## Why
Resumed sessions need the current multi-agent usage instructions when configuration changes or the saved history predates usage-hint tracking.
## What changed
- Store multi-agent usage hints in world state and refresh changed or previously untracked hints before re-emitting the active multi-agent mode.
- Present `wait_agent` polling guidance in the overridable developer instructions only when the tool is enabled, instead of embedding it in the tool schema.
## Testing
- Cover cold resumes, changes to `wait_agent` availability, custom usage hints, message ordering, and tool-schema behavior.
GitOrigin-RevId: a9c6798d5dc2dedc8dadc49e0a80ff50a757ab83
## What changed
- Move tool mention parsing, skill-name counting, and implicit invocation detection into `codex-skills` and expose them through its public API.
- Decouple implicit invocation detection from `SkillLoadOutcome` with an `ImplicitSkillLookup` trait, while preserving the existing `core-skills` interface through re-exports.
- Cover remote plugin attribution for both implicit `SKILL.md` reads and skill script runs.
GitOrigin-RevId: f3da85c9821868638ccfd2fc2e01e0869d2fd437
## What changed
- Add `session_source()` to `McpServerContributionContext` for thread-scoped MCP resolution.
- Propagate each thread's `SessionSource` through initial setup, per-step resolution, and runtime refreshes.
## Testing
- Verify global contexts omit the source while thread-scoped contexts retain it.
- Verify MCP contributors observe distinct sources for a root thread and its spawned subagent.
GitOrigin-RevId: 23da1509858e023fbd9e2d179034ddc83553f17b
## What changed
- Load ordinary recursive host skill roots with the skills extension's host loader.
- Keep plugin-specific roots on the existing loader so plugin snapshot caching and namespacing continue to work.
- Merge both kinds of root snapshots through the existing precedence and deduplication logic, including discovery paths and source filesystem mappings.
- Move `HostSkillsSnapshot` into the skills extension as the owner of host skill loading.
## Testing
- Cover merged host and plugin catalogs, precedence for symlinked roots, discovery paths, and end-to-end rendering and invocation of host, repository, and plugin skills.
GitOrigin-RevId: c18080afb917c33c57bc17e36b40dd0f3c72d1d7
## What changed
- Add end-to-end coverage for a remote environment whose provisioning status is
reported ready before it is selected for a turn.
- Verify that `wait_for_environment` makes remote execution tools and selected
capability-root context available once the transport is ready.
- Update pending-environment and relay tests to cover separate materialization
and readiness reporting, including deferred connection and reconnection.
GitOrigin-RevId: d86d66e8f722e9bafac88b93d93e7c5244a7156e
## What changed
- Move the orchestrator skill catalog from thread context into its own
`orchestrator_skills` world-state section so unchanged catalogs remain
incremental across turns.
- Discover executor, orchestrator, and host catalogs together and render them
against one context-window-aware metadata budget, preserving executor entries
before orchestrator and host entries under pressure.
- Emit catalog metrics and budget warnings when each world-state section is
rendered.
## Testing
- Cover incremental orchestrator catalogs across production turns, combined
catalog budgeting, disabled and unavailable providers, caching, warnings, and
metrics.
GitOrigin-RevId: a0fb25d39dcaf43c978aff9da7972a1206f736e4
## What changed
- Add `include_apps_usage_instructions` to model metadata, defaulting to true for compatibility with existing metadata.
- Emit generic Apps guidance only when apps are available and the selected model enables it.
- Enable the capability for interactive model presets while leaving `codex-auto-review` and local models opted out.
## Testing
- Verify that missing metadata defaults to enabled and an explicit opt-out survives serialization.
GitOrigin-RevId: 6a20573cddfd3b80083d2b4b015829b49e6cc283
## Why
Symlinked skills have a canonical `SKILL.md` path that can differ from the path
under the configured skill root. A catalog entry that advertises the discovery
path must still select the canonical skill.
## What changed
- Track each loaded skill's discovery path alongside its canonical identity.
- Render discovery paths in the skill catalog and accept either form for
structured selections and linked skill mentions.
- Preserve enabled-state checks, deduplication, and product filtering when
resolving discovery paths.
## Testing
Added loader, catalog, mention-resolution, and end-to-end turn coverage for
skills discovered through directory symlinks.
GitOrigin-RevId: bae539b766cc3782c7c121a4276c337699cbb509
## What changed
- Add the disabled-by-default `image_resize_notice` feature.
- When enabled, append a developer message after resized images from user
messages or tool outputs. The notice identifies each image and reports its
original and prepared dimensions.
- Persist notices with new history while avoiding backfilled notices for images
processed when a session is resumed.
## Testing
- Cover notice placement, image numbering, failed images, user and tool image
sources, and resumed-session replay.
GitOrigin-RevId: 11dd808b3ad47eac35e9e2f4a4370f85d0122770
## What changed
- Select `gpt-5.6-luna` for Guardian approval reviews authenticated with an API key, while retaining `codex-auto-review` for ChatGPT authentication.
- Exercise API-key reviews through Responses Lite and fall back to the bundled Guardian policy when the selected model has no catalog-provided review policy.
- Refine the bundled policy's handling of trusted authorization, sensitive-data egress, credential probing, destructive command targets, and post-denial approval.
## Testing
- Cover provider model selection for API-key and ChatGPT authentication.
- Verify Guardian prewarming and review reuse with both Responses Lite and standard Responses requests.
GitOrigin-RevId: e35c43e1692fd643ab8faad1444c0d6a869b3e0c
## Why
The 4 KiB aggregate limit can truncate source metadata before it reaches the
model through the tool search description.
## What changed
Raise the aggregate source-description limit to 512 KiB while retaining the
existing character-boundary-safe truncation and bounded output.
## Testing
Expand unit and RMCP round-trip coverage to verify that substantially longer
Unicode descriptions remain complete within the new limit.
GitOrigin-RevId: 002760cdd2e8741e531b45f75f8624ef84334d24
## What changed
- Propagate the optional `modelSpecialty` field through the model catalog and
app-server `model/list` response.
- When a newly selected model has the `cyber` specialty, default an active TUI
thread to workspace-write permissions with on-request approval. Use auto
review when available, otherwise keep the user as reviewer, while respecting
configured permission requirements.
- Show a notice when auto review is applied and strengthen the full-access
warning for cyber models.
- Preserve explicitly selected permissions when only reasoning settings change.
## Testing
- Cover auto-review defaults, requirement-aware fallback behavior, permission
preservation across reasoning changes, and the new warning notice.
GitOrigin-RevId: 8f25b0082d3051d9ff91f7355921236d0f82f758
## What changed
- Treat different non-empty descriptions for the same tool namespace as a
collision when `error_on_tool_collisions` is enabled, including deferred
tools.
- Continue allowing multiple tools to share a namespace when their descriptions
match or are empty.
- Preserve the first non-empty namespace description when strict collision
checking is disabled.
## Testing
- Cover strict collisions across direct and deferred tools, valid shared
namespaces, relaxed description merging, and failure before model sampling.
GitOrigin-RevId: 788a2ae0eccb64c97995af7a208d9abb79edf2d0
## What changed
- Derive sandboxing, approval, and permission-profile environment variables from the selected turn environment for shell, unified exec, `apply_patch`, image viewing, and network approval flows.
- Materialize project-root permissions against every workspace root in that environment, allowing `apply_patch` to write to secondary workspace roots while preserving protected paths.
- Carry the selected environment's permission profile into spawned or resumed agents and Guardian permission context.
## Testing
- Add an integration test covering `apply_patch` writes in a secondary workspace root.
- Extend tool, Guardian, and multi-agent tests to distinguish environment permissions from thread-level permissions.
GitOrigin-RevId: e22da8974bcd3de1e724978b1a7790307a51e4bc
## What changed
- Track Agent Plugin manifests through plugin, skill, and MCP loading so their capabilities use format-specific behavior without changing legacy plugins.
- Discover only direct-child skills, exclude app and hook capabilities, isolate MCP data, and reject MCP configuration files that are non-regular or resolve outside the plugin root.
- Bound model-visible skill instructions, plugin instructions, MCP descriptions, schemas, individual tools, and the aggregate Agent Plugin MCP tool set.
- Stop MCP and OAuth redirects when Agent Plugins send configured or authorization headers, while retaining existing redirect behavior for legacy MCP servers.
## Testing
- Add coverage for capability filtering, skill discovery, isolated MCP data and reserved-path expansion, unsafe MCP configuration files, context limits, and redirect handling.
GitOrigin-RevId: c9af66b051269f3226628ca280a58d32c808c38f
## What changed
- Group top-level function and custom tool definitions into a single `functions` namespace for Responses Lite providers that support namespaced tools.
- Normalize missing, empty, and explicit `functions` namespaces to the same tool identity across registration, routing, lifecycle hooks, configuration, and tool search.
- Keep default tool names unprefixed in code mode, display output, and dispatch traces while preserving explicit non-default namespaces.
## Testing
- Add coverage for Responses Lite serialization, tool search results, namespace normalization and collision handling, routing, lifecycle events, and code-mode namespace policies.
GitOrigin-RevId: d48414005b5d22d39b11a198e19c47814d3a19f2
## What changed
- Track the first duplicate effective tool name while assembling the tool registry, including collisions introduced by external tools, code mode, and tool search.
- When `[features.tool_registry].error_on_tool_collisions` is enabled, fail the turn with a `duplicate tool: <namespace>.<name>` error before sending a model request.
- Continue allowing identical tool names in different namespaces, and preserve the existing collision behavior when strict checking is disabled.
## Testing
- Cover registry collision tracking, strict tool planning across tool sources, namespace isolation, failure before sampling, and pre-sampling compaction.
GitOrigin-RevId: 4cf83538fb513cec0b9c8b567780caaaadd3243d
## What changed
- Include top-level freeform tools in the tool-search index and mark them for deferred loading.
- Serialize searched freeform tools as Responses API `custom` tools and convert them back into executable tool specs after discovery.
- Cover indexing and the end-to-end search, load, and custom-call routing flow.
GitOrigin-RevId: 124a7705876067542e1bd609546adc9a8d0ef67e
## What changed
- Add the public asynchronous `ModelsCache` contract and cache entry/error types.
- Let model providers and `OpenAiModelsManager` accept caller-provided cache implementations while retaining the existing file-backed cache by default.
- Treat cache read failures as misses, keep cache write failures non-fatal, and support refreshing an entry's TTL after ETag revalidation.
- Continue to prefer an explicitly configured static model catalog over an injected cache.
## Testing
- Cover injected cache hits, read and write failures, TTL refreshes, and end-to-end model selection.
- Verify that the file-backed implementation satisfies the new cache contract and does not serve expired entries before revalidation.
GitOrigin-RevId: eef957a317910b8502a0c20f8fc112f689d5e389
## What changed
- Remove the hidden `PairProgramming` and `Execute` variants from `ModeKind`.
- Delete their unused prompt templates and simplify mode handling to cover only `Default` and `Plan`.
GitOrigin-RevId: 8d24e3f562a09e3071b2decae275d2b3871a1e9c
## What changed
- Add provider-owned remote compaction capabilities for unsupported, v1, and v2 protocols.
- Mark Amazon Bedrock as v1-only so manual and automatic compaction use `/v1/responses/compact`, even when the v2 feature is enabled.
- Preserve v2 support for OpenAI and Azure Responses providers and local compaction for unsupported providers.
## Testing
- Cover Amazon Bedrock manual and automatic remote compaction, including endpoint selection, authentication headers, model selection, and reuse of the compacted summary.
GitOrigin-RevId: d58093178605d1f8155b93d9a161408177f6ac72
## Why
Orchestrator skills with `allow_implicit_invocation: false` were still exposed in the model-visible skill catalog.
## What changed
- Hide explicit-only orchestrator skills from the prompt catalog while keeping them available for direct invocation.
- Include orchestrator resource-access metadata with injected instructions so `skills.read` can load referenced resources.
- Track visible and hidden discovery limits separately, and record discovery duration and resource-count metrics.
## Testing
Add an integration test that verifies an explicit-only orchestrator skill stays out of the prompt catalog, can be selected explicitly, and can read a referenced resource.
GitOrigin-RevId: 9cab5a61eb27c96bec9d1346c7dcdbc97376f7dc
## What changed
- Add a `features.token_budget.mode` setting with `thread` and `name` values.
- Default context-window metadata to the thread ID while allowing configurations to retain the agent name.
- Preserve the selected identity mode when applying model token-budget defaults and replaying locked configuration.
## Testing
- Cover configuration parsing and both identity formats in token-budget context messages.
GitOrigin-RevId: f6e6d2c49e09e70c0352e803bb65784ef77e5bee
## Why
Image resize notices describe the history item immediately before them, so
remote compaction must not preserve or discard them independently of that item.
## What changed
- Group recognized `image_resize_notice` developer messages with their source
history items when filtering compacted history.
- Account for each group together when trimming or truncating history to a token
budget, preserving the notice only when its source is retained.
## Testing
- Extend the remote compaction v2 follow-up test to verify that a retained user
image keeps its resize notice, while notices for discarded items and
unrecognized developer notices are omitted.
GitOrigin-RevId: 82bb70a46eb68b07b4601e7fb345921d0789f442
## What changed
- Generalize `CodexThread::try_start_turn_if_idle` to accept exported
`TurnInput` values, preserving user content and client IDs in the first model
request.
- Allow non-empty user input to start an idle turn in Plan mode while retaining
the Plan-mode rejection for automatic response items and empty user input.
- Process idle user input through the normal user-turn path so connector
selections are cleared, prompt telemetry is recorded, and referenced skill
instructions are included.
## Testing
- Cover idle user input in Default and Plan modes, response-item mailbox
merging, empty-input rejection, and skill instruction injection.
GitOrigin-RevId: 116667d160d4464b525d9e4c7e8cc05d23a94008
## Why
Reviews may use a different model from their parent turn. They need current
thread settings without inheriting defaults resolved for the parent model.
## What changed
- Build review configuration from the parent turn context while preserving
explicitly configured token-budget overrides.
- Use the review model's token-budget defaults and select a supported reasoning
effort when switching models.
- Carry the session's resolved service tier and current environment,
permissions, and approval settings into the review thread.
## Testing
Expanded review coverage for updated thread settings, model-specific
token-budget defaults, service tiers, and reasoning effort selection.
GitOrigin-RevId: cefa0f22cd9b9e28047d94e1d2d49ff83eded43e
## Why
Guardian reviews received sandbox retry reasons, but not the execution-policy reason that triggered an initial approval request.
## What changed
- Propagate approval and retry reasons through the tool approval flow and include the applicable reason in the Guardian prompt.
- Prefer a sandbox retry reason when both are available.
- Truncate approval context to 512 tokens while preserving the start and end of the reason.
## Testing
Added prompt and integration coverage for policy reasons, retry precedence, and truncation.
GitOrigin-RevId: 5b0f4e1c40b792b031bc8f7b31685a61d12b8d92
## Why
Amazon Bedrock supports hosted text web search, but it rejects the
`search_content_types` field used for multimodal search and does not support
external live or indexed web access.
## What changed
- Advertise hosted web search for Amazon Bedrock while marking external web
access as unsupported.
- Resolve unsupported live and indexed modes to cached search, or disable the
tool when cached search is prohibited by managed requirements.
- Normalize built-in and configured Bedrock model catalogs to text-only web
search, and retain the runtime provider in session configuration so turn
setup can apply its capabilities.
## Testing
- Cover cached fallback, managed-mode restrictions, text-only tool payloads,
provider capabilities, and catalog normalization.
GitOrigin-RevId: 310473849257401654388a4ebb42920e03aa3228
## What changed
- Remove the cached permission profile from `TurnContext` and derive the
effective profile, filesystem policy, network policy, and legacy sandbox
policy from `config.permissions`.
- Update sandbox consumers and tests to use the current turn configuration.
- Verify that role-based agent spawning reapplies runtime permissions and that
cold-resumed agents retain their disabled permission profile.
GitOrigin-RevId: d67c5d2bcbe6dc76f15b56defb485b155fd1f138
## What changed
Expand the ChatGPT dual-surface plugin integration test to cover both enabled
and disabled app configurations. Verify that plugin app guidance and searchable
app tools follow `apps.calendar.enabled`, while plugin MCP guidance and tools
remain suppressed.
GitOrigin-RevId: 40c51fbc075e07a8b03b5c520f5c46e360f71ceb
## Why
Thread settings can update the approval policy after a turn context is created.
Keeping a separate copy on `TurnContext` could leave tool approval checks using
the previous policy.
## What changed
- Remove the duplicated approval-policy field from `TurnContext`.
- Resolve the policy through the turn's current configuration everywhere it is
needed, including tool execution, Guardian routing, MCP handling, and
permission requests.
## Testing
- Update the shell escalation test to apply a thread-level policy override and
verify that the next turn rejects an escalation using the updated policy.
GitOrigin-RevId: e5966ba08f179d53fd76871ad904762958c0f5ea
## Why
App-server clients need to advertise structured MCP extension settings, including supported MCP App UI MIME types, rather than only opting into OpenAI form elicitation with a boolean.
## What changed
- Add an `extensions` map to initialize capabilities and preserve `mcpServerOpenaiFormElicitation` as a legacy alias for `openai/form`.
- Capture the declared extension profile when a thread is started, resumed, or forked, propagate it to subagents, and advertise it to downstream MCP servers during initialization.
- Keep the profile stable for the lifetime of the loaded session instead of changing it on later turns or direct tool calls.
## Testing
- Cover extension conversion, downstream MCP initialization, session isolation, legacy form support, and subagent inheritance.
GitOrigin-RevId: fbcedbb74ce788e574b0f884a4c45c4cedb9de54
## What changed
- Apply the session's current approval policy whenever per-turn configuration is built.
- Have review threads inherit the parent turn's full permission settings and approval reviewer.
- Add coverage for settings updated after session startup, including the approval policy, permission profile, and approval reviewer.
GitOrigin-RevId: bff050c6de157dd65d62f7b49651bef3f426ee3f
## What changed
- Allow multi-agent v2 parents to spawn any visible model that has not explicitly disabled multi-agent support.
- Expose collaboration tools to child agents only when their selected model supports multi-agent v2, keeping legacy models as leaf workers.
- Preserve a worker's selected model when reloading it into residency.
- Propagate multi-agent capability metadata to Amazon Bedrock model entries so delegation is gated consistently.
## Testing
- Cover model selection, leaf-worker tool visibility, Bedrock capability handling, and model preservation after reload.
GitOrigin-RevId: 1457adb1a09806d1f0621311d5a42a6815b9dd4e
## What changed
- Keep complete MCP namespace descriptions in tool-search source metadata.
- Raise the namespace tool-spec description limit from 1,000 bytes to 512 KiB,
truncating at a UTF-8 character boundary only when the new limit is exceeded.
## Testing
- Cover descriptions beyond the former limit and multibyte truncation at 512 KiB.
- Update SSE and stdio MCP tests to verify that complete server instructions are
preserved without hiding tools.
GitOrigin-RevId: 000bfcafb3df348065ae451685bfbf978a0e3248
## What changed
- Allow namespace tool specs to contain custom freeform tools alongside function tools.
- Include namespaced custom tools in deferred tool search and expose them to code mode with names such as `editor__apply_patch`.
- Route custom payloads to matching extension tools while preserving function-only payload validation.
## Testing
- Add serialization, tool search, code-mode definition, and end-to-end dispatch coverage for namespaced custom tools.
GitOrigin-RevId: be64d35f6ae54685c5a9fcf45a732320742ea7e5
## Why
Code-mode `wait` and `terminate` requests can remain pending when the host
transport stalls.
## What changed
- Add a 60-second transport allowance to the runtime timeout for `wait`, and
apply the same transport deadline to `terminate`.
- Return a model-visible timeout error and invalidate the connection when the
deadline expires, so the next execution reconnects to the host.
## Testing
- Cover queued `wait` and `terminate` requests that exceed their deadlines.
- Verify that a timed-out `wait` reports the error and reconnects on the next
code-mode execution.
GitOrigin-RevId: 5d772e5a6f3793aa8865a1160639b851fd4824fb
## What changed
- Replace the thread ID in `<context_window>` metadata with the session's canonical agent path.
- Default sessions without an agent path to `/root`.
- Verify that root sessions emit `/root` and subagent sessions emit their own path, such as `/root/worker`.
GitOrigin-RevId: bcf057842ed31f93d554b7de063b6c03403a2594
## What changed
- Store the effective `allow_login_shell` setting on each turn environment, including inherited environments whose child thread has a different policy.
- Expose the `login` argument for shell tools when any selected environment permits login shells.
- Validate each command against the policy of its selected environment instead of the turn-wide configuration.
## Testing
- Cover tool schema generation for single and multiple environments.
- Cover login-shell rejection by both the command handler and the unified exec integration.
GitOrigin-RevId: 5a93149a5c86f4087f2b92d663ebc33feff8a57c
## What changed
- Track approved command prefixes separately from the stable permissions
instructions in world-state snapshots.
- Emit only newly approved prefixes after an exec-policy amendment instead of
appending the full permissions block again.
- Preserve prefix updates when full permissions instructions are disabled and
remain compatible with legacy world-state snapshots.
## Testing
- Cover incremental prefix additions, removals, legacy snapshots, history
rollback, and approval flows with permissions instructions enabled or disabled.
GitOrigin-RevId: e6f68c6a91be82750e70e28456f2b7c58607fbd8
## What changed
- Add `include_plugin_usage_instructions` to model metadata, defaulting to false.
- Emit generic plugin guidance only when plugins are available and the selected model enables it.
- Enable the capability for interactive model presets while leaving `codex-auto-review` opted out.
GitOrigin-RevId: 67f5a97e978033f5f1d533956c0b9deeae610283
## What changed
- Remove `ModelInfo.base_instructions` as an in-memory instruction source and use `model_messages.instructions_template` consistently for bundled, remote, fallback, and overridden model metadata.
- Preserve compatibility by promoting legacy `base_instructions` values when reading model responses and caches, and by including rendered legacy instructions when serializing `ModelsResponse` for older clients.
- Treat templates without instruction variables as literal text and retain the other model-message fields when applying instruction overrides.
This completes the consolidation proposed in https://github.com/openai/codex/pull/31302.
## Testing
- Cover legacy response and cache migration, canonical-template precedence, fallback instructions, personality rendering, overrides, and model switching.
GitOrigin-RevId: 089d986ca5e30da67db2c77a1b6a046d2cff52dc
## Testing
- Verify consecutive tool reviews reuse the same Guardian session without
prewarming and execute both approved commands.
- Verify interrupting a pending Guardian review prevents command execution and
records an aborted tool result for the next turn.
GitOrigin-RevId: b8371ae82b4fef206857012589668f890f9305bf
## Why
MCP tools can be exposed directly, discovered through tool search, or called
from Code Mode. Servers need to be able to opt out of any of these surfaces
without disabling their tools everywhere.
## What changed
- Add `omit_tools_from` to MCP server configuration, accepting any combination
of `direct`, `deferred`, and `code_mode`.
- Apply the exclusions independently when building direct, deferred, and Code
Mode tool surfaces while keeping omitted tools registered for permitted uses.
- Remove client-private `_meta` fields from MCP results returned to Code Mode.
- Include the new setting in the configuration schema and MCP config
serialization.
## Testing
- Cover every exposure combination across Code Mode, Code-Mode-only sessions,
tool search, direct-only namespaces, and prefixed and unprefixed MCP names.
- Verify direct and nested execution paths, parallel-call support, config
round-tripping, and `_meta` filtering.
GitOrigin-RevId: 12dfcb78bb5c5ecf4d70f38a8b5022792463a27f
## Why
Codex Apps tool catalogs can exceed the standard MCP limit of 2,048 items and
still need to remain available through tool search.
## What changed
- Allow up to 8,192 catalog items for host-owned `codex_apps` registrations.
- Keep the 2,048-item limit for other MCP servers, including extensions that
register the `codex_apps` name.
- Restrict the shared Codex Apps tool cache to host-owned registrations.
## Testing
- Cover large catalogs for built-in and hosted-plugin registrations while
verifying that another extension retains the standard limit.
- Verify that a same-named extension cannot read the host-owned Apps cache.
GitOrigin-RevId: 2265693a460e9be63c4c849169cba7324def4f49
## What changed
- Enable `test_sync_tool` for the startup prewarm test.
- Hold the first user-turn response open while Git enrichment completes, then
verify the refreshed workspace metadata on the follow-up request.
GitOrigin-RevId: dd4254fba1ac2f346022988cc21a78648a29556a
## What changed
Move shell snapshot validation in the intercepted `apply_patch` test until after
the turn completes and its patch lifecycle events have been collected.
GitOrigin-RevId: 1c16bcfcc97f09165c4a2eb2d15f4059a14b2a99
## What changed
- Charge `codex_rollout_budget_units` against the shared rollout budget when the provider includes it in response usage.
- Fall back to weighted input and output token accounting when provider units are absent.
- Reject non-finite or negative provider units as a fatal response error.
## Testing
- Cover provider units in reminder thresholds and local and remote compaction budget exhaustion.
- Verify invalid units fail without retrying the response.
GitOrigin-RevId: b452403e365985854d16f298d1ba46383e9892c4
## What changed
- Reattach recorded `executed_tool_calls` metadata when an output is included in a subsequent prompt.
- Bound retained metadata to 32 KiB, prioritizing recent calls and reporting omitted calls in truncation metadata.
- Drop retained entries after their corresponding outputs leave the prompt history.
## Testing
- Cover metadata replay, cleanup after compaction, bounded retained history, and propagation through later tool requests.
GitOrigin-RevId: 1c23a26123be3b7ad51c61f4ad522139b71bb773
## What changed
- Add the optional `delegationAckFiller` field to `thread/realtime/start`.
- Forward explicit `true` or `false` values to V3 Frameless Bidi session payloads as `delegation.ack_filler`; leave the field absent when unspecified so the Realtime API default is preserved.
- Document that V1 and V2 ignore this setting.
## Testing
- Cover both boolean values in Frameless Bidi serialization and verify the WebRTC V3 session creation payload.
GitOrigin-RevId: 9fe8efd14e42738c3e86e3e9a8af81d0f4521244
Adam Perry @ OpenAI