Relaunch live agents when their inference credential rotates

[TheGreatAxios]

Fixes CL-6687 — https://linear.app/abklabs/issue/CL-6687

Problem

Inference sources (decrypted provider key included) are resolved at deploy time and rendered into a run's bytes; nothing re-reads them while the run is resident. After an Anthropic 401, pasting a new key in Settings returned 200 from /connections/anthropic/complete, yet the sidecar kept logging API key is invalid [HTTP 401] credential_failure for the already-open workbench. The send-time drift check (reconcileDriftedRun, CL-6588) compares only foldedBody content, which never carries the secret, so a rotation was invisible to it. Folded runs are self-anchored, so vendor's credential-push skips them too.

Change

• @corbits/folded-runs: extract resolveLaunchSources from deployAtHead; add inferenceSourcesDigest (SHA-256 over the resolved chain, secret included; only the hash leaves). deployAtHead / launchFoldedRun / wakeFoldedRun return the digest.
• @corbits/chat: nullable workbench_launch.sources_digest (migration 0024), recorded on every wake and relaunch. reconcileDriftedRun also relaunches when the stored digest differs from today's resolution; the per-send check is throttled to one resolution per participant per 30s. A row with no digest yet gets today's chain recorded as its baseline. New HubChatPlatform.reconcileInferenceSources(tenantId).
• Standalone launches record the digest too: @corbits/webhook-triggers gains a recordLaunchSources port; the hub's routine launcher calls recordSourcesDigest after launchFoldedRun (their mapping row is written before the deploy resolves the chain).
• apps/hub: the existing onConnected hook kicks reconcileInferenceSources (fire-and-forget, reportError on failure) when an inference provider's credential lands, so the operator's fix reaches the open room without waiting for a message.

Re-saving the same key yields the same digest → no relaunch.

Tests

• packages/folded-runs/src/launch.test.ts: digest is key-order-stable, secret-sensitive, default-source-sensitive, never contains the secret.
• packages/chat/test/platform-adapter.test.ts: rotated key → relaunch on the new key; same key → left alone; pre-column row → baseline recorded, no relaunch; second send inside the throttle window resolves nothing; reconcileInferenceSources sweep relaunches.
• packages/webhook-triggers/test/launch.test.ts, apps/hub/src/routine-launcher.test.ts: the digest is recorded after a standalone launch.

Review

Critique pass (sonnet) found: routine/webhook launches never recorded a digest (fixed above), per-send resolution cost (throttled), duplicated isInferenceProvider (now imported from @workbench/connections).

[TheGreatAxios]

critique · request-changes

Credential rotation relaunches drifted runs, but a connect-path relaunch is still swallowed for 30s.

• packages/chat/src/platform-adapter.ts:533-556 — hasDriftedSources returns false when now - checkedAt < 30_000. A provider connect that lands inside that window never reaches reconcileDriftedRun.
• packages/chat/test/platform-adapter.test.ts — no test drives connect-then-send inside 30s and expects relaunch. Rotate-path tests pass because they send after the cache expires or on a fresh adapter.

CL-6687's connect-path outcome is not met until that cache is gone or invalidated on connect.

GitHub will not accept request-changes on the author's own PR; this comment is the review.

[TheGreatAxios]

critique · approve

Connect-then-send inside 30s is closed. reconcileInferenceSources now deletes sourcesCheckedAt per binding before hasDriftedSources, and the new test stamps the interval then connect+send.

• packages/chat/src/platform-adapter.ts:744 — stamp delete before reconcileDriftedRun
• packages/chat/test/platform-adapter.test.ts:3306 — connecting a provider then sending within the check interval relaunches onto the new key

Should-fix (not blocking): stamp delete is per-iteration after listLaunchesForTenant, so a concurrent send to a later agent in the same tenant can still race. Typical one-agent paste-then-send is closed.

GitHub will not accept approve on the author's own PR; this comment is the review.