[HellHBBD]

Live setup can complete without model output. Use the normal opening generation as one bounded readiness attempt, preserve its events for delivery, and start the interview clock only after the first model output. Startup failures do not retry.

Carry outcomes on the existing interviewer_state message and degraded-start banner. Name provider-specific errors only when the provider identifies them.

Validation

• ./scripts/test.sh passed after rebasing onto 546f559, with Chromium, commentflow, and actionlint enabled. Startup tests cover setup without output, timeout, a single generation attempt, event preservation, and timer ordering.
• The offline Rust-server interview check passed with the existing web/audio/test-voice.wav capture fixture.
• The contributor ran release build and tests successfully at df170a8; the release binary reports that revision.

Browser tests used a temporary Liberation Sans alias; the default-font lobby failure also reproduces on base a3f3e80. Live provider integration was not run.

Closes #159

---

Summary by cubic

Live setup can finish without model output, so the interview clock can run against a silent provider. This change starts the interview timer and hard deadline only after the first model output—nonempty audio, model text, output transcription, or a tool call—using the opening greeting as a single bounded readiness attempt that never retries. A failed check ends the interview; practice mode that continues stays untimed, and Practice End during the startup wait closes without requesting a report.

Failure reporting

• Readiness outcomes publish on the existing interviewer_state control topic with startup: "ready" or startup: "failed" plus a classified reason (quota, billing, credentials, refused, or unavailable), shown in the degraded-start banner.
• An exhausted key rotation or a cooldown that cannot fit the first-open budget keeps its quota or billing cause instead of degrading into a generic transport timeout.
• The page sends a bounded startup_received receipt for failed readiness, hashed and published under the same bound as report receipts, and keeps the diagnosis through room teardown.

Socket and client behavior

• Socket setup keeps its bounded key-failover and restart budget; only readiness generation is a single attempt.
• Keepalives keep a responsive socket alive while waiting, and a GoAway advisory sent before output replays with its original timestamp and no further generation.
• Queued opening events are preserved in order for the normal room loop, and the reply-wait clock starts at readiness so replay cannot expire an already answered greeting.
• Observed usage is accounted once even when readiness publication fails; strict mode stays gated until readiness.
• Editable work before the clock starts and after a failed opening warns on unload until its report persists; early interviewer departure, recording delays, and premature report requests are handled without blocking cleanup or losing context.

^{Written for commit 3fe5839. Summary will update on new commits.}

[ColtenOuO]

Changes look good to me, before merge this PR, please squash your commit.

[HellHBBD]

The browser CI failure also occurs on upstream main at fd7c23c, before the Gemini changes in this PR:
tests/browser/lobby.test.js:6457
problem cards span desktop panels and return home on narrow screens

AssertionError: 57 !== 4

The same test passed at the preceding main commit, 0116b9b, then failed after merging #401 at fd7c23c. This PR's browser job reports the same assertion failure.

Code inspection points to an interaction between #410's layout test and #401's new .problem-entry wrapper. The test temporarily changes .problem-card.hidden and later clears it, but #401's filtering also hides the enclosing .problem-entry. Clearing the button's hidden property therefore does not necessarily make it visible. The subsequent .problem-card:not([hidden]) query still includes buttons inside hidden wrappers; their zero bounding rectangles cause them to be counted as sharing one row.

The layout test needs to account for the wrapper when changing visibility and count cards that are actually rendered. This analysis is based on the CI logs and source comparison; I have not rerun a corrected test yet.

[jserv]

After a failed startup check, connect() returns null, so init() stops at if (!startup) return; and never reaches updateBeforeUnloadGuard(). Meanwhile the banner tells the candidate "You can still work the problem offline." shouldWarnBeforeUnload only checks endsAt, so a candidate who keeps coding offline can close or reload the tab and lose everything without a prompt. Before this PR, a degraded start always set endsAt and turned the guard on. Turn the guard on for this path without starting the clock, for example with a flag passed to shouldWarnBeforeUnload. The same early return also hits any non-startup failure after state.joinedRoom is set (for example, publishIntegrityEvent throwing). If that is intended, the comment should say so, because it currently mentions only readiness.

[jserv]

Outside Strict mode, nodes.end and nodes.pause stay enabled during this wait, which now lasts the whole server-side startup check. Pressing End here sends end_interview to an agent that only defers it until readiness, so when the check then fails the candidate is left on the ending overlay waiting for a report the agent will never send, and when it succeeds they get a report for an interview whose clock never started. Keep both buttons disabled until readiness settles, the way init() already does for Strict, or have endInterview skip the report wait while state.interviewerReadiness is still pending.

[jserv]

isolate_local_agent removes duplicate agent participants during startup. That runs before this agent publishes its audio track, so the browser has usually not adopted an agent yet. When the duplicate's ParticipantDisconnected lands, !state.agentIdentity && isAgent(participant) holds, so fail() settles readiness as failed. The candidate then gets the "did not answer the startup check" banner just before the real interviewer would have reported ready. Fail only when no interviewer is left in the room.

Suggested change

	if (
	participant?.identity === state.agentIdentity ||
	(!state.agentIdentity && isAgent(participant))
	)
	if (
	isAgent(participant) &&
	!roomInterviewer(roomParticipants(), state.agentIdentity)
	)

[jserv]

Before asking someone to review, you MUST resolve any regressions flagged by GitHub Actions.

[jserv]

RECONNECT_NOTICE_LIMIT is one second, sized for a reconnect notice that retry_cold_open logs and drops on timeout. Here the same bound makes the ready notice fatal: when publish_data takes longer than a second on a congested data channel, an interview whose greeting Gemini already generated (and billed) is torn down, and the failure notice that follows carries unavailable, which the page words as "Gemini did not answer the startup check". Give the ready notice its own bound that fits inside the browser's readiness wait, and do not report a publication failure as a provider failure.

[jserv]

The handler now calls roomInterviewer and roomParticipants, but this context still passes only state, isAgent and updateAgentState, so this test and the next one throw ReferenceError: roomInterviewer is not defined in the browser gate. Pass roomInterviewer and a roomParticipants stub that returns who is still in the room, and rewrite the cases for the new rule: an agent's departure fails readiness only when no interviewer remains, so another agent leaving while the pinned one stays settles nothing, and the pinned one leaving fails it only when no other agent is left.

[jserv]

init now leaves End enabled outside Strict mode during the readiness wait, so this helper fails both practice-mode tests that call it (false !== true in the browser gate), and its click on End would start the ending flow rather than do nothing. Assert End disabled and inert only when the page is in Strict mode, and for practice mode assert what pressing End before readiness is meant to do, so the test pins the behavior the open thread on End during the startup wait settles on.

[jserv]

Rebase latest main branch and resolve conflicts.

[jserv]

Keep End available during startup in practice mode. If readiness is pending, End should leave without entering report generation or waiting for a report. Keep End disabled during startup in Strict mode. Rebase onto main, which has removed Pause, and align the implementation, tests and documentation with this behavior.

[HellHBBD]

Got it. I've identified a few additional issues during testing and working on addressing them. I'll need a bit more time before pushing the updated changes.

[jserv]

Rebase the current branch onto the upstream default branch and resolve any conflicts. Run /codetrial-review in AI agents (use $codetrial-review in Codex) and resolve all identified issues. Rework the series into minimal, logically coherent commits, squash closely related commits, and ensure all commit messages follow the project’s rules.