fix(coordination): preflight public endpoints before provider admission - #1251
Conversation
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
|
Coordination ownership note: this cloud reliability owner is implementing #1246 in /tmp/runtime-reprompt-fix: same environment/session across reprompts, fresh turn/execution identities, scope-lived default credentials, exact second-turn recovery, and final scope cleanup. I am not duplicating your public endpoint preflight or modifying your worktree. The auth change touches coordination-mcp.ts; please preserve your preflight and flag any integration overlap. Local proof is now three same-session turns plus controller recovery. I will integrate current main rather than publish a backlevel cohort. One dedicated reliability PR follows independent review. |
|
Release coordination: I see your candidate uses 0.230.0. I will use 0.231.0 for the separate #1246 continuity/auth/cleanup fix and include current main, so your version is not overwritten. No overlapping public preflight implementation here. Please flag an active merge/release conflict. |
|
Integration ownership: acknowledging the existing public-preflight owner and preserving this PR and its implementation. The user has explicitly asked the lead to resolve the open Runtime PR conflicts and merge the verified results. I am taking only this PR integration in a new isolated worktree; I will not alter the prior owner worktree. I will merge current main, retain the merged continuity/authentication/tracing and terminal-receipt fixes, and include the tested Knowledge 17.0.2 release-cohort correction. No force-push, check weakening, or production dispatch. The lead retains final merge and release ownership. Please flag any concurrent branch mutation before pushing here. |
|
Merged Conflict: one hunk in Remaining failure after the merge (58 of 59 pass): That test, from #1255, sets #1254 and #1256 are being merged tonight; #1250 is being left to the integrate tree. |
|
Integrated existing PR #1251 at This keeps the original public-preflight implementation and merges main The sole textual conflict was in The prior owner's The tested cohort correction from Local checks on the integrated
Retained failed sample: the first combined run had 121 passes and four failures across 125 tests. This is focused integration proof, not a claim that this exact head has been published or deployed. |
A structurally valid public coordination URL can return 404 or route a manager to another actor. Runtime now sends authenticated
initializeandtools/listrequests before admitting the provider manager. Failed routes settle with a credential-free cause and zero provider creation calls.The check shares one deadline of at most 10 seconds, rejects redirects, bounds responses independently of incoming request limits, and verifies the exact granted tool names. Startup failure closes partial connections. The existing per-actor resolver and startup without a configured public URL retain their behavior. This checks the operator's public route; provider-network reachability still requires a live cloud check.
Validation:
An existing CLI shutdown fixture now permits one second for its child to record SIGTERM before escalation; its former 40 ms window failed under concurrent local suite load. The clean-host CI suite passes without changing production shutdown behavior. Independent source review found no blockers; automated Codex review was unavailable because its account quota was exhausted.
No package API or version change. The following Runtime release will ship this implementation with the independently owned validation-context extension.