Repository navigation
CONTRACT: feat(ui-server): add GPT-Live asynchronous client delegation - #2070
Conversation
schlessera
added
type: feat
e525be3 to
5d1d977
Compare
|
Additional reproduction from voice stack #2135, without another main refresh:
This matches #2141's native merge-chain shape. Metadata/immutable-object evidence is recorded here; no independent per-job identity reconciliation or replacement validation policy is claimed. The tested heads remain intact. No unrelated-main rebase or aggregate bypass is being used. |
5d1d977 to
da85ded
Compare
|
Gemini PR #2063 is merged as Current-head confirmation 38933150766 is pending with its first job waiting for a hosted runner. Contract run 38933095529 is queued. The earlier upper-layer confirmation's fourteen selected jobs passed, but its native parent-identity aggregate failed under #2141; that historical run is not substituted for this new candidate. The PR remains ready, milestone 0.44.0, awaiting current hosted qualification before the next dependency-order native merge. |
da85ded to
1a6cb2c
Compare
|
Continuing delivery on 2026-10-17. Prior confirmation 38933150766 completed with unit failures and a stale-base aggregate; it is no longer described as queued or passing. The failed profile/model, cancellation, bridge snapshot and admission expectations already have repairs on main (#2183, #2193, #2199 and #2251). I reproduced the four stale profile assertions before integration, then rebased the six GPT commits onto fixed base Local preflight now passes completely: 541 fast invariant tests, then 25 import-graph feedback tests. The targeted eleven-file regression run passes 149 cases, including both async adapters, registry/model expectations, cancellation, runtime records, admission, snapshots and isolated unit-worker controls. Two interactive policy-probe cases failed because the fixture selected a host-home Python unavailable inside the worker; unchanged main reproduces both. Selecting system Python through a command-scoped PATH passes the entire 19-case policy boundary suite / 587 assertions with all namespaces, mounts, nonempty attack checks and offline guards intact. The fixture portability issue is filed separately as #2304; no product guard or test assertion was changed here. Fresh confirmation 39023395726 passed metadata/lint/retention and has selected the full categories. Its metadata job 117369650156 actually checked out UI #2142 and permission #2145 have also been rebased in their own layers and pushed, resolving the API-report and documentation conflicts while preserving main's Mermaid signature and current contracts. The top layer again passes 39 pinned real-browser scenarios and both native approval-recovery cases, whose setup completes the current package build. Those layers remain draft for dependency-order delivery; no native aggregate waiver is implied. All existing release milestones remain 0.44.0. |
|
Current voice adapter confirmation run 39023395726 at head Unit (3), job Unit (4), job Resume #2070 by incorporating the delivered restart repair and #2293 after it lands, verifying both named failures and the voice contracts, then requesting current selected confirmation. Keep the diagram proof correction with its existing owner; no gate weakening or unchanged failing-job rerun is proposed. |
1a6cb2c to
6c06f6b
Compare
|
Pushed the prepared voice stack in one atomic update on fixed base
Range-diff verifies every feature commit unchanged. Delivered fixture repairs #2305 and #2317 are incorporated. Current GPT local preflight passes 541 invariant tests plus 25 affected tests; both loopback adapters pass 42 tests / 123 assertions, including async host work with continuing chat progress and interactive questions. The broader scope command's unexplained SIGKILL is recorded under #2074; it is not a full-suite pass. Historical full confirmation 39023395726 finished with thirteen selected jobs passing, the two diagnosed main fixture failures, and a stale-base aggregate failure. All fifteen actual selected checkout logs independently agree on The deterministic diagram binding now has a smaller repair, #2322 / #2314. It remains draft behind existing discovery failure #2153; no need to wait for the larger #2293 retention change. These three voice PRs are draft, with milestones 0.44.0, to avoid another full run against the known-broken checker. Current recovery setup hits its existing 180-second package-build deadline before assertions; #1600 now contains named phase evidence (ui-react begins at 173.8 seconds). The queued browser attempt then skips all 39 cases because that interrupted build left styles.css absent. Neither is passing coverage. A separate full Resume once the reviewed source-binding repair is delivered: incorporate that specific repair, verify the named binding/restart cases and current async behavior, obtain current selected hosted proof, then merge the eligible native prefix in dependency order. Keep actual tested identities; unrelated main advancement alone does not require refreshing. No guard, timeout or CI-policy relaxation is included. |
Use the primary Live v1 connection and server Bearer auth, preserving host-owned commit, backend execution and interactive chat progress. Correlate original transcript fragments by echoed audio command IDs and reject uncorrelatable multi-utterance input; no complete-turn or per-output cancellation signal is fabricated. Proof: 19 real loopback adapter/host tests and 49 assertions pass. Epoch and delayed-input identity mutations fail their intended rejection/misattribution assertions, then pass after restoration. Focused preflight passed 448 tests / 2793 assertions; strict types passed before final resync fixture addition. No paid or live provider calls.
…phone input Ignore zero-duration appended text and require command correlation after a historical resync. Removing the zero-duration guard makes the named history test fail by adding Historical context to the fresh input; restored suite passes 20 cases and 50 assertions. This tightens the original-input and resync criteria of #959.
…ities Apply the same explicit audio identity validation to delegation metadata as to transcript deltas. An unknown supplied command id must not fall back to implicit host binding. The 21-case real loopback suite passes 51 assertions.
…nstalled types Use the installed OpenAI Live event shapes for nonempty session.started, input transcript and client delegation fixtures. The session expiry uses the pinned fictional fixture date. The 21-case loopback suite remains green.
…redentials Exercise native Bun redirect refusal against two loopback servers, preserve complete WebSocket descriptors, and make history/foreign-identity assertions wait for their actual event evidence. The focused suite passes 22 tests and 54 assertions.
…ission The provider must retain historical command identities without treating settled utterances as live host slots. Use the existing bounded history capacity; the host still enforces active utterance limits. A 17-request loopback regression first failed at the expected resolves assertion because input 17 rejected with identity capacity exceeded; after the fix the suite passes 23 cases and 72 assertions. Audited other historical/pending ledger limits in the adapter.
6c06f6b to
3dc08df
Compare
|
Delivery resumed after #2326 landed the diagram binding repair. Current pushed heads: #2070 Current focused adapters/binding proof passes 52 cases / 172 assertions; the named restart case also passes. GPT complete local preflight passes 542 invariant tests plus 25 affected feedback tests. Top-layer actual Chrome recovery passes both cases / 16 assertions (163.71 seconds), and all 39 pinned Chromium conversation cases pass (28.93 seconds). Delivered forward-closure builds complete under the unchanged setup budget; no guard or timeout change is included here. Upper-layer local preflights remain in progress, with UI fast invariants already passing 724 cases. GPT is ready and confirmation 39032664018 is running all fourteen selected product categories. Contract and metadata pass. Actual checkout |
|
All fourteen selected product jobs in 39032664018 pass. Actual checkout Main additions were assessed: #1870 changes HTTP fixtures/documentation; #2335 changes restricted relay IPC validation. The composed current-main candidate uses base The additional native qualification first reproduces the existing #2315 SDK whole-profile parent timeout: 24 identity plus 13 egress controls pass before its unchanged 25-second caller exits143. This is not a passing full native receipt. A separately labelled diagnostic changes only that caller to #2315's selected four-turn composed allowance; all existing inner turn/cancellation deadlines and every assertion remain. It completes both SDK profiles, all 28 restricted-relay checks, actual managed probes, prepared native consumers and cleanup-failure controls, with zero owned resources and complete loop/mount cleanup. The fixture is then restored byte-for-byte and the combined tree is clean. This is controlled affected-runtime evidence, not clean-source whole-suite qualification or delivery of #2315. Following the maintainer instruction to push tested work and avoid repeated full refreshes for main advancement, retain these actual identities and use the normal native prefix merge with final head/main/milestone readback. No old passing run is relabelled as the newer combined tree; the freshness failure and diagnostic limitation remain explicit. |
1 participant
Closes #959
What changed
Adds
createGptLiveConversationProvider(apiKey)for explicit registration beside Gemini in this voice stack. The primary Live API v1 WebSocket uses server Bearer authentication, gpt-live-1, PCM16/24 kHz, marin, client delegation and store:false. Native delegation remains metadata; the host owns semantic commit, its independently selected backend, tools, permissions and ordinary chat progress/questions. Later input and output continue while backend work genuinely awaits.Correlation and lifecycle
Audio commands retain their own event identities. Delayed transcript/delegation events use echoed client_event_id; ambiguous or unknown explicit correlation closes visibly. Historical zero-duration text never becomes microphone input. Host results retain native delegation_id, use bounded quiet thinking chunks and at most one commentary append, and reject stale/consumed/replaced/aborted handles. Native redirects are refused before Bearer credentials can be forwarded. session.closed acknowledges whole-session close only; no output-done or cancellation acknowledgement is invented.
Proof
Current head
3dc08df98e155d08b9feda38ceb05ce68e82415f, repaired main base7cc86931e763f0eca636b89abd78b3b5371ec027, tree040dcebe25ead4aecb7cc22f32a99e3119185e7a:Local preflight passes: 542 invariant tests / 3635 assertions, then 25 affected feedback tests / 105 assertions.
Both async adapters and diagram input binding pass 52 focused cases / 172 assertions. GPT alone has 23 cases / 72 assertions; both adapters have 42 / 123. The real host holds work pending while later speech, chat progress/tool activity and an ordinary interactive question continue. The named real-host restart regression also passes (1 case / 7 assertions) with the delivered fixture repair.
Hosted confirmation 39032664018: all fourteen product jobs pass, including 4,706 unit cases with no failures/skips. Every selected job and private unit-worker checkout matches
f9889eb78e00c38f1f94b2013242176d079adb72. The final aggregate rejects freshness after main advancement; the whole workflow is failed and is not relabelled green. Assessed main integration and separate affected-runtime/diagnostic receipts are recorded in the merge comment.Retained failing-first receipt: after completing 16 inputs, a 17th was incorrectly rejected by the host active-utterance limit. Historical identities are now bounded separately (4096), preserving the host's 16 active slots; the named regression and later input pass.
Retained mutations: removing the epoch check makes stale publication resolve; borrowing the latest input identity mislabels a delayed utterance; removing the history guard turns replay text into microphone input. Each failed its named assertion and was restored. These feature commits compare unchanged after base integration.
No live provider session or paid experiment ran.
Stack and release
Merged normally as
18a65206d4ed5f6fc4790728293ac2ce72858026, parentbc9ee2b25eb0687fa66b01536a6e15aae9c010cd, treeb671d5deeceb2b7b82418599aab567c72147808a, matching the computed merge tree. Ancestry on main and #959 closure are verified. Automatic main-push CI39034501043 is separate and remains pending. Gemini #2063 is also merged; UI #2142 and permission #2145 remain above this layer in native stack #2135. The diff adds only the GPT-Live slice and its tests/docs/minor ui-server changeset. Hosted proof retains its actual parent/head identity; unrelated later main advancement is assessed without repeatedly refreshing. Milestone 0.44.0 remains assigned to the PR and #959, with final readback and actual squash reconciliation required at merge.Checklist
Anything left open
#961 owns separately authorized live handshake, device/audio, latency and safety qualification. Transcript deltas lack a finalization event and origin/certainty remain unknown, so the client must require manual review. Exact permission speech is unsupported; echo isolation and word alignment remain unproven. store:false does not certify zero retention. Client capture/playback and permission-specific behavior are separate stack layers.