providers.md

Provider architecture

For maintainers. Using T3 Code? See docs/user.

A provider is the agent runtime that does the actual work. T3 Code supports several, and the orchestration layer does not know which one is behind a thread.

Built-in drivers

builtInDrivers.ts exports BUILT_IN_DRIVERS with five entries:

Driver kind	Driver source
codex	Drivers/CodexDriver.ts
claudeAgent	Drivers/ClaudeDriver.ts
cursor	Drivers/CursorDriver.ts
grok	Drivers/GrokDriver.ts
opencode	Drivers/OpenCodeDriver.ts

Each driver declares its driverKind, a configSchema, and a create function that builds an adapter in a child scope. Adapter implementations live beside them in apps/server/src/provider/Layers/ (CodexAdapter.ts, ClaudeAdapter.ts, and so on) and conform to ProviderAdapter.ts. Read the driver plus its adapter to see how a specific agent's transport, config, and event shapes are mapped.

Registry and routing

Two registries separate configuration from live processes:

• ProviderInstanceRegistry keys configured instances by ProviderInstanceId. Creating one looks up the driver by driverKind, decodes entry.config with that driver's schema, opens a child scope, and calls driver.create.
• ProviderAdapterRegistry resolves an instance ID to its live adapter via getByInstance.

ProviderService sits on top. It combines the adapter registry with the provider session directory to route session and turn operations for a thread, so callers name a thread, not an agent.

Adding a driver means writing the driver plus adapter and adding it to BUILT_IN_DRIVERS. No orchestration, contract, or client change is required for the common case.

Model manifest

The model picker's legacy section is driven by apps/server/src/provider/model-manifest.json, which lists the current (non-legacy) model slugs per driver kind. The ModelManifest service (apps/server/src/provider/ModelManifest.ts) refreshes that data from the same file on main via raw.githubusercontent.com, so moving a model in or out of the legacy section is a commit, not a release. Preference order is remote fetch, then the on-disk copy of the last successful fetch (in the state directory), then the bundled copy. Fetches are TTL-gated, run concurrently with provider probes, respect the enableProviderUpdateChecks setting, and never fail a provider check. The Codex and Claude drivers apply the classification to every snapshot with applyModelManifest; driver kinds absent from the manifest have no legacy concept.

Secret references in provider environments

A provider instance's environment is merged into the child process env by mergeProviderInstanceEnvironment, once per driver inside create. A value that starts with op:// is not passed through: it is a secret reference, and ProviderSecretResolver swaps it for the value the 1Password CLI returns before the merge happens.

The parsing half lives in ProviderSecretReference.ts and knows nothing about how a secret is fetched, so the registry can ask "does this instance read from a secret store?" without depending on the resolver. ProviderSecretResolverLive is the half that shells out to op read --no-newline.

Three decisions are load-bearing:

• A failed read unsets the variable. It never substitutes an empty string, because an empty ANTHROPIC_API_KEY reads to the provider as a configured-but-broken credential rather than an absent one, and the status badge would go back to lying about it. Unsetting is stronger than leaving the name out of the resolved list: the child environment starts from the server's own, so a name left alone keeps whatever the server inherited under it, and the agent would quietly run as a different account than the one the instance names.
• Reads are batched across instances, and sequential within one. The store charges an unlock per op invocation, not per secret, and every instance resolves its own environment inside create, so a fleet would otherwise cost one prompt per provider. prime reads the whole set in a single op inject before the builds start, called from the settings watcher (which covers boot) and from reloadSecretBackedInstances (which covers the refresh button). Whatever prime misses, resolve still walks with a plain loop rather than Effect.forEach with concurrency, so it produces one prompt rather than several simultaneous ones.
• Priming can only ever help. It is best effort and never fails: op inject resolves the whole template or fails it, so one bad reference would take the batch down with it. A batch that does not come back leaves the cache exactly as cold as it found it and resolve falls back to reading one reference at a time, which is both where the per-variable failure isolation lives and how the user learns which reference is the broken one. A separator is generated per call because a secret can contain anything, newlines included.
• Failures are cached alongside successes. The Cache holds Exits, so a locked vault costs one prompt per refresh cycle instead of one per thread start. Recovery is the refresh button, not a timeout: the cache is built without a timeToLive.

Why a refresh has to rebuild the instance

A driver resolves its environment once, at create time, and makeManagedServerProvider re-probes using that captured processEnv. Dropping the cached secret therefore changes nothing on its own, because the running instance still holds the value it was built with.

So ProviderRegistry.reloadSecretBackedInstances invalidates the cache and then calls rebuildInstanceWhen on each instance, passing hasProviderSecretReference as the predicate. The registry owns the child scopes and already stores each instance's config, so it can close and rebuild one entry in place; passing a predicate rather than exposing the entries keeps secret-store policy in ProviderRegistry and keeps the instance registry ignorant of 1Password. reconcile cannot do this job, because it diffs the config envelope and a rotated secret leaves that envelope byte-identical.

A rebuild takes as long as the secret read does, and a locked vault can park it on a person at a biometric prompt. Three things follow from that window being long:

• The instance leaves the map before its scope closes. Otherwise every lookup for the whole window hands back a bundle whose scope is already gone. Its last snapshot stands in for it meanwhile, because ProviderRegistry prunes ids it finds in neither list, and the card must not blink out of Settings while 1Password waits on a fingerprint.
• Rebuilds and reconcile take turns. Both read the instance map, do slow work, then write it back, so a settings change landing mid-rebuild would be overwritten by the map the rebuild read before it. Reads stay outside the lock: getInstance and listInstances never wait on a 1Password prompt.
• A rebuild that fails is retryable. The registry keeps the config envelope of an instance it could not bring back, so the next refresh retries it, and a refresh with no explicit target covers the unavailable instances as well as the live ones. Without both halves, a vault that happened to be locked would cost the user the instance until settings changed. The envelope is recorded before the build starts and the map writes around the build are uninterruptible, so a refresh whose caller walked away mid-read is retryable on the same terms as one that failed.

This hangs off the three refresh entry points (refreshAll, the kind-scoped refresh, and refreshInstance), all of which are reached only by a user action: the Settings refresh button, and the post-update verification in providerMaintenanceRunner. The periodic provider health loop is not one of them. It lives inside makeManagedServerProvider and calls refreshSnapshot directly, which is what keeps a resolved secret alive between refreshes instead of re-reading it every few minutes.

Claude credential liveness

The Claude capability probe reports which credential the CLI found, which is a different question from whether Anthropic still honours it. A revoked or expired setup token still reports tokenSource: "CLAUDE_CODE_OAUTH_TOKEN", so Settings kept showing a green badge until the first turn failed.

ClaudeCredential.ts closes that gap with one authenticated GET /v1/models, chosen because it is the cheapest first-party endpoint that accepts the token and costs no message quota. The status check runs it from the same path that already spawns the probe, so it inherits the provider health cadence and needs no cache of its own.

Four decisions carry the behavior:

• Only a value shaped like an Anthropic credential is ever sent: the sk-ant- prefix followed by key material and nothing else. A placeholder, or a secret reference nothing resolved, earns a 401 that says nothing about the token, and acting on it would blame the credential for a problem one layer up. The tail is checked because a placeholder can wear the prefix (sk-ant-oat01-${MY_TOKEN}), and rejecting a real token by mistake only costs an unknown.
• Only an explicit 401 counts. Timeouts, transport errors, 403, and every 5xx answer unknown, because a proxy or an outage must never sign a working install out of Settings.
• The check only ever downgrades. It runs after the probe has already concluded the instance is authenticated, so it can turn a green badge red but never the reverse.
• It is gated on the CLI reporting that it took its token from the environment. An install that authenticates through a keychain login, an API key, a router, or a cloud backend is never judged by a variable it ignores, even when that variable happens to be set.

Note that Bearer authentication with a Claude Code OAuth token is not part of Anthropic's public API surface. It is verified working, not contractually stable, which is the other reason every unexpected answer is treated as unknown.

How provider work is requested

Clients never call a provider directly. They dispatch orchestration commands over the RPC method orchestration.dispatchCommand, defined with the rest of the orchestration surface in orchestration.ts. The client-dispatchable provider-facing commands are thread.turn.start, thread.turn.interrupt, thread.approval.respond, thread.user-input.respond, thread.checkpoint.revert, and thread.session.stop, plus the mode setters thread.runtime-mode.set and thread.interaction-mode.set.

The engine persists an event for the command, and a server-side reactor performs the provider call. Provider output comes back as internal commands such as thread.message.assistant.delta and thread.session.set, which clients observe through orchestration.subscribeThread. See overview.md for the command/event loop.

Server-side workers

Provider work flows through three queue-backed workers. All three are built with makeDrainableWorker from DrainableWorker.ts and expose drain for deterministic test synchronization.

1. ProviderRuntimeIngestion consumes provider runtime streams and emits orchestration commands.
2. ProviderCommandReactor reacts to orchestration intent events and dispatches provider calls.
3. CheckpointReactor captures workspace checkpoints on turn start and completion, and performs reverts.

Buffered assistant delivery

A thread in buffered assistant delivery mode accumulates assistant text instead of streaming each delta. The buffer is not held until turn completion. In ProviderRuntimeIngestion, MAX_BUFFERED_ASSISTANT_CHARS is 24,000: the append that would exceed it invalidates the buffer and spills the whole accumulated text as one delta. The buffer also flushes at interaction boundaries, when a request opens (approval) or user input is requested, via flushBufferedAssistantMessagesForTurn.

Stalled prompt detection

session/prompt is a long-lived request: ACP agents answer it only once the whole turn is done, and nothing in the protocol says how long that takes. An agent whose upstream connection dies mid-turn never answers and never errors, so AcpSessionRuntime races the RPC against a liveness watchdog and fails the turn instead of waiting forever.

Liveness is traffic for this session plus outstanding work, and both halves are load-bearing.

Scoping matters because one runtime projects one root session: a child session chattering on the same pipe says nothing about whether the root prompt is alive, so only updates that pass the root-session check refresh the stamp. Counting outstanding requests matters because silence is often our fault, not the agent's. Every request the agent makes of us is held open while it runs, including the extension requests, which is the case worth stating: cursor/ask_question and x.ai/ask_user_question park on a human, and a user who takes fifteen minutes to answer must not look like a dead agent. The same holds for a twenty-minute terminal/wait_for_exit.

A stall therefore needs no root-session traffic and nothing of ours outstanding, for promptStallTimeout (ten minutes by default).

On a stall the runtime sends session/cancel so the agent can release the dead prompt and stay usable, then fails with an AcpTransportError. ProviderCommandReactor turns that into a thread session error with a provider.turn.start.failed activity and clears activeTurnId, so the working indicator stops and the reason is visible in the timeline.