Órrery v1.2.0

Every model named by its exact version, work sent back to the plan when a premise proves false, and spend counted once per model.

• Every model is named by its version on every surface. The configuration page offers each exact version under the CLI's own name ("Opus 5.5") and says what a floating alias runs now ("Opus (latest) · Opus 5.5"); an alias the CLI lists no row for reads as the version a delegated run last reported, dated, or "not yet observed". Two model ids are compared by identity, never by prefix.
• New models appear on the configuration page as soon as the installed CLI lists them, and the page now names a newer copy of a CLI installed elsewhere, such as an editor extension that updates itself, with the command that brings Orrery's level.
• Premise-first handoffs: every task handed to a worker names the premises its plan rests on, each with the check that would show it false. A worker that finds one false stops and reports it, and the principal verifies the report and revises the plan instead of sending the same task out again. The flowchart draws that loop.
• Spend is measured across the whole account, delegated and nested runs included, and one Claude model is one key whatever its spelling (dated, Bedrock, Vertex, [1m]); the stored rollup moves to version 3 in place, keeping its history. The doctor no longer counts routine spend records as incidents.
• A failure is classified from the provider's own error, not from what a delegate happened to read.
• The standing-approval and pickup stores open their files without following symlinks. An insecure standing store is skipped with a warning instead of stopping a launch, and revocation reports every store it could not clear.
• The no-delegation guard is documented as best-effort rather than a boundary, the hook residual is restated against current CLIs, and the README and technical overview say plainly that no speed, cost or quality claim is made.
• Validated against Claude Code 2.1.283 and Codex 0.157.0; the doctor warns on newer CLIs until they are re-validated. 554 deterministic tests spending no model credits, from 478.

MIT licence. Provider products and model names remain the property of their respective owners.

Órrery v1.1.0

Containment of what the principal itself spends, and configuration that no longer lives in a tracked file.

• A delegate is never substituted onto the provider its principal spends from, and any substitution now needs a verified failure on record; a per-provider spend ceiling refuses dispatches and warns the principal's own turns, measured from local session logs including cache reads.
• The shipped default and the machine's configuration are separate: global/orchestration.json stays as shipped, your choices live in ~/.config/orrery/config.json, and orrery-config --import moves a configured manifest's choices across.
• Every session opens with the configured roles, provider, model and thinking level, two to a line, on every surface including the VS Code extension.
• Parked pickup: quota-stopped work is parked and re-dispatched when the provider's limit resets. Stall detection stops a delegate that is looping rather than merely slow. Containment claims are enforced and their limits named.
• The model catalogue keeps itself current per client version and says when it cannot; delegated Claude runs stream and survive Leave No Trace.
• The configuration page's Apply is discoverable and sits beside the legend it acts on, with a coloured preview diff; the README lists every command.
• Measured, not assumed: changing the thinking level inside a session does not rebuild the prompt cache, and a principal pinned at max has no in-session lever, which the policy now states.
• tests/run-pinned.sh verifies against the committed default on a configured machine. 478 deterministic tests spending no model credits, from 357.

MIT licence. Provider products and model names remain the property of their respective owners.

Órrery v1.0.0

First stable release. The feature roadmap is complete.

• Provider-neutral orchestration over Claude Code and the Codex CLI: one principal, four delegated roles, on any mix of Anthropic, OpenAI, or third-party and local endpoints.
• Task control plane: sealed contracts, isolated worktrees, parallel dispatch with admission control, and merges gated on recomputed evidence, re-running the union of acceptance checks at integration.
• Review pipeline: schema-validated findings with evidence, carried findings that outlive a reviewer, and a ranked queue that explains what awaits a decision and verifies every artefact it cites.
• Containment measured, not assumed: kernel-enforced read-only delegate runs probed before each use, with honest degradation and a named escape where a host cannot enforce.
• Consent-gated fallback with standing approvals that disclose themselves; token and cost accounting with ceilings; a governed memory whose every fact carries the command that re-checks it; a local incident log.
• 357 deterministic tests spending no model credits; lint and suite on CI for every push, plus a weekly drift run.

MIT licence. Provider products and model names remain the property of their respective owners.