Changelog

Unreleased

• Made Main→Terra and Terra→Luna/Reviewer waits event-driven with an explicit 60-minute timeout that returns early on mailbox updates.
• Prohibited short polling and routine list_agents calls during healthy Runs.
• Added staged timeout diagnostics: lightweight Maestra status for Main, Task-local evidence for Terra, and a single path-scoped list_agents call only when liveness remains unclear.
• Added regression coverage for both orchestration layers.

0.3.2 — 2026-08-15

Rollout-backed Doctor verification release.

• Added deterministic verify-routing-proof support for Codex rollout metadata.
• Bound each route to the recorded parent/child thread chain, exact spawn_agent arguments, and observed child turn_context.model / turn_context.effort.
• Distinguished affirmative routing mismatches (fail) from unavailable or ambiguous metadata (unverified).
• Kept prompts, messages, summaries, and reasoning out of routing evidence.
• Added regression coverage for successful routing, observed model mismatch, and missing rollout metadata.

0.3.1 — 2026-08-15

Explicit opt-in routing release.

• Added $maestra:using-maestra as the only full-workflow entry.
• Set all seven Maestra Skills to policy.allow_implicit_invocation: false.
• Kept Spec, Plan, Run, Gate, and Finish independent; Main now loads each exact sibling SKILL.md after explicit activation instead of requiring a user command at every phase.
• Made Doctor and standalone phase recovery workflow-neutral.
• Added no hooks, session-wide injection, activation database, or custom router runtime.
• Preserved v0.3 context isolation, exact routes, Finding Freeze, bounded review, candidate-tree/commit binding, Run Integration Review, Main Gate, and risk-based Final Review.
• Expanded layout, metadata, hook-absence, validator, and both-ZIP packaging coverage.

0.3.0 — 2026-08-15

Planning ownership release.

• Main Sol now owns the full detailed implementation Plan, including architecture, Run/Task decomposition, target components, implementation steps, design decisions, verification strategy, and review focus.
• Plan approval is the natural context boundary; /compact is recommended immediately afterward, and execution re-reads approved Spec/Plan artifacts instead of relying on the compact summary.
• RunPacket schema v2 requires detailed Task fields and rejects unresolved open_questions before Terra is spawned.
• Added fail-closed PLAN_GAP: Terra must stop instead of inventing missing architecture, implementation strategy, migration, or test strategy. Main resolves the gap through REPLAN or USER_DECISION.
• Terra is now explicitly execution-only: sequencing, exact command resolution, verification, reviewer lifecycle/adjudication, Git commits, integration review, and RunReport.
• Runtime artifacts moved out of the product tree to git rev-parse --git-path maestra, including linked-worktree support.
• Routing proof schema v2 binds successful Doctor results to both Codex version and Maestra version.
• Run completion now mechanically requires passing Run-level deterministic verification and fresh Sol Integration Review bound to the same candidate tree.
• Normal $maestra:run flows directly into the Main semantic gate; $maestra:gate remains available for recovery/explicit invocation.
• Whole-branch final Sol review is risk-based rather than mandatory for every low-risk single-Run change.
• Retained Terra review adjudication, Finding Freeze, maximum two Task review rounds, and one-Task-one-commit candidate-tree protection.
• Added/updated regression coverage for detailed Plan validation, PLAN_GAP, integration gating, version-bound routing proof, and linked worktrees.