Skip to content

Latest commit

 

History

History
335 lines (299 loc) · 18.1 KB

RELEASE_NOTES.md

File metadata and controls

335 lines (299 loc) · 18.1 KB

Release Notes

0.3.0

Release-readiness memo for the current 0.3.0 local-first CLI release line of the Multi-Agent Coding Orchestrator.

Implemented Scope

  • Crate packaging metadata now includes crates.io keywords and categories for command-line and development-tool discovery.
  • Task-path proposal helpers are more precise: word-boundary symbol matching, no arbitrary first-file fallback, Git-ignore-aware planning repository walks, and degraded-scan diagnostics surfaced in autopilot and inbox reports.
  • Supervisor child-report enforcement is hardened with lenient JSON report recovery from noisy child output, structural failure for assigned workers that do not report, parent audit of zero-worker children with non-empty diffs, primary-worktree integrity rechecks, opt-in corrective retry through max_child_retries, and claim-conflict diagnostics that name current owners.
  • Supervise evidence handling now includes worker-evidence cross-checks, canonical child-report artifacts, per-assignment task override, and stable child-diff baselines for auditable child results.
  • maco supervise run now defaults --max-concurrent-children to auto, deriving a resource-bounded child-execution limit from measured host capacity. A positive numeric value remains available as an explicit override, with --max-concurrent-children 1 selecting serial execution. Disjoint path sets can run concurrently; overlapping assignments serialize with scan-ahead scheduling under the sync claim overlap rules. Retries and parent audits retain their assignment slot, ordinary failures remain isolated, fatal aborts stop new starts and join active calls, plan-indexed outcomes preserve deterministic report order, journal appends are synchronized, and concurrent invocations use unique scratch roots. Zero is rejected before artifacts are reserved. max_child_assignments remains the plan fan-out budget rather than the concurrency bound.
  • New terminal cross-runtime CONSULTANT role with maco consult ask and maco consult artifacts list/latest/prune. Consultation is fake-first by default, supports explicit read-only Codex and Claude CLI adapters, and can be wired into supervise planning only by opt-in configuration.
  • Focused inbox unit tests cover the reaction loop and fixed two exposed bugs: token-like privacy_scan predicate matching and .maco runtime-path filtering in lock-overlap checks.
  • maco inbox run / inbox watch and maco inbox workspace run / workspace watch execute. inbox run accepts optional rolling-quota ceilings --max-rolling-tokens, --max-rolling-cost-usd, and --rolling-window-seconds. A quota is bound only when at least one ceiling is set; the window defaults to 86400 seconds. inbox watch and workspace run/watch do not expose those flags. Non-dry-run github_git, github_pr, and github_full still fail closed until an external reviewer is bound.
  • maco supervise run and maco autopilot run accept optional --role-category (delegating_coordinator, non_delegating_terminal_worker, read_only_researcher, read_only_review_auditor). Resume of an existing supervise run refuses a new override. Later retired: autopilot plan/run fail closed; use maco <instruction>.
  • maco eval-harness run-v2 always parses the v2 manifest schema. maco evaluation rescore re-scores a stored results document under a named objective profile without overwriting the stored file. Real network providers remain refused.
  • maco worktree remove --force has a hardened removal fallback and idempotent re-removal behavior for safer cleanup of stale or partially removed agent worktrees.

Verification Expectations

This document does not authorize tagging, pushing, publishing a crate, or creating a GitHub release. Before any separate release-manager operation, run the project checks from a clean working tree:

cargo fmt -- --check
cargo test
cargo clippy --all-targets -- -D warnings

If Rust is not available globally, run the same commands through the documented Nix shell.

Known Limitations And Risks

  • Richer merge conflict classification remains a known limitation. Current apply uses Git apply safety checks and reports structured blockers, but does not classify conflicts by symbol or dependency impact.
  • Full semantic task planning remains roadmap work. Current task-to-path proposals are conservative helpers for autopilot and inbox defaults; claim gates remain authoritative.
  • Consultant results are advisory evidence only. Real Codex and Claude consultant adapters are explicit local-process opt-ins, remain read-only, and do not override assigned ownership, validation, review, or merge gates.
  • Effectful maco supervise run executes through the live supervisor gates after acquiring the capability-bound repository-cleanliness input and creating managed child worktrees. A dirty primary fails with the required remedy before worktree creation. The in-process Fake runtime is always non-publishable. Plan, status, collect, resume, re-audit, and artifact inspection remain available.
  • PR and issue publication remain intentionally narrow and opt-in for live GitHub paths. Fake forge and fake inbox data stay the no-network defaults.
  • Real network LLM providers remain absent from maco agent run and must be separately approved, adapter-bound, and covered by no-network invariant tests before introduction.
  • Automatic merge remains intentionally absent. Autopilot and inbox can record merge intent, but leave human review and merge as the next action.

Release Operations

This memo does not tag a commit, push a branch, publish a crate, create a GitHub release, or publish any release artifact. Those operations remain separate release-manager actions.

0.2.0

Release-readiness memo for the current local-first CLI release line of the Multi-Agent Coding Orchestrator.

Implemented Scope

  • Local Git repository initialization and linked worktree management.
  • Durable exclusive path claims for coordinating edit ownership.
  • Local JSON orchestration plan validation and command-backed execution with dependency ordering, timeouts, per-agent validation commands, optional patch output, checkpoint writes, checkpoint resume, completed-agent skipping, and path-boundary checks.
  • Safe reuse=reset handling for clean, unclaimed stale worktrees, with refusal for dirty, untracked, or actively claimed worktrees.
  • Read-only repository mapping, including parser-backed Rust semantic maps and symbol/path/risk queries. Risk reports include touched symbols, dependency impacts, and impacted files for changed paths. Repository intelligence now excludes local-only .agents/temp, .agents/storage, and .agents/live paths while preserving durable .agents documentation.
  • Standalone repo-local semantic intent coordination commands: maco coord preview/claim/status/release/release-agent for path, module, and symbol intents.
  • Worktree diff collection, orchestration result collection, merge preview, and guarded merge apply with dirty-primary, stale-base, unclaimed-edit, validation, and apply-check gates.
  • Direct merge preview/apply --validation-report <file> support for external validation JSON, including machine-readable blocked apply reports with blocker details, related paths, and distinct missing, not-run, skipped, and failed validation blockers.
  • merge preview/apply --require-validation requires at least one passed validation report. merge apply --validation-command <command> validates a temporary merged candidate before applying to the primary worktree; failure blocks and leaves the primary worktree unchanged.
  • Provider-neutral LLM provider listing and prompt preview without network calls or provider credentials.
  • Local fake-provider-backed maco agent run execution in isolated worktrees with durable claims, boundary checks, validation, merge-preview reporting, and no real network providers by default.
  • Opt-in process-level supervisor-of-orchestrators mode with maco supervise plan/run/status/collect, serial Codex CLI child subprocess execution, isolated child worktrees, durable path claims, semantic coordination metadata, structured reports, and no-network fake subprocess tests by default.
  • maco supervise run verifies actual child worktree Git-visible changes against the supervisor's starting primary HEAD, rejects unauthorized paths, and uses max_child_assignments as a child assignment fan-out budget rather than a parallel execution limit. max_child_processes remains accepted as a legacy JSON alias.
  • maco supervise artifacts list/latest/prune, maco autopilot artifacts list/latest/prune, and maco inbox artifacts list/latest/prune inspect and prune durable run artifacts with generated run ids, explicit run-id reuse refusal, dry-run pruning, and keep-count retention.
  • First local-first autopilot workflow with maco autopilot plan/run/status/collect. It normalizes task files or JSON plans, writes durable .maco/autopilot/runs/<run-id>/ artifacts, launches supervised child work through a fake local subprocess by default, publishes through the PR safety gates, runs an independent reviewer, and records repair attempts for validation failures or blocking review findings.
  • Autopilot and inbox use path-scoped safety refusals before launching work: overlapping durable sync claims, semantic coordination intents, and active/blocked live claim locks block only the overlapping target paths and return machine-readable path and lock details.
  • Autopilot task-to-path proposals are now conservative helpers informed by repository paths, Rust semantic names, and task wording instead of defaulting only to README.md; claim gates remain authoritative.
  • Standalone maco review pr <number|url> command with a deterministic fake structured review report by default. Review findings include severity, path, summary, suggested fix, and blocking status. CI reaction is reported as unsupported with ci_reaction_supported=false.
  • Fake-first inbox reaction loop with maco inbox scan/run/status/collect/watch. It uses deterministic local fake issue and PR data by default, redacts public JSON, skips unsafe or duplicate items, converts issue intake plus PR review and failing CI context into autopilot plans, writes .maco/inbox/runs/<run-id>/ artifacts, and preserves the no-network, no-credentials, no-automatic-merge default.
  • Cross-repository inbox workspace supervision with maco inbox workspace scan/run/watch --config <path>. Workspace configs select enabled repositories, per-repo or default permission modes, per-repo item limits, issue/PR inclusion, labels, strict failure behavior, and safety flags. Aggregate reports use public-safe config and artifact paths, include repo counts and per-repo scan or run reports, continue across repository refusals in non-strict mode, fail the aggregate command in strict mode, and write workspace artifacts under .maco/inbox-workspace/runs/<run-id>/.
  • Fake PR review and failing CI context enrich inbox-generated repair plans with paths, reasons, and validation expectations. Richer live GitHub review and CI reaction remains explicit opt-in and future hardening work.
  • Opt-in PR and issue publication adapters. maco pr preview and maco issue preview are non-creating previews. maco pr publish --forge fake|github and maco issue create --forge fake|github either use the deterministic local-only fake forge or, with explicit --forge github, shell out to local git and gh. PR preview and publish support --require-validation and block missing validation evidence.
  • Public maco live liveness commands for repo-local Markdown claims: status, validation, heartbeat refresh, and project-owner override release to handoff with audit logging.
  • Command-backed orchestration and opt-in provider-proposed commands execute trusted local shell commands. Worktrees and path claims are repository change controls, not OS or filesystem sandboxes; path claims enforce Git-visible repository changes.

Verification Expectations

This document does not authorize tagging, pushing, publishing a crate, or creating a GitHub release. Before any separate release-manager operation, run the project checks from a clean working tree:

cargo fmt -- --check
cargo test
cargo clippy --all-targets -- -D warnings

If Rust is not available globally, run the same commands through the documented Nix shell.

Known Limitations And Risks

  • Merge apply uses Git apply safety checks and reports blockers, but conflict classification is not yet semantic or symbol-aware.
  • Full semantic task planning and automatic path-claim proposal are not implemented. Current task-to-path defaults are conservative helper proposals only; they do not replace human review or claim gates.
  • Bounded parallel supervisor child execution is explicitly deferred. Supervisor child execution remains serial, and max_child_assignments is a fan-out budget, not a concurrency limit.
  • PR and issue publication are intentionally minimal. The fake forge is local-only and covered by no-network tests. GitHub mode depends on local git and gh setup and is selected only with explicit --forge github for direct PR commands or forge_mode: "github" for autopilot plans. Issue triage metadata is limited to title, body, and labels. Inbox GitHub intake is also explicit opt-in; deterministic fake data remains the default. Richer live GitHub PR status, check, review import, stale metadata handling, and CI reaction remain future work.
  • Autopilot intentionally omits automatic merge. It accepts and reports auto_merge=true, but always writes auto_merge_performed=false and leaves human review and merge as the next action. Inbox reactions keep the same boundary and do not apply or merge repaired work automatically.
  • Workspace inbox is cross-repository supervision, not approval or merge automation. Safety flags do not enable automatic approval or automatic merge; workspace reports keep auto_approval_performed=false and auto_merge_performed=false. GitHub PR publication and source comments remain limited to explicitly configured permission modes, and Git-only publication does not create GitHub PRs.
  • Live claim liveness uses human-readable Markdown parsing. Missing timestamps are reported as stale risk instead of blocking command execution, and malformed timestamps are reported as unknown liveness.
  • Real network LLM providers are still absent. maco agent run currently supports only the deterministic local fake provider.
  • Real Codex authentication and provider execution for maco supervise are opt-in operator responsibilities. Default tests use fake subprocesses and do not require network access, provider credentials, or a real Codex login.
  • Repo-level validation commands run in the primary worktree after agent commands complete; use per-agent validation for checks that need to see unmerged agent worktree changes.
  • merge apply --validation-command validates the candidate merged state before primary mutation, but merge apply does not automatically run post-apply validation after a successful primary apply. Release managers should run final project checks after accepted changes are applied.

Release Operations

This memo does not tag a commit, push a branch, publish a crate, create a GitHub release, or publish any release artifact. Those operations remain separate release-manager actions.

0.1.0

Release-readiness memo for the first local-first CLI slice of the Multi-Agent Coding Orchestrator.

Implemented Scope

  • Local Git repository initialization and linked worktree management.
  • Durable exclusive path claims for coordinating edit ownership.
  • Local JSON orchestration plan validation and command-backed execution with dependency ordering, timeouts, per-agent validation commands, optional patch output, checkpoint writes, checkpoint resume, completed-agent skipping, and path-boundary checks.
  • Safe reuse=reset handling for clean, unclaimed stale worktrees, with refusal for dirty, untracked, or actively claimed worktrees.
  • Read-only repository mapping, including parser-backed Rust semantic maps and symbol/path/risk queries. Risk reports include touched symbols, dependency impacts, and impacted files for changed paths.
  • Worktree diff collection, orchestration result collection, merge preview, and guarded merge apply with dirty-primary, stale-base, unclaimed-edit, validation, and apply-check gates.
  • Direct merge preview/apply --validation-report <file> support for external validation JSON, including machine-readable blocked apply reports with blocker details and related paths.
  • Provider-neutral LLM provider listing and prompt preview without network calls or provider credentials.
  • Local fake-provider-backed maco agent run execution in isolated worktrees with durable claims, boundary checks, validation, merge-preview reporting, and no real network providers by default.

Verification Expectations

This document does not authorize tagging, pushing, publishing a crate, or creating a GitHub release. Before any separate release-manager operation, run the project checks from a clean working tree:

cargo fmt -- --check
cargo test
cargo clippy --all-targets -- -D warnings

If Rust is not available globally, run the same commands through the documented Nix shell.

Known Limitations And Risks

  • Merge apply uses Git apply safety checks and reports blockers, but conflict classification is not yet semantic or symbol-aware.
  • Semantic task planning and automatic path-claim proposal are not implemented.
  • Real network LLM providers are not configured by default. maco agent run currently supports only the deterministic local fake provider.
  • Repo-level validation commands run in the primary worktree after agent commands complete; use per-agent validation for checks that need to see unmerged agent worktree changes.

Release Operations

This memo does not tag a commit, push a branch, publish a crate, create a GitHub release, or publish any release artifact. Those operations remain separate release-manager actions.