Orchestra 0.8.0

Follow-ups from the first real end-to-end runs, plus a resume path for runs held for approval. Everything new is opt-in or backward compatible.

• orchestra resume <runId> finishes a run that orchestra approve finalised with work still owed, on the same run id and integration branch. It runs the tasks that were blocked behind the approved task, then the deferred mutation pass and learning. Resume inherits the run's recorded worker isolation and refuses weaker settings unless --override-isolation is given. A lock stops two resumes of the same run (#81).
• Footprint scheduling: shared append-only files (throughput.sharedFiles, default ["CHANGELOG.md"]) no longer count as overlap, and they merge with git's union driver (#83).
• Budgets: each worker attempt logs the exact budget line its prompt carried (#84).
• Sandbox: opt-in sandbox.allowLocalBinding. Pre-flight checks argv pass-through for srt and docker (#77, #78).
• Fixed: runner accounting for sessions with more than one result message, the integrity monitor false positive on test fakes, and docker pre-flight tests that depended on writable /var/tmp (#78, #82).
• Bench: optional h2-none and h2-agentic arms, and the first H1/H2 findings in docs/status.md (#80).

The Claude Code sandbox backend has now run real workers. srt, docker and sandbox-exec still have not. Full details are in CHANGELOG.md.

Orchestra 0.7.0

Every milestone in the Closing the Gaps plan (M1 to M6) plus the Repo Atlas. All new capabilities are opt-in and off by default, so a default run behaves as in 0.6.1.

• M2 Measurement: orchestra bench with paired statistics, ablation arms and judge calibration.
• M3 Evolution: time-split eval sets with a locked holdout, orchestra evals add|harvest|list|audit, cascade screening, sequential testing, a cost objective and a Pareto archive.
• M4 Verification: judge panel, agentic judge, machine-readable contracts, oracles and a mutation score.
• M5 Isolation: an OS sandbox for workers (--sandbox with srt, the Claude Code sandbox, docker or sandbox-exec), env scrubbing and pre-flight probes.
• M6 Throughput and oversight: footprint scheduling, a batch merge queue with bisect, risk tiers with orchestra approve, budgets, a planner cascade, best-of-n and oversight metrics.
• Repo Atlas (W9): a repo map and orchestra atlas find|neighbors|impact|tests.
• Fixed: the srt backend now passes the worker command after -- (#74).
• Docs: whitepaper and blog updated (#75).

None of these has yet been exercised against real sandbox backends, and hypotheses H1 to H8 have not been measured. See docs/status.md. Full details are in CHANGELOG.md.

Orchestra 0.6.1

Docs: adds the Orchestra whitepaper (docs/whitepaper-orchestra.md) and a Substack-style blog post (docs/blog/orchestra-substack-post.md), linked from the README. No code changes.

Orchestra 0.6.0

Multi-candidate evolve (#8, evolve.candidates), wider evolvable harness surface (#9, tools.md, style.md, checklist.md), and opt-in contrastive repair without tests (#11, contrastive config). See CHANGELOG.md.

Orchestra 0.5.0

Per-domain learning libraries (#10): domain-tagged runs, learnings/.json plus _global.json, evidence on every lesson, per-domain caps, and orchestra learnings --domain. See CHANGELOG.md.

Orchestra 0.4.0

Adds orchestra serve --tui (alias orchestra tui): interactive full-screen terminal dashboard, no browser or HTTP port needed. See CHANGELOG.md.