Changelog

All notable changes to this project are documented here. The format follows Keep a Changelog, and the project adheres to Semantic Versioning.

[Unreleased]

0.3.0 — 2026-06-19

Two surfaces for the validated generator: an editor integration and a public, comparable scoreboard. Both build on the existing engine — generation and the static lint gate are unchanged, and the base install stays GMAT-free.

Added

• VS Code extension — generation commands inside the editor, driving the engine through a thin stdio JSON-RPC worker launched in your own Python environment. A draft is shown as a reviewable diff and applied only on accept (never auto-applied); lint and dry-run findings surface as inline diagnostics, and the provider/model is chosen explicitly from a reachable-providers quick-pick. Every .script language feature (highlighting, hover, formatting) stays with the gmat-script extension, which it depends on. Published to the VS Code Marketplace and Open VSX in lockstep with the package (#72, #73).
• Per-model leaderboard — a harness that scores each seeded provider:model over the evaluation suite and writes a ranked leaderboard.json. It ranks on a never-committed held-out set as the headline, with the committed public set alongside as the offline-reproducible anchor, so a model that overfits the public prompts gains no rank; the board carries aggregates only, and a verify step is the firewall that keeps held-out golds off it. Hosted as a static Hugging Face Space — the project's one hosted artifact — rebuilt by a gated CI job and a refresh workflow (#74, #75).

0.2.0 — 2026-06-18

Closes the loop from a natural-language intent to a script validated against a real GMAT. Generation and the static lint gate stay GMAT-free; the new dynamic validation, the repair loop, and provenance are additive — the dry-run behind an optional extra.

Added

• Dynamic GMAT dry-run tier — behind the optional [gmat] extra, a lint-clean draft is loaded in a real GMAT (and run when a Target/Optimize solver is present) to catch the runtime errors a static parse cannot. Strictly additive: the strict/permissive lint contract is unchanged and the base install stays GMAT-free (#58).
• Bounded repair loop — draft() can feed a failing draft's lint (and, with the dry-run, runtime) diagnostics back to the model and regenerate, lint-first. Opt-in; it stops at the first runnable draft, on budget exhaustion, or when a regenerated draft stops changing (#59).
• Provenance — every result carries a versioned record of how it was produced (the request, the model, the retrieval trace, the per-attempt draft history, and the outcome), always populated in memory and optionally serialised to a credential-free .copilot.json sidecar next to a saved script (#60).
• CLI --dry-run / --repair N / --provenance — drive the dynamic tier, the repair loop, and the sidecar from the command line; the summary line reports the lint, dry-run, and retry outcome (#61).
• Eval close-the-loop tier — the evaluation suite gains a dry-run-agreement tier and measures the repair loop's lift, with a workflow_dispatch input to select the live suite in CI (#62, #65).

0.1.0 — 2026-06-17

Initial public release. gmat-copilot turns a natural-language request into a GMAT mission .script, retrieval-grounded in the GMAT documentation and validated against a static linter, through a model you choose. Generation and validation need no GMAT install.

Added

• draft() and CopilotResult — the public entry point. It retrieves grounding context, builds the prompt, calls the chosen provider, extracts the fenced .script from the response, lint-checks it, and returns a CopilotResult carrying the script, the lint report, and the retrieval trace. CopilotResult.save() writes the script to disk; a strict gate (the default) raises DraftRejected when the draft does not lint clean (#24).
• Model-agnostic provider abstraction — adapters for Anthropic (Claude), OpenAI, Ollama, and GitHub Models, selected explicitly as provider:model. There is no default model and no silent fallback: with none chosen, the tool lists the providers reachable from your configured credentials rather than picking one. A recorded test provider keeps generation tests offline and deterministic (#25).
• Retrieval-grounded context — a build tool ingests GMAT help pages, sample scripts, GmatFunctions, and a curated set of domain notes into a corpus with a prebuilt FAISS index that ships inside the wheel; a query-time retriever assembles a token-budgeted grounding context for each request, so the model writes against real syntax rather than from memory (#26, #27).
• Static lint gate — every draft is checked by the gmat-script linter before it is returned. Strict mode rejects a script that reports any error or warning; permissive mode returns the best-effort script with every diagnostic attached (#28, #30).
• CLI (gmat-copilot) — generate a .script straight from an intent, write it to -o (or stdout) and print a concise lint summary; gmat-copilot validate <file> lints an existing script. Strict mode exits non-zero on a non-clean draft; --permissive writes it anyway. API keys are read from the environment, never committed (#32).
• Two-layer evaluation suite — a ~50-prompt set with an LLM-as-judge and a scorer, runnable reproducibly and wired into CI behind a gated, budget-aware job (#31).
• Documentation — an MkDocs-Material site (getting started, the provider and auth model, the validation contract, the result schema, the evaluation protocol, the corpus and its licences, worked examples, an API reference, and the design decisions), auto-deployed to GitHub Pages on tag pushes (#23, #54).
• Packaging — a GMAT-free base install with [anthropic], [openai], [ollama], and [gmat] extras; CITATION.cff metadata; and a release workflow that builds, publishes to PyPI via trusted publishing, and creates the GitHub release on v* tags (#23).