Commit Graph
118 Commits
Author SHA1 Message Date
jif f8e58064fc Preserve remote compaction summaries when interrupting child agents (#61916)
## Why

Child agents configured to interrupt at item boundaries should retain the summary when interrupted during remote compaction. An interrupt during post-turn compaction should also preserve the answer already completed.

## What changed

- Keep sampling interrupt registration active through summary installation, then honor the pending interrupt without running post-compaction hooks.
- Deliver an already completed answer when post-turn compaction is interrupted.
- Keep queued follow-ups for a new turn instead of continuing the interrupted turn.

## Testing

Add a streaming scenario test and context snapshot covering interrupts during mid-turn and post-turn remote compaction. Verify that resumed requests retain the summaries, the completed answer is delivered, and each queued follow-up starts a separate turn.

GitOrigin-RevId: 939eb8b3ed24dbdbff10c9f62226e38a0198fd6f
2026-11-16 23:58:12 +00:00
tongzhou wang eaa25e1f43 Refactor rollout budget checks and turn completion state (#61583)
## What changed

- Centralize generation admission checks before request contributors run for HTTP and non-warmup WebSocket requests.
- Route pending input and inline compaction through dedicated budget checks, and centralize post-compaction budget handling.
- Preserve whether any checkpoint was installed when remote compaction combines a background candidate with foreground compaction.
- Track accepted final answers separately from trailing commentary and encapsulate the response state shared by stop and after-agent hooks.

## Testing

Update the plan-mode contributor test to verify that contributed commentary is tracked separately from final answers, and adapt stream-event assertions to the phase-aware assistant text representation.

GitOrigin-RevId: 43b663e92bfd24ec7e5a0778b711938f16a7fea4
2026-11-15 00:03:59 +00:00
tongzhou wang ce82e38b16 Gate new rollout work on shared budget exhaustion (#61582)
## Why

A budget can be exhausted while another thread waits to start work. New requests, tools, and automatic turns must observe that stop without discarding an already-consumed response whose cost exhausted the budget.

## What changed

- Share a work gate across the rollout tree while preserving native and host accounting ownership.
- Recheck the budget before HTTP and WebSocket requests, hook execution, tool dispatch, and compaction. Prevent automatic wakeups and late input from restarting stopped work, leaving pending input queued.
- Record response costs without rejecting the consumed response solely because its bill exhausts the budget, and stop new work when enforced accounting overflows.
- Expose a thread-registration receipt that remains committed after unload.

## Testing

Extend rollout-budget coverage to verify that accepted response text remains in history after exhaustion and that waiting tools observe the shared stop while completed tools retain their results.

GitOrigin-RevId: e583d657c6f29e374d374cdb014bf23fd6c28614
2026-11-15 00:03:23 +00:00
tongzhou wang 2e61e8ec4e Record neutral interruption markers for rollout budget stops (#61581)
## Why

Turns stopped by an exhausted rollout budget need a history marker without attributing the interruption to the user.

## What changed

- Record and flush a `<turn_aborted>` marker before the terminal event for `SessionBudgetExceeded`, respecting the interrupt-message setting.
- Use neutral guidance in a `developer` message for automatic budget stops.
- Centralize tool-start budget checks across parallel dispatch, tool handlers, and MCP handoff, retaining the check after tool-start observers finish waiting.
- Share interruption classification and history recording across turn and compaction paths.

## Testing

Extend the rollout-budget integration test to assert that both the budget-crossing turn and a later denied turn persist `developer` interruption markers.

GitOrigin-RevId: 2a49ac9824a4e95e4799f2d0f638f8c46889e8e8
2026-11-15 00:02:01 +00:00
tongzhou wang c8c3f70893 Account for synchronous MCP post-hook costs in rollout budgets (#61580)
## What changed

Allow successful synchronous `PostModelRequest` MCP hooks to return `extra_codex_rollout_budget_units` in `structuredContent`. Add these charges to rollout spending while preserving provider usage separately, including completed responses without token usage and incomplete responses with reported usage.

Validate prices as finite and nonnegative. Preserve valid sibling charges when another hook reports an invalid price, and surface accounting failures without retrying inference. Carry hook charges and overflow information through host accounting and enforce local budget limits.

## Testing

Extend coverage for hook-driven budget exhaustion, failed hooks, overflow, interrupted responses, and foreground and background compaction accounting.

GitOrigin-RevId: dc54ba53478d82848bcafb55ecef88215019d09f
2026-11-14 23:53:53 +00:00
tongzhou wang f226c9ccfb Recheck rollout budgets before starting tool work (#61578)
## Why

A rollout budget can be exhausted while a tool call waits for an execution slot, approval, or a lifecycle callback. Checking only when scheduling the call can allow new work to start after the budget is exhausted.

## What changed

- Recheck the budget after acquiring an execution slot, immediately before starting a handler, and immediately before handing an MCP request to the server.
- Report refused calls as aborted in tool outputs, lifecycle callbacks, and dispatch traces, while allowing already-running work to complete.
- Track handler start time separately from dispatch waiting and preserve terminal outcomes when cancellation arrives during completion callbacks.

## Testing

Add regression coverage for queued calls and calls paused in start callbacks when the budget is exhausted. Verify MCP requests are blocked before handoff and retain successful results after handoff, and assert budget-stop state during remote compaction.

GitOrigin-RevId: 573730391452256592d3512f15929a2fbdf03c57
2026-11-14 23:52:50 +00:00
tongzhou wang 92c1a4dc2e Add response processing after completion and usage accounting (#61577)
## Why

Streaming can dispatch tools before a response finishes and its usage is accounted for. Allow callers to wait for completion and synchronous post-hooks so budget enforcement can block tools from the response that exhausts the budget.

## What changed

- Add `responses.process_mode = "after_complete_response"` to buffer response items and release them in order after completion, synchronous post-hooks, and usage accounting. Keep `streaming` as the default and deliver safety-buffering notifications immediately.
- Track whether response usage has already been charged to avoid duplicate accounting, including during compaction.
- Stop new tools across the agent tree when the rollout budget is exhausted. Coordinate steering with accounting and prevent buffered tool calls from dispatching after steering or cancellation.
- Bound buffered responses to 65,535 events and reject overflow.

## Testing

Add configuration coverage and extend integration tests for buffered output ordering, synchronous hooks, interrupted responses and resume, steering, invalid usage, and budget exhaustion preventing tool dispatch.

GitOrigin-RevId: bd06811cfd50cb701babe6353465c57dfed2a37a
2026-11-14 23:50:12 +00:00
tongzhou wang 81b995a6c9 Unify model request hook lifecycles across sampling and compaction (#61562)
## What changed

- Pair `PreModelRequest` and `PostModelRequest` hooks per sampling or compaction attempt, with distinct lifecycle IDs and cancellation reporting.
- Add `thread_id` and `kind` to hook inputs, distinguishing `sampling`, `foreground_compaction`, and `background_compaction`. Include the provider's `response_id` in post-hook inputs when available.
- Wait for synchronous post-hooks before releasing terminal response events, accounting for usage, or installing compaction results. Allow streaming output and tool execution to proceed before the hook finishes, and cancel synchronous post-hooks when their owning request is interrupted.
- Preserve incomplete-response, budget-exceeded, and hook-denied outcomes across sampling retries and compaction model fallback.

## Testing

Add coverage for streaming and tool ordering, interrupted post-hooks, per-attempt metadata, and foreground and background compaction accounting and installation ordering.

GitOrigin-RevId: 3e808c5b3bf2d5a6a326d9f5be5bbd0359b2d16e
2026-11-14 19:40:12 +00:00
Vivian Fang c33002259e Add per-item retention policies for injected history (#61344)
## What changed

Allow `thread/inject_items` to accept wrapped items with `retention: "default"` or `retention: "always"`, while preserving support for raw response items. Persist explicit policies separately from legacy client-authored metadata.

Carry explicitly retained items through local and remote compaction, context-window changes, and child-agent creation, subject to retention budgets. Preserve history order and keep retained tool exchanges together; require matching calls and outputs to be adjacent in the injection request and both marked `always`.

Update Quinn dependencies and record exceptions for additional Hickory DNSSEC advisories where DNSSEC is disabled.

## Testing

Update injection tests for explicit retention metadata and extend the compaction scenario to cover a retained tool exchange across repeated compaction and restart.

GitOrigin-RevId: 1195a4ed62dc162a62e58bc2ad733262fbabea27
2026-11-13 18:46:13 +00:00
cassirer-openai b8dde1eb40 Move rollout trace recording to a background worker (#61253)
## Why

Recording inference starts synchronously serializes prompts and writes trace payloads before model work can proceed. Defer that work so inference dispatch can continue while the trace is recorded.

## What changed

- Use one worker per trace bundle to write events in submission order, with up to eight pending jobs.
- Share prepared prompts and encoded requests with the worker, and build remote compaction snapshots there. Write payloads before publishing events that reference them.
- Add `TraceWriter::flush` and `ThreadTraceContext::flush` for live reads. Drain accepted jobs and seal the bundle on shutdown; suppress the closing marker after a recording failure.

## Testing

Add tests for deferred prompt serialization, ordered start and completion events, shutdown draining, and serialization failures that preserve earlier evidence without publishing a closing marker. Verify compaction snapshots are built on the worker and skipped when tracing is disabled.

GitOrigin-RevId: 32fd787bb1d281f1814df2f08c73b667628f8add
2026-11-13 04:36:03 +00:00
Alex Kira 64b8d019c8 Record reasoning replay removal metrics (#61017)
## What changed

Parse reasoning removals from response `input_transformations` and carry the counts through completion events to sampling telemetry.

- Add `codex.responses.reasoning_replay` to distinguish unavailable reports, no reported removals, prefix mismatches, and other removals.
- Add `codex.responses.reasoning_items_removed` histograms by bounded removal reason, grouping unknown or missing reasons under `other`.
- Include zero counts for usable reports; exclude missing, null, or malformed reports from histograms without failing the response.
- Use the sampling step's session telemetry when recording response events.

## Testing

Add SSE and WebSocket integration coverage for report availability, removal counts, unknown reasons, zero observations, and `product_sku` tags in exported metrics.

GitOrigin-RevId: 83c70ca6067fa4b34bf5c463958b4632a7a0a188
2026-11-12 04:31:25 +00:00
cooper-oai faeb379646 Bind runtime auth to the provider's selected credential source (#59823)
## Why

Runtime authentication must use the credential selected by the model provider. Explicit API keys and command credentials need independent ownership so cached connections and in-flight work cannot cross credential sources.

## What changed

- Register and reuse runtimes per credential source, supporting explicit provider keys and command credentials without falling back to saved login credentials.
- Track `AuthOwner` by credential manager and owner generation across WebSocket state, compaction, and scoring.
- Preserve cached identity and runtime registration state during credential-source outages or failed storage reloads. Require a successful auth reload before logout revokes credentials.
- Reject requests whose production or staging environment differs from runtime registration, including model catalog requests.

## Testing

Add regression coverage for credential selection and runtime reuse, owner isolation, source outages, failed reloads and logout recovery, and environment mismatches.

GitOrigin-RevId: 7bab55e0ea502593393df10046309fd8502edbd0
2026-11-06 04:33:14 +00:00
jif 48524f337a Preserve pending agent messages across compaction (#59404)
## Why

Agent messages that arrive after a model request starts must survive compaction until a normal inference request includes them. Compactor output alone does not establish that those messages were delivered.

## What changed

- Persist history sequences and sampling checkpoints in harness metadata, capturing the input prefix for each normal inference request.
- Retain pending agent messages within existing size budgets during local and remote compaction, placing the summary before pending messages and subsequent caller input.
- Preserve sampling progress across history replacement and resume, and drain queued mail after interrupted sampling even when compaction leaves context above the token limit.

## Testing

Add regression coverage for request-boundary sampling checkpoints, pending messages across local compaction and repeated remote compaction with resume, and injected messages that interrupt sampling before mid-turn compaction.

GitOrigin-RevId: 26c04ab96b5a1c3cc1d6625ccc74a0f00201d14c
2026-11-04 13:03:01 +00:00
Ruslan Nigmatullin d84c872464 Route response item IDs through a shared generator (#59312)
## What changed

Introduce `ItemIdGenerator` and pass the runtime-owned generator through sessions, model clients, history, hooks, tool outputs, and memory requests. Keep UUIDv7 generation as the default and replace direct calls to `ResponseItemId::new` with the shared callable.

Preserve existing item IDs and backfill missing or empty IDs through the same generator, including forked history before reconstruction and persistence.

## Testing

Extend coverage for UUIDv7 IDs on user messages, agent messages, and additional tools, and assert that forked model context matches the backfilled items persisted to history.

GitOrigin-RevId: 0dd9ccf0f46f0f1005a6acfc740133b03ce82259
2026-11-04 04:02:00 +00:00
Hung 1685f2d621 Use measured durations to budget uploaded audio (#59268)
## Why

Uploaded audio always reserved 12,000 tokens per file, even for short clips. This could discard recordings during compaction that would fit within the retained-message budget.

## What changed

- Add optional audio duration metadata to attachments and cache measured durations per session, resolving metadata for new input and replayed history.
- Use measured durations for history estimates and compaction budgets at 10 tokens per second, rounded up and capped at 12,000 tokens. Keep the full budget for missing or unusable metadata.
- Bound metadata lookups to 250 milliseconds per batch with up to eight concurrent requests, and record accepted input before lookup so interruption preserves the batch.

## Testing

Add coverage for duration rounding and fallback estimates, mixed measured and unknown audio, retention through compaction and cold resume, and interruption during metadata lookup.

GitOrigin-RevId: 26fe320c7024944d9d4c371d5fdcda84442b5cc5
2026-11-04 00:19:12 +00:00
Hung 3780a5f750 Budget retained file audio during remote compaction (#59237)
## Why

Remote compaction did not charge retained file audio against its token budget, allowing audio-heavy history to exceed the intended limit.

## What changed

- Charge file audio using `FILE_AUDIO_TOKENS_ESTIMATE`, independently of image budgeting.
- Retain the newest media content that fits, keeping images and audio together with their adjacent labels and preserving content annotations. Stop at oversized media instead of backfilling older content.
- Share media-label detection with user-input event mapping.
- Add download timeouts and a Debian mirror fallback for the musl build's `libcap` source archive.

## Testing

Add unit coverage for audio budgeting in both image-budget modes, atomic media labels, and truncation boundaries. Add integration coverage showing that compaction retains the newest five of seven file-audio items within the retained-message budget, including a mixed-media case.

GitOrigin-RevId: 4b7cd69f53966461fdeec14096712671d480187b
2026-11-03 21:53:26 +00:00
Hung f1b0721c88 Enable uploaded audio with durable rejection recovery (#59234)
## Why

Uploaded audio was rejected even for audio-capable models. Invalid or expired recordings also need to be excluded from later requests so users can continue the conversation.

## What changed

- Accept uploaded audio when the selected model supports audio; otherwise, preserve valid sibling input and report the unsupported attachment before inference.
- Track successful audio decoding by model and provider, and attribute provider rejections against the exact submitted request. Preserve attachments when attribution is ambiguous.
- Persist confirmed rejected file IDs across compaction and resume, omitting them from future requests without rewriting retained user history.
- Check audio availability before remote compaction and retry once without an unambiguously rejected recording. Include guidance to remove or replace affected attachments in audio errors.

## Testing

Add unit and scenario coverage for uploaded audio across follow-up turns, compaction, cold and paginated resume, model switching, ambiguous rejections, expired files, and queued or interrupted input.

GitOrigin-RevId: e701901e7b1c0865c9f2a5918197215f549b06c0
2026-11-03 21:24:16 +00:00
Charlie Marsh 4460449921 Keep HTTP fallback active for the rest of the session (#58989)
## What changed

Once Responses requests fall back to HTTP after exhausting WebSocket retries or receiving HTTP 426, keep subsequent requests and turns on HTTP for the remainder of the session. Clear cached WebSocket state and remove background recovery probes, recovery trials, and their acknowledgement timeout.

## Testing

Update fallback tests to verify that a second turn stays on HTTP without additional WebSocket attempts. Adjust retry and `Retry-After` request-count assertions to exclude background probes.

GitOrigin-RevId: a4f6376604527c14db8d60f79bd21c3aaf289f70
2026-11-02 21:37:17 +00:00
rhan-oai c725e472e5 Add usage accounting for host-owned agent runtimes (#58975)
## Why

Host-owned runtimes skip controller rollout-budget accounting. Give hosts an optional callback to account for inference spending and signal budget exhaustion.

## What changed

- Add `AgentUsageAccounting`, `AgentUsageReport`, and `UsageAccountingDecision`, captured at startup and supported by host child spawn reservations.
- Await per-response usage reports with thread and turn attribution, including provider budget units and foreground and background compaction usage.
- Map host exhaustion to `SessionBudgetExceeded` and accounting failures to fatal errors without retrying inference. Preserve the original error for incomplete responses.
- Keep controller accounting unchanged and leave accounting disabled for host-owned runtimes without a provider.

## Testing

Add coverage for usage attribution, child accounting, budget exhaustion, nonretryable accounting failures, incomplete responses, and compaction reporting, including background usage charged once to its originating turn.

GitOrigin-RevId: 887703f50fe882497e4e8cb33259072935ec82e0
2026-11-02 20:37:55 +00:00
Charlie Marsh 0a2dda4b5e Recover WebSocket transport after transient HTTP fallback (#58950)
## Why

A transient WebSocket failure previously kept the session on HTTP for its remaining lifetime, even if WebSocket connectivity recovered.

## What changed

- Probe WebSocket connectivity in the background during HTTP fallback, allowing a subsequent request to use a ready connection without delaying HTTP requests.
- Verify recovery with a real request and bound its first-response wait to five seconds. Keep HTTP 426 responses and recovery timeouts on HTTP for the rest of the session.
- Keep retries of an HTTP request on HTTP. If a recovered connection fails before output, replay the request over HTTP without consuming its ordinary retry budget; failures after output use normal retry accounting.
- Rate-limit recovery probes and honor `Retry-After` deadlines.

## Testing

Extend integration coverage for successful recovery within a turn, disconnects before and after output, full-context replay followed by incremental WebSocket requests, persistent HTTP 426 fallback, and retry deadlines.

GitOrigin-RevId: 1c139a4bbce1a67442db5622fd70ff0cc8056c9c
2026-11-02 19:02:50 +00:00
cassirer-openai 7656a23b8d Preserve incomplete response output in failed inference traces (#58456)
## Why

When a response fails with incomplete output, attempted tool calls can appear only in the terminal response payload. Failed inference traces previously omitted those items, losing evidence of the attempted call.

## What changed

Retain output items marked `incomplete` on `ResponseIncompleteError` and include recognized items in failed inference traces, deduplicating by item ID. Keep these items out of streamed output, tool dispatch, and the continuation cache, and omit their payloads from the error's debug representation.

## Testing

Add tests for retaining only incomplete output from terminal responses and recording attempted calls in failed traces without streaming them or creating tool calls. Cover unknown response item types being skipped.

GitOrigin-RevId: 594f6a37a5af543694f19fcffe0c164cb2addb36
2026-10-30 23:55:47 +00:00
jif e742db9f7b Preserve parent routing affinity for Guardian reviews (#57710)
## Why

Shared routing cookies can change after a parent connection is established. Guardian reviews need the routing affinity captured with their parent response so unrelated traffic does not replace that hint.

## What changed

- Capture `__oailb` and `__cflb` cookies from successful ChatGPT WebSocket handshakes and pair them with the parent response ID.
- Retain the snapshot throughout a review and its retries, and apply it only to Guardian requests with the matching parent response. Reconnect cached reviewer sockets when the parent affinity changes.
- Preserve cookie expiry and restrict reuse to the original secure origin and endpoint path. Replace routing cookies while retaining unrelated cookies, and keep the captured affinity out of prompts, history, and serialized metadata.

## Testing

Add coverage for cookie scope, expiry, duplicate rejection, and merging; replaying a captured pin after shared cookies change or are deleted; and retaining the matching parent snapshot through compaction without exposing cookies in review prompts or metadata.

GitOrigin-RevId: d180679e6265b612a3617f20ca1f5dec89e73aa8
2026-10-28 11:48:00 +00:00
rka-oai 9e03960cf8 Wire background compaction into the turn lifecycle (#53800)
## What changed

- Launch background compaction at eligible pre-turn, mid-turn, and post-turn checkpoints when `AsyncCompaction` is enabled, capturing the current world state and client routing state.
- Adopt ready results at compaction checkpoints while preserving newer messages, tool outputs, and world-state updates. Continue normally if an optional result is rejected.
- Run one foreground compaction if an adopted result still exceeds the context limit.
- Prevent cancelled turns from launching new background work while preserving already-running work for later reuse.

## Testing

Add unit and scenario coverage for stale history and account changes, cancellation, pending-result reuse and fallback, retained tool outputs, compaction event pairing, and token usage accounting. Add a context snapshot covering a collaboration-mode change while compaction runs.

GitOrigin-RevId: 04ac028d9fe2e3ea64aacaaeb351c3c776cf3940
2026-10-13 23:51:47 +00:00
jif 816fa702ec Account for token usage from incomplete responses (#53670)
## Why

Terminal `response.incomplete` errors discard reported token usage. Counting that usage as a new context baseline can also double-count input and trigger unnecessary compaction on the next turn.

## What changed

- Carry the response ID and valid reported usage with incomplete-response errors.
- Record usage for sampling and foreground compaction in cumulative totals, rollout records, budgets, and token-usage contributors without replacing the context baseline. Preserve the original terminal error, including when accounting exhausts the budget or cancellation races the error.
- Charge incomplete background compaction requests to the rollout budget without publishing foreground token usage.
- Ignore missing or malformed usage rather than estimating it.

## Testing

Extend coverage for SSE and WebSocket incomplete responses, rollout usage records, compaction accounting, deferred token-count events, and token-usage contributors. Add regressions for unnecessary compaction on the next turn and background compaction publishing foreground usage.

GitOrigin-RevId: bf1ee6df3912052ab7c19e1c6af8ad3dc4b07675
2026-10-13 13:36:24 +00:00
rka-oai 9167a0b2d4 Reuse background results for remote V2 compaction (#53620)
## What changed

- Await a compatible pending background request when foreground compaction is needed, and share the result installation path with foreground requests.
- Validate the captured history generation, prefix length, model, and authentication owner before installing a prepared result. Fall back to foreground compaction when required if the result is rejected.
- Replace only the captured history prefix, preserving later messages, retained tool outputs, goal edits, and the current world-state baseline. Record installation analytics and completion events only after a result is installed.

GitOrigin-RevId: 798963f2fc603c3febba2c21eafa1bcb094b3ebf
2026-10-13 09:18:45 +00:00
rka-oai 1a69ad3237 Add infrastructure for speculative remote V2 compaction (#53614)
## What changed

- Capture history, prompt metadata, routing state, authentication ownership, and context-window identity for background compaction requests.
- Send background requests over HTTP while preserving foreground routing, and retain the latest candidate in session memory.
- Add feature and threshold eligibility checks and deduplicate attempts by history version, model, compaction hash, and authentication ownership.
- Cancel and drain background tasks during session shutdown, preserving completed usage accounting.

The background launch helper is not yet connected to the turn loop.

GitOrigin-RevId: 0e5e6ab7dac4d98a836733668a775ac48b5a59f1
2026-10-13 08:44:23 +00:00
rka-oai be5296d467 Prepare compaction helpers for background requests (#53606)
## What changed

- Extract an HTTP Responses executor that can run without owning a turn's WebSocket state, with optional authentication-owner checks.
- Add cancellation-aware compaction output collection that records background token usage against the rollout budget.
- Support installing a compacted history prefix while preserving later history items and placing initial context around real user or agent messages.
- Allow callers to supply a context window for history trimming and a captured window for compaction request metadata.

GitOrigin-RevId: 29f3ca0ae8841f295058feb207e3a225af9a1487
2026-10-13 08:13:34 +00:00
Won Park 6c0c6759d0 Retain client-marked tool outputs across compaction (#52937)
## Why

Client-supplied tool outputs can carry task instructions. Even when marked with `retain: true`, these instructions could be trimmed before compaction and disappear from the compacted history.

## What changed

Add the disabled-by-default `retain_client_tool_outputs` feature to honor retention for explicitly marked, named `function_call_output` items without a `call_id`.

- Preserve eligible outputs in remote compaction v2 and when starting a new context window, keeping their original role.
- Stop emergency tool-output trimming at a retained output.
- Share the existing 64,000-token retention budget, keeping outputs whole and dropping an oversized output and older items rather than partially truncating it.
- Place initial context before retained standalone outputs and account for their images in retained-image counts.

## Testing

Add a scenario covering enabled and disabled retention, emergency trimming, oversized-output handling, and preservation of the latest instruction after compaction and cold resume. Verify that retention annotations stay out of model requests.

GitOrigin-RevId: 449179670e4d2be2623928700738758e43b5dc1a
2026-10-10 17:30:25 +00:00
pakrym-oai 5ef96ab278 Add opt-in output token replay for OpenAI requests (#52742)
## What changed

- Add the disabled-by-default `output_token_replay` feature to request `output.encrypted_content` from OpenAI providers.
- Preserve encrypted message and tool-call output, message and function-call status, and output-text `annotations` and `logprobs` through history and replay. Update the generated protocol schemas and SDK types.
- Strip output ciphertext from requests to other providers while retaining it in stored history, including after replay is disabled. Remove encrypted content from memory extraction inputs.

## Testing

Add integration coverage for the feature/provider combinations, metadata replay across follow-up requests, and history preservation when resuming with replay disabled.

GitOrigin-RevId: 77729a021b6771933402aebe74ece8eb39c057e2
2026-10-10 01:25:00 +00:00
pakrym-oai 062439b2f8 Record base instructions in incremental tool history (#51188)
## What changed

- Record nonempty base instructions as a developer message after tool declarations at the start of each incremental tool context window, and omit the request-level `instructions` field.
- Preserve this prefix when rebuilding history after local or remote compaction, while keeping the compaction summary last.
- Count recorded base instructions once when estimating context usage and trimming tool outputs.

## Testing

Add regression coverage for instruction placement, omission of request-level instructions, compaction ordering, and token accounting. Exercise remote compaction trimming with both request-level and recorded base instructions.

GitOrigin-RevId: 6a6fda92c5c6eb33852a77b38d34c028cf78b164
2026-10-05 22:20:38 +00:00
pakrym-oai 8571b9eaa4 Install full context in compaction replacement history (#51117)
## Why

Manual, pre-turn, and post-turn compaction previously deferred rebuilding initial context until the next turn. Their saved replacement history could omit refreshed instructions and skill catalogs.

## What changed

- Install initial context during local and remote compaction, before the last real user message or summary, keeping the compaction summary last.
- Persist matching full world-state and turn-context baselines with every replacement instead of retaining only partial extension metadata.
- Build replacement context and recompute token usage for the current model even when the previous model performs compaction. Reuse that captured context for eligible current-model retries.
- Capture manual compaction context centrally and propagate preparation errors to the task runner.

## Testing

Update assertions and snapshots for context ordering, refreshed instructions, resume and fork history, current-model checkpoint baselines, and cloud skill catalog retention across compaction and resume.

GitOrigin-RevId: 8572a76c31c6884a3c580b3bbb350d49af876bd1
2026-10-05 16:10:27 +00:00
felixxia-oai de02016798 Persist explicit user goal edits in model history (#49598)
## Why

User goal edits and clears need to reach the model as user instructions, including when a thread is unloaded or history is being compacted. Automatic goal lifecycle updates must not supply user authorization.

## What changed

- Add optional `origin` values of `user` and `automatic` to `thread/goal/set` and `thread/goal/clear`, and mark TUI goal requests as user actions. Omitted origins do not supply user authorization.
- Persist explicit user objective, status, and clear instructions before changing goal state. Recording failures leave goal state unchanged.
- Coordinate goal edits with thread resume, reload stored history under writer ownership, and preserve edits accepted after the compaction input snapshot.
- Update generated protocol schemas and TypeScript and Python bindings.

## Testing

Add coverage for user edits and clears on loaded and unloaded threads, concurrent path-based resume, exclusion of automatic goals from user instructions, preservation of edits during compaction, and recording failures that leave goal state unchanged.

GitOrigin-RevId: d779d7c678a0c32a6ef76975dbe5b1400d1df245
2026-09-30 11:58:03 +00:00
Won Park a9118edae8 Move content-filter guidance into the shared Responses retry handler (#49130)
## What changed

Move content-filter guidance recording from the sampling loop into
`handle_response_stream_error`. Pass `StepContext` so the handler can resolve
model-specific guidance and record it before retry decisions for sampling
requests. Keep remote compaction excluded from guidance recording and update
callers and the existing retry-deadline test to use the new signature.

GitOrigin-RevId: 030bde1d4f1597fc4f3971ad837478e5304c5aa9
2026-09-29 02:05:56 +00:00
pakrym-oai 12de0e395d Preserve WebSocket continuations when steering a turn (#48508)
## Why

Steering an active WebSocket response previously dropped the connection and
resent the full history. Draining the response preserves the connection and
allows the follow-up request to continue with `previous_response_id`.

## What changed

- Drain WebSocket responses when steering. For models using
  `use_responses_lite`, first send `response.interrupt` with
  `mode: "discard_partial_items"` once the response ID is available.
- Treat `response.incomplete` with reason `interrupted` as completion with
  `end_turn: false`, preserving token usage and allowing the turn to continue.
  Other incomplete reasons remain errors.

## Testing

Update the steering integration test to cover completed and discarded reasoning
items, asserting connection reuse, incremental follow-up input, and token usage
from the interrupted response.

GitOrigin-RevId: bc714b713e797cf1a67e3b26ffbf7664cb92eb33
2026-09-26 18:36:55 +00:00
felixxia-oai d5cbfe1455 Deduplicate retained instructions against Guardian transcripts (#48109)
## Why

Synchronous Guardian reviews can include the same user instruction in both the retained instructions section and the conversation transcript. Avoid this duplication while preserving complete authorization evidence and its original acceptance order.

## What changed

- Carry retained source metadata into transcript rendering. Omit a retained instruction when the transcript delivers its complete matching source revision.
- Label qualifying transcript entries with retained source order and protect them from budget truncation. Keep separate retained instructions when source metadata is missing or incomplete.
- Track delivery of source-order guidance so it remains available even when all retained instructions are delivered through the transcript. Invalidate that delivery proof when messages are shortened.

## Testing

Add coverage for complete-source deduplication, missing or incomplete metadata, budget protection, guidance redelivery, and source metadata preservation across resume. Update Guardian review assertions and snapshots to verify instructions are delivered once.

GitOrigin-RevId: 20ad0304501fa49f1596c4cfaf3ef89d64815753
2026-09-25 14:54:51 +00:00
felixxia-oai 86be5320b0 Deduplicate retained instructions across Guardian reviews (#48060)
## Why

Reused Guardian sessions already contain previously delivered instructions. Follow-up reviews can avoid repeating that evidence, but must deliver it again if it changes or compaction removes it.

## What changed

- Track retained source identity, revision, and completeness in host-owned history metadata, preserving revisions during replay.
- Omit retained instructions from synchronous review inputs only when the same complete revision remains in reviewer history.
- Restore missing evidence after reviewer compaction, enforcing the request budget before persisting additions. Classify restored evidence as context rather than new user authorization.
- Keep source-order labels stable across eviction and retain full delivery for legacy records without reliable provenance.

## Testing

Add coverage for consecutive-review deduplication, changed revisions, missing or incomplete metadata, and restoring compacted evidence exactly once. Extend compaction and resume coverage to verify retained revisions.

GitOrigin-RevId: e90fca16cde0a0970da36dd873af2ca1dbaaf4d3
2026-09-25 10:49:50 +00:00
felixxia-oai 920e1a34dc Make Guardian thread context capture unconditional (#47689)
## What changed

Always use `GuardianContextMode::ThreadOwned` for session capture and subagent history forks, regardless of `guardianv2.thread_context`. Continue selecting reviewer policy from saved checkpoints, preserving legacy review for incompatible checkpoints and removing the feature gate from reviewer compaction compatibility checks.

## Testing

Update coverage for unconditional retained context across compaction, resume, and subagent forks. Exercise legacy review with saved checkpoints, and verify that local and managed opt-outs do not disable sender context capture or checkpoint migration.

GitOrigin-RevId: cd74e2029c83f12bc655b6563242d06f29dd46af
2026-09-23 22:50:48 +00:00
jif 27969c0ae9 Route session agent operations through AgentControl (#47520)
## What changed

- Store `Arc<dyn AgentControl>` in session services and route agent tools, turn admission, usage accounting, completion notifications, and shared settings through the trait.
- Add `AgentControlInit` so delegated and guardian sessions inherit the parent's controller and local runtime.
- Move local registry and context helpers to `LocalAgentRuntime`, keep resume, close, inspection, and status subscriptions on the local controller, and simplify synchronous and infallible trait methods.

GitOrigin-RevId: ad3f213eb483ac524bde56d769f20a78f20989a0
2026-09-23 09:39:30 +00:00
jif c117207a6f Drop descendant channel posts from retained remote compaction v2 history (#47249)
## Why

Remote compaction v2 excluded descendant progress messages from retained history but kept `CHANNEL_POST` messages from those same agents.

## What changed

Treat descendant `CHANNEL_POST` messages like `MESSAGE` progress updates when selecting history to retain after remote compaction v2.

## Testing

Extend the remote compaction test to verify that a child channel post is included in the compaction input and omitted from the follow-up request.

GitOrigin-RevId: 204bf44b8cddf1e6dc1e48234f51d4ac2aaf2e9f
2026-09-22 10:44:00 +00:00
Henry Levy 49e248d4c3 Add opt-in compaction after final responses (#46541)
## What changed

Add `model_post_turn_compact_threshold_percent` to trigger compaction after a final response when context usage reaches the configured percentage of the usable window or an existing auto-compaction limit. Accept values from `0` to `100`; omitted or zero disables the feature.

Skip post-turn compaction when input is pending, the turn is cancelled, or token-budget mode is enabled. Disable it for approval reviewers to avoid delaying approval completion, and record a distinct `PostTurn` analytics phase.

Buffer local post-turn summaries until compaction succeeds and require a nonempty assistant summary before replacing history. Log compaction failures without failing the completed turn, while still propagating interruptions and turn aborts.

GitOrigin-RevId: b616994175829bd83b5d7974f95dabb042a39a89
2026-09-19 00:52:56 +00:00
Adam Perry @ OpenAI d5b29951ac Centralize retry decisions and delays in CodexErr (#46540)
## What changed

Replace `is_retryable()` with `retry_delay(retry_count)`, which returns `None` for terminal errors and a delay for retryable errors. Prefer server advice when present; otherwise use exponential backoff with jitter. Keep `server_retry_delay()` for callers that need only the server-provided delay.

Route sampling and remote compaction errors through the shared response stream handler, which now checks retryability and obtains the delay together. Preserve server-only retry timing when the retry budget is exhausted.

## Testing

Add coverage for attempt-dependent backoff, zero-duration server advice, and terminal errors that remain non-retryable despite server advice. Update API mapping and provider tests to use the combined retry API.

GitOrigin-RevId: 903f92935e71dea043afe1fad8ccb9389471aa75
2026-09-19 00:46:30 +00:00
felixxia-oai c51cb968e4 Preserve Guardian evidence during checkpoint migration (#45789)
## Why

A checkpoint produced by a previous model may be incompatible with the selected Guardian reviewer. Legacy review must retain user restrictions and verified answers across compaction and restart until a compatible checkpoint is available.

## What changed

- Expose retained user evidence independently of whether Guardian reviews the legacy transcript or parent context.
- Resolve the selected reviewer's compaction compatibility hash during resume and remote compaction. Preserve the legacy transcript when checkpoint compatibility is unknown or mismatched.
- Activate parent-context review immediately after compatible compaction and invalidate pending reviews bound to the previous evidence policy. Ordinary compaction preserves pending reviews.
- Keep strict compatibility checks for previously migrated checkpoints whose complete legacy transcript is no longer available.

## Testing

Add coverage for mismatched checkpoint hashes, retained evidence during legacy review, and pending-review cancellation on migration. Add a request-history scenario covering a model switch, incompatible automatic compaction, restart, and compatible manual compaction with preserved user restrictions and verified answers.

GitOrigin-RevId: b52bfd7c21ce7ebed482174c8d7c327e50b1ada8
2026-09-15 20:41:39 +00:00
Krish Chainani 5a66d460d3 Refactor image content to use a shared ImageReference type (#45543)
## What changed

Represent images in `ContentItem` and `FunctionCallOutputContentItem` with `ImageReference::Inline`, flattened to preserve the existing `image_url` wire format. Update image producers and consumers and regenerate app-server schemas and SDK artifacts.

Preserve the Python SDK's `InputImageContentItem` and `InputImageFunctionCallOutputContentItem` class names during generation.

## Testing

Add a regression test for stable Python image class names and adapt existing image tests to the shared representation.

GitOrigin-RevId: c38a780ac3314c2ac2deb3afc1b93b94b6f93fec
2026-09-14 23:25:09 +00:00
Owen Lin 196964ef10 Preserve root turn attribution in turn-start events (#44611)
## Why

Persisted turn-start events lack the originating root turn ID, so they cannot directly associate child turns with the root turn that initiated them.

## What changed

- Add optional `root_turn_id` to `TurnStartedEvent`. Use the inherited root turn ID when available, otherwise the turn's own ID.
- Centralize turn-start emission for regular turns, shell commands, and manual compaction so they record attribution consistently.
- Carry attribution into `ThreadHistoryTurnChange` when processing turn-start events and retain it through completion in the history builder.
- Keep older records compatible by defaulting missing attribution to `None`.

## Testing

Add coverage for root and child turn attribution, persistence when tool collisions fail a turn before sampling, history change propagation, and deserialization of older turn-start records.

GitOrigin-RevId: 54a80b828ca7d1c4c19f13d468c01d052e1b9130
2026-09-10 18:24:13 +00:00
jif ee93abb690 Preserve incoming prompts when pre-turn compaction fails (#44487)
## Why

Pre-turn compaction runs before incoming input is recorded, so failures could leave an accepted prompt out of conversation history. Reporting the error before prompt hooks finish also lets clients steer follow-up input into a turn still preserving its prompt.

## What changed

Record input and run prompt hooks on every pre-turn compaction failure. Defer local and remote compaction error events to `run_turn`, after prompt preservation completes, while retaining the remote error prefix.

## Testing

Extend the remote compaction output-validation test to trigger automatic compaction and assert that the prompt is emitted before the error, saved exactly once, and followed by no additional model request. Retain the assertion that usage is recorded for invalid compaction output.

GitOrigin-RevId: ab7ccb259836c084173e0e81dbbd4c610e95a6c8
2026-09-10 10:08:11 +00:00
rka-oai 1ac689cc7d Remove the unused legacy remote compaction implementation (#44273)
## Why

Remote compaction already uses the normal Responses API, leaving the legacy `/responses/compact` runner unused.

## What changed

- Remove the legacy runner, endpoint client, request types, analytics variant, and obsolete test helpers and snapshots.
- Move tool-output trimming and its metadata tests into `compact_remote_history` for reuse by remote compaction v2.
- Consolidate retained-history filtering in the v2 implementation and remove redundant `RemoteCompactionV2` settings from tests.

## Testing

Extend the retained-history unit test to explicitly check that hook prompts survive compaction.

GitOrigin-RevId: 999bb1391aa22c3c281d6a96f98817d6c9e46ed9
2026-09-09 18:40:53 +00:00
Felipe Coury 9ec33e1926 Show streaming reasoning summaries in the TUI status row (#43921)
## What changed

- Update the status heading from the latest usable reasoning line and retain it through tool activity and empty reasoning items. Keep completed reasoning in the expanded transcript only.
- Restore active reasoning after resume, thread switches, and session refreshes, including when earlier events have left the replay buffer.
- Default new embedded TUI threads to `model_reasoning_summary = "detailed"`, honoring explicit summary settings. Keep `concurrent_reasoning_summaries` opt-in and disable it when summaries are `none`.
- Persist the request's captured reasoning summary setting in context updates and local and remote compaction checkpoints.
- Animate status text with a smooth, grapheme-aware shimmer that respects terminal colors and reduced motion.

## Testing

Add regression coverage for reasoning status updates, transcript visibility, resume and buffered replay, outbound summary configuration, and summary settings across all three compaction modes. Add shimmer snapshots and checks for smooth brightness changes and Unicode grapheme preservation.

GitOrigin-RevId: 725c7de20f23166d7a3da6daeca617add5cbc600
2026-09-08 21:41:02 +00:00
felixxia-oai 35d9e4bc4d Preserve reasoning effort through compaction and reset it on success (#43796)
## Why

With reasoning effort overrides enabled, compaction used the selected effort even when sampling requests still used an earlier pinned effort. Successful compaction also left that old pin active in the new context window.

## What changed

- Use the pinned request effort for local and remote compaction when it matches the model.
- Resolve fallback models' effort without mutating the live pin, preserving it if compaction fails.
- Reset the pin after successful compaction so the next sampling request establishes the selected effort as its baseline without a redundant `configuration_update`.

## Testing

Add regression coverage for remote compaction success and failure, fallback model effort selection, pin preservation during fallback lookup, and compaction after resuming with overrides disabled.

GitOrigin-RevId: 35e666b0c8026553500715ddbe7bef02cd759f4a
2026-09-08 10:46:07 +00:00
felixxia-oai ad8ee16a5f Require Guardian review for incompatible compaction checkpoints (#42588)
## Why

Guardian can reuse an opaque parent compaction only when the checkpoint's
recorded producer hash matches the scoring model. Missing or incompatible
provenance must not allow an asynchronous score or cached approval to omit
compacted context.

## What changed

- Record the producing model's compatibility hash on compaction history items
  and expose the latest checkpoint hash through conversation snapshots.
- In thread-owned context mode, fail closed and require synchronous review when
  checkpoint reuse is enabled but the latest checkpoint is incompatible or has
  unknown provenance.
- Reject incompatible checkpoints in the sampler while preserving the legacy
  context behavior.

## Testing

Add coverage for local, remote, resumed, compacted, and rolled-back histories,
including matching, mismatched, missing, and empty compatibility hashes.

GitOrigin-RevId: 1ce959cb276edc602fd7dfa442133c7fe5e53857
2026-09-03 17:48:15 +00:00
sayan-oai e017e93ace Preserve raw response usage metadata (#41980)
## What changed

- Include the complete upstream `response.usage` object in
  `ResponseUsageMetadata.metadata` while retaining the existing `amount` field.
- Expose the metadata through `rawResponse/completed` notifications and update
  the generated JSON and TypeScript schemas.

## Testing

- Cover SSE, WebSocket, turn, and compaction completion paths.

GitOrigin-RevId: 6d79b1b4c92ac11abf43ec98f380b4868288a701
2026-09-01 05:17:11 +00:00