Lightbridge Governance — VS Code extension

Contributes governed language models to VS Code chat through the Lightbridge AI gateway, using VS Code's LanguageModelChatProvider API (finalized in 1.104).

Status: verified against the live gateway. All 18 production models map, with the catalogue fetched and a real credential resolved through governance-auth token. A chat turn has been streamed end to end in a real editor. Not yet published to the Marketplace.

Why this exists

docs/matrix-row.md is the RFC-0003 §9 source declaration and is the thing to read first. The short version: serving the model rather than observing it makes attribution come from the credential instead of from a payload field that carries no identity.

Design decisions worth knowing before editing

No credential of its own. src/auth.ts shells out to governance-auth token. This extension implements no OAuth, stores no refresh token and has no keychain. governance-auth already does authorization_code + PKCE, holds the token at 0600, locks against concurrent refresh and never emits a stale credential (ADR-0010, ADR-0012).

Fail closed, everywhere. Every path out of getToken is a token or a throw. provideLanguageModelChatInformation contributes zero models when there is no gateway URL, no credential in silent mode, or no reachable catalogue — with no stale-catalogue fallback, because serving a cached list after the gateway stops answering is how a model that policy withdrew stays selectable.

Capability comes from the gateway, never from here. src/catalogue.ts maps /v1/models/info. A model whose context window the gateway does not report is skipped, not defaulted. If entries come back but none map, that is logged as an error naming the likely cause — a bare "0 models" is the hardest possible symptom to debug, and it is what a schema drift produces.

The context window is not passed through verbatim. VS Code renders maxInputTokens + maxOutputTokens as the total window, while the catalogue's context_length is that total. Passing both through advertised 264k for a 200k model. The input budget therefore excludes the output reserve, so the displayed sum equals the real window.

modelOptions is filtered against the catalogue's supported_parameters. Copilot Chat populates modelOptions with its own internal fields (_capturingTokenCorrelationId, _otelTraceContext, _telemetryTurn, _enableThinking). Forwarding those relayed Copilot's telemetry identifiers to our gateway. An empty allowlist forwards nothing rather than everything.

Nothing logs a body. src/redact.ts is deliberately free of any vscode import so redaction is testable with plain node --test; a control exercisable only by launching an editor stops being exercised.

Commands

just ext-install     # npm ci, inside ide/vscode only
just ext-check       # typecheck + unit + integration
just ext-build       # bundle to dist/
just ext-package     # .vsix

just all-checks does not run these — the extension has its own path-filtered workflow, vscode-extension.yml. See AGENTS.md for the rule that fence enforces.

Testing

src/test/ — unit tests for SSE framing and redaction. No vscode import, so they run under plain node --test.
tests/ — integration scenarios driving the real provider against an in-process fake gateway, with vscode aliased to a stub at bundle time. Covers catalogue mapping, streaming, tool-call reassembly, the modelOptions allowlist and every fail-closed branch.

Two tests exist because the obvious version of them proved nothing, and both are worth understanding before editing:

The SSE tests were rewritten after the first four passed with a deliberate \n\n → \n framing bug injected — their payloads happened to contain no internal newline. Two of them now pin the terminator specifically.
The fail-closed scenarios point at a permissive control probe, not the strict gateway. Against the strict gateway they passed even with a fail-closed bypass injected into auth.ts, because the gateway's own 401 produced exactly the empty list and the throw they asserted on. They were measuring the gateway's strictness, not the extension's refusal. The suite also asserts the probe was never contacted at all.

If you change either, re-inject the corresponding bug and confirm the test fails for the reason you predict.

Known limits

Inline completions are unreachable. BYOK covers chat, agent mode and utility tasks; ghost-text completions stay on Copilot's infrastructure (microsoft/vscode#318545). No accept/reject, edit-distance or time-to-decision signal.
A Copilot Business/Enterprise admin can disable the BYOK policy, which makes this provider unselectable regardless of what we ship.
Images and prompt-tsx parts are dropped in src/messages.ts rather than half-encoded, even though the catalogue can advertise imageInput.
provideTokenCount is a documented estimate, not a tokenizer. It over-counts on purpose; the two errors are not symmetric. It is never a billing input — cost is the gateway's to compute in integer µUSD (ADR-0008), which is also why pricing from the catalogue is deliberately not read.

Lightbridge Governance