent usage in pinned state and usage verdict comments when an issue ends.
  • Run-level trajectory token/cost summaries and Claude per-turn usage detail in recorded trajectories and the viewer.
  • Local Codex skill-catalog discovery, trigger-rate columns, and sortable per-skill matrix views.
  • Decomposer explanations are surfaced when an issue is judged to be a single task.

Improvements and changes

  • Issue-author allowlisting becomes a comment trust boundary for prompts, requirement drift, resume signals, PR feedback, and retry controls. Only orchestrator-authored pinned comments supply durable state.
  • Authenticated Git operations reject local proxy/TLS overrides, and the launcher keeps running when self-update fails.
  • hold_base_sync was removed. Workflow and observability internals were refactored, and CI gained line-length checks.

v0.6.0

Published 2026-06-30. Sources: release notes and tagged changes.

Added functionality

  • Opt-in agent trajectory recording through TRAJECTORY_LOG_PATH, with redaction and truncation, separate from analytics.
  • A Streamlit trajectory viewer for tool calls/results, text turns, outputs, run selection, and fixture-marked runs.
  • Opt-in TRACK_SKILL_TRIGGERS parsing for Claude and Codex, skill fields on agent-exit records, and skill_triggered audit events.
  • Repository skill-catalog records, dashboard skill-trigger rates, and a per-skill trigger matrix with run totals.
  • Bounded shutdown through SHUTDOWN_GRACE_SECONDS, including process-group cleanup of agents and verification commands. Shutdown-killed runs are classified as interrupted, and resume stages ignore their partial results.

Improvements and changes

  • Clean implementation commits stranded after an agent timeout can be recovered and published.
  • Cached label objects and CLOSED_ISSUE_SWEEP_EVERY_N_TICKS reduce GitHub reads in multi-repository operation.
  • Dashboard analytics default to seven days; truncated PR output is identified, and Claude message fallback is safer.

v0.5.0

Published 2026-06-23. Sources: release notes and tagged changes.

Added functionality

  • Prompts can expose tracked sibling repositories, local roots, and base branches for read-only cross-repository reference. EXPOSE_TRACKED_REPOS disables this context globally; single-repository deployments do not receive it.
  • DEV_SESSION_MAX_RESUMES proactively rotates long-lived developer sessions. Claude context-overflow failures get one immediate replacement session, grounded in the issue, comments, branch state, and available repository context.
  • Commit-producing prompts, PR titles, and squash commits follow the target repository's recent commit style, including repository-specific prefixes and a documentation pass without a mandatory docs: prefix.

Improvements and changes

  • Clean committed fixes stranded locally are published even when a resumed developer makes no additional commit; this recovery takes priority over a fixing acknowledgment.
  • Commit-producing agents receive foreground-execution guidance for builds, tests, and servers.
  • Bot-authored PRs are excluded from community-contribution routing; locked dependencies and operator references were refreshed.

v0.4.0

Published 2026-06-09. Sources: release notes and tagged changes.

Added functionality

  • Dashboard timezone selection, agent-role cost splits, and cached versus uncached cost breakdowns by review round and workflow stage.
  • Daily analytics rollups, batched sync inserts, existing-record hash checks, and sync progress/timing output.
  • Typed workflow labels, typo detection, and a transition-table guard against invalid workflow-state changes.
  • Community-contribution PR labeling and routing to human reviewers.

Improvements and changes

  • Dashboard layout and styling were redesigned. Reused database connections, parallel reads, consolidated queries, static metadata caching, and rollup-backed widgets reduced dashboard work.
  • Reviewer-requested developer changes use fixing, with recovery for stuck, unacknowledged, and nothing-to-fix rounds.
  • PR branches gained repository namespaces; backlog, umbrella, and no-agent family buckets received corrected capacity accounting. Closed sweep-labeled issues can finish as done.

v0.3.0

Published 2026-06-03. Sources: release notes and tagged changes.

Added functionality

  • A Streamlit analytics dashboard with Plotly charts, shared styling, and views focused on analyzing workflow behavior.
  • JSONL-to-Postgres synchronization and a bundled Docker Compose Postgres service for local analytics storage.
  • A Postgres dashboard read model with aggregate queries, an agent-run view, hot-path indexes, and GPT-5.5 pricing.
  • A process-local issue scheduler with family-aware dispatch and completion reaping once per polling pass.

Improvements and changes

  • Reviewer approval hands work to a final documentation pass before the human merge handoff. Linked PRs already merged are reconciled to done.
  • Automatic merging was removed: in_review became permanently dependent on a human merge decision.
  • Agent and verification subprocesses receive environments stripped of production secrets.
  • Lockfile-based installation, Dependabot, dependency review, and read-only CI permissions strengthened dependency handling.
  • Git and analytics responsibilities were split into focused modules; basic and advanced environment templates were separated. Launcher pull-failure handling and documentation-stage base synchronization were corrected.

v0.2.0

Published 2026-05-26. Sources: release notes and tagged changes.

Added functionality

  • A dedicated documenting stage for documentation passes after implementation and later code-changing fixes.
  • A fixing stage that waits for a quiet feedback window, tracks unread PR feedback, and resumes the developer session.
  • Read-only issue Q&A through the question label, including follow-up conversations without opening a pull request.
  • Local verification through VERIFY_COMMANDS and VERIFY_TIMEOUT before approved work reaches in_review.
  • Concurrent work within and across repositories, bounded by per-repository and global issue limits.
  • backlog to postpone pickup, plus the historical hold_base_sync control to suspend base synchronization. The latter was removed in v0.7.0.
  • Automatic base refresh into issue worktrees and routing or session resumption when humans edit issue requirements.
  • /orchestrator add-review-rounds N to extend an exhausted review allowance, and one human notification when a PR is ready for merging.
  • Shell-like agent command specifications with model/effort arguments, persisted across session resumes, and configurable Git remote names globally or per repository.
  • Audit events and an analytics JSONL sink with retention, stage evaluation timing, agent-run records, parsed token usage, and estimated costs.

Improvements and changes

  • Stage-specific workflow modules and tests, expanded operator documentation, stale Claude-session replacement, rejected-branch cleanup, authenticated target fetches, and correct launcher shutdown on Ctrl+C or SIGTERM.

v0.1.0

Published 2026-05-19. Sources: release notes and tagged history.

Added functionality

  • GitHub issue processing through workflow labels and a pinned JSON state comment, allowing progress to survive orchestrator restarts without a workflow database.
  • Configurable local codex and claude backends for implementation, review, and decomposition, with reusable agent sessions and human notifications.
  • Isolated per-issue git worktrees, branch publication, idempotent PR creation, and branch cleanup after merge.
  • Manifest-based decomposition into child issues, dependency ordering, umbrella parents, and a final documentation child for split work.
  • A bounded reviewer/developer loop, approval squashing, conflict-resolution handling, and done / rejected outcomes.
  • Optional automatic merging after approval and green CI, disabled by default. This option was removed in v0.3.0.
  • Multi-repository operation through REPOS, separate target clones, repository-specific credentials, and worktrees namespaced by repository.
  • Configurable daily implementation retries and review/conflict round limits, issue-author allowlisting, and multiple human handles for requests requiring intervention.
  • Token storage outside .env, credential stripping from agent environments, authenticated pushes hardened against repository-controlled Git configuration, and lease-protected push retries.
  • A polling launcher with self-update/restart behavior, file logs, and GitHub Actions lint and test checks.