## What changed
- Add callback lifetimes to extension `ToolCall`, `ToolEnvironment`, turn-input context, and skill-read request types.
- Require extension tool executors to handle calls for any invocation lifetime and tie their returned futures to that lifetime.
- Update built-in extensions, adapters, and tests to use the lifetime-scoped APIs without retaining host-owned environment capabilities.
GitOrigin-RevId: aca9ce3a1b6870df550bc84ec9308cc65533f161
## Why
Guardian treated all skill instructions as untrusted, so it could not use the
invocation of a user-owned skill as authorization evidence.
## What changed
- Record explicit and implicit skill invocations through extension contributors.
- Send Guardian a bounded, deduplicated list of canonical skill paths only when
they resolve under the user's Codex or `.agents` skill directories.
- Keep repository skills and symlinks that escape trusted roots untrusted, and
send only verified paths rather than skill contents to the classifier.
- Avoid recording host skill prompts that were superseded by core skill
injection.
## Testing
- Added unit coverage for trusted roots, symlink escapes, deduplication, and
size limits.
- Added integration coverage showing that Guardian trusts invoked user skills
while rejecting a forged repository skill.
GitOrigin-RevId: 1cc0be230f7e6649110201c23c3de8f26447ff36
## What changed
- Add `plugin_id`, `model_slug`, and `reasoning_effort` dimensions to
`codex.skill.injected` metrics for explicit and implicit skill invocations.
- Propagate plugin IDs from orchestrator skill metadata so resource-backed skill
invocations can emit the same attribution.
- Record `codex.skill.turn.duration_seconds` once per plugin used in a turn,
tagged with the model, reasoning effort, and completed, aborted, or error
status.
GitOrigin-RevId: d249a9c4e85682b1106ac49d9e598535c6cbb06d
## What changed
- Add the under-development `skip_host_skill_discovery` feature to bypass host
skill snapshots during session warmup and turn setup.
- Let skill invocation contributors declare whether they require host-owned
skills. Preserve host discovery when no contributors are registered or any
contributor requires it.
- Keep executor and orchestrator skill catalogs and instruction loading
available when host discovery is skipped.
## Testing
- Cover executor-only, orchestrator, mixed-contributor, and legacy host-skill
behavior.
GitOrigin-RevId: 48169403090b234e1a304c6523633fd867df454a
## What changed
- Require each `ContextualUserFragment` to provide a stable `<feature>.<name>`
`ContentItemKind`.
- Add `AnnotatedContent` and `RenderedFragment` so rendered text, its role, and
its classification can travel together to API boundaries.
- Derive extension-owned world-state classifications from the extension ID and
keep the skills catalog classification with its fragment implementation.
## Testing
- Verify that an extension-owned world-state section renders with an
`<extension-id>.instructions` content kind.
GitOrigin-RevId: e46b74a0bb41e0b6112667c9d36bc9e7f2714451
## What changed
- Require each `PromptFragment` to include a producer-owned `ContentItemKind`.
- Re-export `ContentItemKind` from `codex_extension_api` and assign stable kinds to skills, memories, examples, and test contributors.
- Limit prompt slots to developer policy and developer capabilities, and combine contributed fragments into a single developer update.
## Testing
- Update extension registry tests to verify fragment ordering with the new content-kind metadata.
GitOrigin-RevId: ef8624a9b4ae9dce025bca9c0954d91dd8075e85
## What changed
- Carry selected plugin identities from MCP discovery into per-turn extension data, limited to capability roots that are ready for the turn.
- Annotate executor skill catalog entries with the matching plugin ID and `user` scope.
- Include that attribution in `skill_invocation` analytics for both explicit and implicit executor skill invocations.
## Testing
- Extend the executor skills app-server test to verify plugin ID, skill scope, and invocation type for explicit and implicit analytics events.
GitOrigin-RevId: 9345600ae29bcb8eb894fffd2d6f7b05f22ebf74
## Why
Short continuation prompts such as `continue` do not contain enough context for
skill selection on their own.
## What changed
- Add the `task_context_fusion_v1` shadow selector, combining the current request
with up to two prior substantive requests and recently relevant skills.
- Record explicit skill intent and successful skill invocations for future turns,
while excluding same-turn observations from predictions.
- Bound retained requests, augmented queries, and skill history, including safe
truncation at UTF-8 character boundaries.
## Testing
Add unit and extension tests for continuation prompts, explicit intent, turn
isolation, cold thread state, bounded history, and unchanged control selectors.
GitOrigin-RevId: 72eca3f0d64620a0d24e95d5675f126ac982f8c7
## What changed
- Add `[skills].max_context_tokens` to override the token budget used to render the available-skills catalog.
- Require a positive value and cap configured budgets at 10,000 tokens.
- Preserve the existing default of 2% of the model context window when the option is unset.
## Testing
- Cover configuration parsing, override behavior, the 10,000-token cap, and catalog rendering within the configured budget.
GitOrigin-RevId: f29ddcb57fc2d800c0006d03e1c33cbe3ab6272d
## What changed
- Track the 50 most recently invoked skills for each thread and evaluate them
as the `lru_v1` shadow-selection method.
- Filter stale and duplicate entries while preserving recency order.
- Increase the shadow result limit to 50 and add a `21_50` rank bucket.
## Testing
- Add unit coverage for recency refresh, eviction, filtering, limits, and rank
buckets.
- Add an extension test showing that `lru_v1` recovers a skill invoked on an
earlier turn.
GitOrigin-RevId: 7aa56514db57cf180cc3f57b4d493dee3304e766
## What changed
- Emit skill invocation analytics when a resource-backed skill is explicitly selected or its main resource is successfully read from the first page.
- Preserve provider-supplied skill IDs and user scope when available, and derive a stable ID from the main resource otherwise.
- Deduplicate repeated implicit reads of the same main resource within a turn.
## Testing
- Extend the orchestrator skill integration test to cover explicit and implicit invocation events, stable fallback IDs, failed reads, and repeated-read deduplication.
GitOrigin-RevId: 7d8453b31b5a2205847e65cda2c46183d6c97b45
## What changed
- Add `HostSkillsSnapshot::load_skill_prompts` so the skills extension owns reading and rendering selected host skills.
- Keep explicit invocation telemetry, unreadable-skill warnings, plugin prompt ordering, and suppression of host prompts superseded by provider prompts.
- Remove the legacy prompt injection and fragment implementation from `codex-core-skills`.
## Testing
- Cover prompt truncation without an installed skills extension, skill/plugin ordering, selective provider supersession, unreadable skills, and contextual fragment detection.
GitOrigin-RevId: 9c8b84ad54b90d174abc287ab49cc5a231e9e1aa
## Why
Long executor and orchestrator resource identifiers can consume enough of the
skills context budget to omit otherwise available skills.
## What changed
- Add source-aware root aliases for executor, orchestrator, and host skill
catalogs, including catalogs with a single shared root.
- Select the best combination of aliased catalogs under the shared metadata
budget while accounting for root tables and alias usage instructions.
- Explain how to expand shortened `skill://` locators and accept the resulting
description-shortening warning in app-server coverage.
## Testing
- Add rendering and production-turn coverage for singleton, executor,
orchestrator, host, and combined catalog aliases.
GitOrigin-RevId: c1bc1f3d371eae9526232674060235272cc28f16
## 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
- 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
## 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
- Load and cache the host skill provider catalog during world-state contribution.
- Render host skill listings from that catalog with Core-compatible budgeting, while preserving Core's full prompt injection for selected host skills.
- Reuse the cached host catalog for shadow selection, including when skill listings are disabled.
- Preserve an empty Core-compatible skills fragment when every entry exceeds the metadata budget, and make budget warnings independent of the configured percentage.
## Testing
- Cover provider-backed host listings, selected-skill prompt injection, shadow selection, disabled listings, oversized prompts, empty fragments, metrics, and warnings.
GitOrigin-RevId: b6b2d8ea14386356094cff5783d121d74657e9e3
## Why
Executor capability discovery can traverse plugin and skill roots, including
symlinks. Under restricted filesystem permissions, discovery must not expose
files outside the permitted paths.
## What changed
- Pass each environment's filesystem sandbox context through capability root
discovery and apply it to metadata, directory walks, and file reads.
- Enable discovery for restricted sessions so permitted executor skills remain
available while inaccessible roots and symlink targets are omitted.
- Key discovery caches by sandbox context and reject sandboxed discovery on
executors that do not advertise support for it.
- Split requests with more than 128 roots into supported-size batches.
## Testing
- Cover permitted and denied external symlink targets.
- Verify restricted skill listing excludes inaccessible skills.
- Verify cache separation across permission contexts and discovery of 129 roots.
GitOrigin-RevId: 44d16468ca003403bdb8b71a04ae8c9ff94ed494
## Why
Executor-backed skill resources must honor the active filesystem permission profile instead of reading outside its allowed roots.
## What changed
- Propagate each environment's filesystem sandbox context into `skills.read` calls.
- Apply turn-scoped permission grants when reading executor skill resources.
- Fail closed when no matching sandbox context exists or a restricted Windows read cannot be sandboxed.
- Preserve the existing resource size limit for both sandboxed and streamed reads.
## Testing
- Cover reads within permitted roots and reads that succeed after a permission grant.
- Verify denied references do not expose their contents and disabled Windows sandboxing is rejected.
GitOrigin-RevId: 5cc679b9385b0621665ff20cf5848863cc6396a4
## What changed
- Allocate 2% of the resolved model context window to skill metadata without
imposing the previous 4,000-token ceiling.
- Keep the 8,000-character fallback when no context window is available.
- Extend catalog rendering coverage to verify that a 400,000-token context
window receives an 8,000-token metadata budget and includes more skills.
GitOrigin-RevId: 5120d462de22fbfe5f9c4ab0d1bb97e51950d4b2
## Why
Rendering host and executor skill catalogs independently allows their combined
metadata to exceed the skills context budget.
## What changed
- Allocate one budget across both catalogs when both contain model-visible
skills.
- Prefer retaining executor skills when budget pressure requires omissions,
while preserving an executor omission marker and reporting when all host
skills were omitted.
- Evaluate absolute and aliased host paths under the shared budget and select
the rendering that best preserves executor entries, total entries, and
descriptions, in that order.
## Testing
Added renderer and production-turn coverage for shared-budget description
shortening, host-first omission, executor omission markers, and alias selection.
GitOrigin-RevId: 30041736a76aa173cb641aa98a876a306cb437d2
## Why
Host skill catalogs supplied through extension world state should report budget warnings only when their catalog update is actually rendered.
## What changed
- Stop rebuilding an extension-owned host catalog in core.
- Emit host catalog budget warnings when the corresponding world-state fragment is rendered, and deduplicate them with executor catalog warnings.
- Cover full, shortened, and omitted host and executor catalogs, including repeated turns and resumed persisted snapshots.
GitOrigin-RevId: ba28b6e337281b49ed4da2175c193fd4a6898ab7
## What changed
- Add a character n-gram shadow selector that combines skill descriptions with host interface metadata and tool dependency names and descriptions.
- Bound routing fields, dependency records, and candidate counts before ranking.
- Exclude explicitly selected skills from shadow candidates and only record invocations for skills eligible for the experiment.
## Testing
- Add unit coverage for host interface metadata, connector dependencies, baseline description matches, and long routing metadata.
GitOrigin-RevId: 36ef0cb4a68601f2d9d41c3344c88f7927c32bf4
## Why
Empty rendered skill catalogs skipped catalog telemetry instead of reporting
zero-valued counts.
## What changed
- Record a default render report when thread, turn-input, or executor catalog
rendering produces no fragment.
- Record zero-valued host world-state metrics when skill instructions are
enabled but no host skills are available.
- Preserve the `catalog_surface` tag so empty host and executor catalogs remain
distinguishable.
## Testing
Add coverage for empty catalog rendering and mixed nonempty-executor,
empty-host world state.
GitOrigin-RevId: f68d65806703c0ded1ece1e5ae1ffaf6d6045126
## What changed
- Add an optional `ExtensionMetrics` capability to thread, turn-input, and world-state contributors, backed by session telemetry in the host.
- Record skill counts, omissions, and truncated description characters for extension-rendered catalogs, tagged by the catalog surface.
- Preserve host session attribution on extension metrics, including the effective model for each turn, and prevent extension tags from overriding host metadata.
- Avoid duplicate host-catalog samples by recording world-state metrics only when the section is published or changes.
## Testing
- Cover metric values and surface tags, session metadata preservation, turn-level model changes, and host world-state publication behavior.
GitOrigin-RevId: 9059e521943a8f81df25a38b03692eddbc1d990d
## Why
Executor skills that disallow implicit invocation are omitted from `skills.list`, so an explicitly selected skill did not provide the authority and package identifiers needed to read its referenced resources.
## What changed
- Add `resource_access` metadata to injected instructions for explicit-only executor skills.
- Teach the `skills.read` description to accept that metadata as well as values returned by `skills.list`.
- Keep discoverable executor skill instructions unchanged.
## Testing
Add an app-server integration scenario that selects an explicit-only executor skill, confirms it remains absent from `skills.list`, and reads a referenced file through `skills.read`.
GitOrigin-RevId: 97d3b93d1618c013363788452303c8a31e32cc56
## What changed
- Add executor authority support to `skills.list` and `skills.read` for skills from selected capability roots.
- Allow `skills.read` to load package-relative resources referenced by an executor skill while keeping reads within the selected package.
- Paginate skill listings and resource contents with bounded response sizes and stale-cursor validation.
- Resolve executor filesystems per sampling step so skill tools use the active selected capability roots.
## Testing
- Extend the app-server executor skill test to list an executor skill, read its main `SKILL.md`, and read a paginated reference resource.
- Cover visibility filtering, metadata bounds, pagination, invalid cursors, package boundaries, and resource size limits.
GitOrigin-RevId: 1b5de95d5d1b9f79c34d2879d6e67f24401e9c29
## Why
Extension warnings need a stable thread target even when they are emitted outside an active turn.
## What changed
- Add `ExtensionWarning` and a dedicated `ExtensionEventSink::emit_warning` path with an explicit thread ID and optional turn ID.
- Deliver extension warnings as thread-scoped app-server `warning` notifications, preserving listener ordering and waiting briefly for a subscriber when no listener is active.
- Move skills warnings to the new path, limiting provider catalog warnings to four messages of 256 bytes each.
## Testing
- Cover warning routing, ordering, UTF-8-safe truncation, subscriber fallback, and invalid thread IDs.
- Verify skills warning bounds and app-server delivery of executor catalog budget warnings.
GitOrigin-RevId: 4dcb1ba7f17c0edb1a9d548f9568e0c734bffc15
## Why
Skill catalog rendering can shorten descriptions or omit enabled skills to fit
the model context budget. Surface these reductions so users know when the
model-visible catalog has lost detail.
## What changed
- Emit a warning with the omitted skill count whenever entries do not fit.
- Warn when description shortening averages more than 100 characters per skill.
- Deduplicate executor catalog warnings across repeated world-state builds in a
turn.
## Testing
- Cover the description-shortening threshold and omission warning text.
- Verify warnings through extension and production-turn catalogs, including
per-turn deduplication.
GitOrigin-RevId: bd7fc3482e5dfd8c79072f772a2f54aa502478d9
## What changed
- Make `McpRuntime` own the published MCP configuration, connections, elicitation routing, and selected capability roots for a thread.
- Capture immutable MCP bindings for model steps and tool calls so in-flight work keeps a consistent connection set and approval authority while refreshed state is published atomically.
- Mark MCP state dirty when relevant configuration, plugins, environments, authentication, or elicitation capabilities change, then rebuild it before the next sampling request or out-of-band MCP operation.
- Separate config reloads from server invalidation: `ReloadMcpConfig` applies resolved MCP inputs, while `RefreshMcpServers` requests reinitialization from the thread's latest state.
## Testing
- Cover refreshed state visibility for existing turns, stable step bindings, current approval authority, cancelled refresh retries, resource-client reconciliation, and Apps recovery between sampling requests.
GitOrigin-RevId: 59eabb1aa8dc083426bd18ef4d3630508f376401
## What changed
- Remove the step-scoped `ExtensionData` argument from context, turn-input, and tool contributors.
- Pass the host's optional `McpResourceClient` through `ThreadStartInput` so extensions can retain session capabilities explicitly.
- Keep the MCP resource client in skills-owned session state for catalog loading and skill tools.
GitOrigin-RevId: bafa77bcd998aff408d6a396c5fd9ac268c4cce4
## Why
A fixed character limit does not account for the different context-window sizes supported by models.
## What changed
- Budget extension-rendered skill metadata at 2% of the resolved model context window, capped at 4,000 tokens.
- Keep the existing 8,000-character fallback when model context metadata is unavailable.
- Include the omission marker in the budget and still emit it when no skill entry fits.
- Apply the same resolved budget to executor and host skill catalogs assembled for a turn.
## Testing
- Cover proportional and capped budgets, multibyte fallback accounting, and omission-marker behavior.
- Verify through the production turn path that larger model context windows include more catalog entries without exceeding the computed budget.
GitOrigin-RevId: 4667293f1594de4dd605b32b9fa4255d26772c0f
## What changed
- Parameterize skill catalog rendering with explicit core- and extension-compatible description policies.
- Use full skill descriptions for core-compatible output, while extension-compatible output prefers `short_description` and falls back to the full description.
- Keep extension prompt and world-state rendering on the extension-compatible policy.
## Testing
- Add a renderer test covering description selection and fallback for both policies.
GitOrigin-RevId: 5382075f88c25a49a715feb0702a9ab888ec0123
## What changed
- Add an `ExtensionData` store to each `StepContext`.
- Pass the step store to context, world-state, turn-input, and tool contributors so extensions can use capabilities bound to the current sampling step.
- Preserve the captured step context when rebuilding initial context during compaction.
GitOrigin-RevId: 51f3105d3a793f60534f122f9be8247c131c2ca5
## Why
Selected capability roots can contribute plugins, MCP servers, connectors, and
skills. Discovering each contribution separately requires repeated access to the
executor filesystem.
## What changed
- Add the `capabilityRoots/discoverV1` exec-server RPC to scan selected roots and
materialize recognized plugin manifests, configuration files, skill
instructions, and skill metadata in one bounded request.
- Add the opt-in `executor_capability_discovery` feature, with a thread-scoped
cache and per-step snapshot shared by MCP and skill discovery.
- Parse MCP, connector, and skill contributions from the materialized snapshot,
including serving cached skill instructions without another filesystem read.
## Testing
- Cover discovery limits, manifest precedence, root-local failures, cache reuse,
plugin contributions, and parity with the existing environment skill loader.
GitOrigin-RevId: f98fd2321cafb58c596db02da1f83c09d8eb375d
## Why
Host skills can change after a thread starts. Subsequent turns need the updated
catalog without reinjecting an unchanged catalog on every turn.
## What changed
- Register the host skill provider in app-server and project its catalog through
a dedicated World State section.
- Render host skills with the existing metadata budget and usage instructions,
including skill-root aliases when available.
- Suppress the legacy host catalog and contextual fragment when World State owns
the catalog, while preserving host skills for explicit and shadow selection.
## Testing
Added an app-server test that installs a host skill at runtime and verifies the
catalog updates once and remains stable on the following turn.
GitOrigin-RevId: 4cacc6b2258fc3cab8499fbdf9aa6ab0324ec10d
## What changed
- Add an opt-in `skill_search` feature that ranks prompt-visible skills against each turn's user input with a bounded weighted lexical selector.
- Keep the ranked selection out of model-visible context and record metrics for selection cost, catalog reduction, and whether later implicit or `skills.read` invocations matched the ranked candidates.
- Include host-provided skills in the experiment catalog without changing the rendered skill catalog.
## Testing
- Add selector unit tests covering ranking, limits, truncation, stop words, and deterministic tie-breaking.
- Add extension tests covering turn-local invocation recording and host-skill shadow selection.
GitOrigin-RevId: 4d00a1c805ea8b391d6c6ac6a8450afa88ca3e25
## Summary
- add a false-by-default `include_skills_usage_instructions` model
metadata field
- enable the field for the bundled `gpt-5.5` model metadata
- consume the metadata in both core and extension skill rendering
- remove hardcoded legacy-model matching and its marker plumbing
## Why
A selected executor environment can be unavailable in one model step and
ready in the next. The model should see its skills only while that
environment is ready, without rescanning stable files on every sample.
The product assumption is simple:
- an environment ID names one stable logical environment;
- the selected root contents do not change during the thread.
## Behavior
```text
E1 unavailable -> do not show E1 skills
E1 ready -> discover once, cache, show through World State
E1 unavailable -> hide skills, keep cache
E1 ready again -> reuse cache, show skills again
resume -> create a new thread cache and discover again
```
The cache key is the full `SelectedCapabilityRoot`. Availability does
not invalidate it; dropping the extension's thread state does.
The step supplies the ready selected roots directly. They do not have to
be turn environments:
```text
turn environment: laptop
selected root: worker:/plugins/lint-fix
worker ready -> lint-fix skills are visible
```
## What changes
- Keeps executor skill catalogs in the existing skills extension.
- Passes the roots resolved as ready for the step into World State
contributors.
- Loads each ready selected root at most once per thread.
- Contributes the executor catalog as the `skills` World State section.
- Uses the exact step catalog for explicit skill selection and body
reads.
- Leaves host and orchestrator skill behavior where it already lives.
Taking a step snapshot itself does not add an RPC. Executor filesystem
calls happen only on the first discovery of a stable root for that
thread.
## What does not change
- No filesystem watcher or content-based invalidation.
- No retry/generation framework.
- No skill runtime migration into core.
- No general rewrite of the skills extension.
## Stack
1. Extension-owned World State sections.
2. **This PR:** project cached executor skills through World State.
3. Pin one MCP runtime to each model step.
4. Project selected MCP/app/connector metadata by environment
availability.
5. One end-to-end integration scenario.
## Why
Orchestrator-provided skills and Codex Apps MCP tools add model-visible
instructions, resources, and tools beyond the local workspace. Hosts
need config-level switches to disable those orchestrator-owned surfaces
independently, without disabling regular skills or regular MCP servers.
## What changed
- Adds `[orchestrator.skills].enabled` and `[orchestrator.mcp].enabled`
config entries, both defaulting to `true`.
- Includes the new settings in `config.schema.json` and in the config
lock so resolved thread configuration preserves the same orchestrator
exposure decisions.
- Threads `orchestrator.skills.enabled` through the app-server skills
extension so disabled orchestrator skills do not expose the `skills`
namespace or inject orchestrator skill context.
- Gates Codex Apps MCP exposure, app instructions, and app auth
eligibility on `orchestrator.mcp.enabled` while leaving non-Codex-Apps
MCP tools available.
- Updates the thread-manager sample config to disable both
orchestrator-owned surfaces.
## Verification
- Added config parsing, loading, defaulting, and schema coverage for the
new settings.
- Added MCP exposure coverage that `orchestrator.mcp.enabled = false`
removes Codex Apps tools while preserving regular MCP tools.
- Added app-server coverage that `orchestrator.skills.enabled = false`
prevents orchestrator skill tools, prompts, and resource reads from
reaching the model turn.
## Summary
- keep context injection on a single ContextContributor trait
- split context injection into thread-scoped and turn-scoped
contribution methods
- wire turn-scoped fragments into initial context assembly so extensions
can contribute context from turn-local state
## Why
Host skill discovery was still exposed as a manager even though it is a
process-owned service shared by sessions, the app-server catalog, and
file-watcher invalidation. The skills extension also consumed an ad hoc
loaded-skills wrapper instead of a named immutable snapshot.
## What changed
- replace `SkillsManager` with concrete `SkillsService`
- make the service cache and return immutable `HostSkillsSnapshot`
values
- migrate the skills extension host provider to the snapshot boundary
- migrate app-server catalog, watcher, and invalidation paths to the
service
This keeps the service limited to host discovery, caching, roots, and
invalidation. Catalog rendering and invocation remain extension
responsibilities for the next stacked change.
## Why
Hosted orchestrator skills are read through the remote MCP resource
server. Within one thread, the same catalog or skill resource can be
requested multiple times by prompt injection and the `skills.list` /
`skills.read` tools. Re-fetching adds latency and can make those
surfaces observe different remote contents during the same thread.
This is a follow-up to #28333: orchestrator skills remain limited to
threads without a local executor, and those threads now get a stable
per-thread view of the remote skill data they use.
## What changed
- Reuse the existing per-thread orchestrator catalog snapshot for
`skills.list` and `skills.read` availability checks.
- Cache successful orchestrator resource reads by authority, package,
and resource so prompt injection and tool calls share the same contents.
- Keep the cache memory-only and bounded to 100 resources and 8 MiB per
thread.
- Leave host and executor skill reads unchanged, and do not cache failed
remote reads.
## Verification
- Extended the app-server MCP resource integration test to read the same
hosted skill resource twice and verify that the remote server receives
one read.
- The same test verifies that catalog discovery and the selected skill's
main prompt are each fetched only once per thread.
## Why
App-server threads without a local executor need orchestrator-owned
skills from the hosted `codex_apps` MCP server. Threads with the local
executor already discover installed skills from the local filesystem.
After the orchestrator skill provider was enabled for every app-server
thread, local-executor threads also received the hosted skill catalog
and the `skills.list` and `skills.read` tools. This changed the existing
local behavior and could expose a second hosted copy of a skill that was
already installed locally.
## What changed
- Expose the thread's selected execution environments to extensions at
thread startup.
- Enable orchestrator skills only when the reserved local environment is
not selected.
- Apply that decision consistently to hosted skill catalog discovery,
explicit skill injection, and the `skills.list` and `skills.read` tools.
## Verification
- The existing no-executor app-server test continues to verify hosted
skill discovery, invocation, and child-resource reads.
- A new app-server test verifies that local-executor threads do not
receive hosted skill context or `skills.*` tools.
## Why
`ext/skills` currently depends on `codex-core` for two host concerns:
reading the concrete `Config` type and borrowing core-owned
model-context fragment types. That coupling prevents the extension from
being assembled independently above core and leaves context that belongs
to the skills feature owned by core.
This stacked PR introduces the host boundary needed for the broader
extension migration while intentionally preserving existing skills
behavior. It is stacked on #27404.
## What changed
- Adds a small public `SkillsExtensionConfig` view and makes skills
installation generic over the host config type.
- Requires the host to map its config into that view; app-server
supplies the current `Config` values.
- Moves the available-skills and selected-skill context fragment
implementations into `ext/skills`, preserving their roles, markers, and
rendered bytes.
- Removes the direct `codex-core` dependency from
`codex-skills-extension`.
- Keeps local discovery, invocation, side effects, and the
`codex-core-skills` compatibility types unchanged for later staged PRs.
## Behavior
This adds no capability and is intended to have no user-visible or
model-visible behavior change. The install API and ownership boundary
change internally; emitted skills context remains byte-for-byte
compatible.
## Validation
- Updates the skills extension integration coverage to use a host-owned
test config.
- Asserts the complete rendered catalog and selected-skill fragments,
including their roles and markers.
- `just bazel-lock-check`
- Rust tests and Clippy were not run locally per request; CI will run
them.
## Why
PR #27387 makes backend plugin skills discoverable and invocable without
an executor, but resources referenced by those skills still sit behind
the generic MCP resource surface. The model needs a skills-owned API
that preserves the provider authority and package boundary instead of
treating remote resources like local files.
This is stacked on #27387.
## What
- Adds one `skills` namespace with bounded `list` and `read` tools for
remote skill providers.
- Revalidates `authority + package` against the live remote catalog on
every read, then routes the opaque resource ID back through that
provider.
- Allows the backend provider to read canonical child `skill://`
resources while rejecting cross-package, non-canonical, query, fragment,
and traversal-shaped URIs.
- Caps each serialized tool result at 8 KB. Lists are paginated; reads
return an opaque continuation cursor.
- Marks the JSON output as external context so memory generation can
apply its normal suppression policy.
- Deliberately does not add `skills.search`; that waits for a bounded
plugin-service search contract.
## Tool contract
Pseudo-Python matching the wire shape:
```python
from typing import Literal, NotRequired, TypedDict
class RemoteSkillAuthority(TypedDict):
kind: Literal["remote"]
id: str # e.g. "codex_apps"
class RemoteSkill(TypedDict):
authority: RemoteSkillAuthority
package: str # opaque provider-owned package ID
name: str
description: str
main_resource: str # opaque provider-owned SKILL.md ID
class SkillsListParams(TypedDict):
cursor: NotRequired[str]
class SkillsListResult(TypedDict):
skills: list[RemoteSkill]
next_cursor: str | None
warnings: list[str]
truncated: bool
class SkillsReadParams(TypedDict):
authority: RemoteSkillAuthority # copied from skills.list
package: str # copied from skills.list
resource: str # provider-owned child resource ID
cursor: NotRequired[str] # copy next_cursor to continue
class SkillsReadResult(TypedDict):
resource: str
contents: str
next_cursor: str | None
truncated: bool
class Skills:
def list(self, params: SkillsListParams) -> SkillsListResult: ...
def read(self, params: SkillsReadParams) -> SkillsReadResult: ...
```
There is one namespace for all remote skills, not one tool or MCP server
per skill. No resource ID is converted into a filesystem path.
## Backend dependency
`/ps/mcp` must support direct reads of child resources such as
`skill://plugin_demo/deploy/references/deploy.md`. This PR implements
and tests the Codex side of that contract; production child reads remain
dependent on the corresponding plugin-service support. Search remains
out of scope until that service exposes a bounded search/resource API.
## Validation
- Added an app-server integration test covering `skills.list` followed
by `skills.read` with no executor.
- Ran `just fmt`.
- Ran `just bazel-lock-update` and `just bazel-lock-check`.
- Did not run Rust tests or Clippy locally, per request; CI will run
them.
## Summary
- cache the first remote skill catalog outcome per thread, including
failures
- preserve discovery errors as catalog warnings
- update the existing cache regression test to verify failed discovery
is attempted once
## Why
A failed or hanging `codex_apps` `resources/list` call could run once
while building initial context and immediately again while contributing
first-turn input. With the discovery timeout, an ordinary Apps turn
could wait up to 20 seconds before inference and retry again on later
turns even when no remote skill was mentioned.
Caching a warning-only empty catalog preserves graceful degradation
while preventing repeated synchronous discovery attempts.
## Testing
- `just fmt`
- Tests and Clippy not run per request; CI will validate the change.
## Why
#27198 made the extension-owned `codex_apps` MCP connection the hosted
plugin runtime, but its `mcp/skill` resources still bypassed the skills
extension. App-server could list and read those resources through
generic MCP APIs, but a thread with no selected environment did not
expose them in the model's skills catalog or load their `SKILL.md`
through `$skill`.
Hosted skills should stay remote while using the same typed catalog,
source authority, deduplication, bounded contextual catalog, and
selected-skill prompt injection as host and executor skills. They should
not be downloaded or exposed as ambient filesystem paths.
## What changed
- Add a session-scoped `McpResourceClient` over the replaceable MCP
connection manager so resource list/read calls follow startup and
refresh replacements.
- Add a `BackendSkillProvider` that pages `codex_apps` resources,
accepts bounded and validated `mcp/skill` entries, and reads a selected
skill's `SKILL.md` through the same MCP connection.
- Register the remote provider in app-server and include it in the
skills catalog even when a thread has no selected capability roots or
executor.
- Contribute hosted skill metadata through the bounded
`AvailableSkillsInstructions` developer-context path, exclude remote
entries from per-turn catalog injection, and classify `<skills>`
messages as contextual developer content so rollback can trim and
rebuild them correctly.
## Testing
- Extend the app-server MCP resource integration test with
`environments: []` to exercise two-page discovery, filter a
non-`mcp/skill` resource, verify the escaped developer catalog entry and
user-role `<skill>` fragment containing the fetched `SKILL.md`, and
preserve generic MCP resource reads.
- Add core event-mapping coverage that classifies `<skills>` developer
messages as contextual history.
## Why
Extension contributors are registered behind `dyn Trait` objects, so
native `async fn`/RPITIT methods would make these traits
non-object-safe. Spell out the boxed, `Send` future contract directly so
`extension-api` no longer needs `async-trait` while retaining the
existing runtime model.
## What changed
- add a shared `ExtensionFuture` alias and use it for asynchronous
contributor methods
- migrate production and test implementations to return `Box::pin(async
move { ... })`
- remove `async-trait` dependencies where they are no longer used,
keeping it dev-only where unrelated test executors still require it
## Behavior
No behavior change is intended. Contributor futures remain boxed,
`Send`, dynamically dispatched, and lazily executed; cancellation and
callback ordering stay unchanged.
## Testing
- `just test -p codex-extension-api` (11 passed)
- affected extension crates (64 passed)
- targeted `codex-core` contributor tests (14 passed)
- `just fmt`
- `just bazel-lock-update`
- `just bazel-lock-check`
A broad local `codex-core` run compiled successfully but encountered
unrelated sandbox and missing test-binary fixture failures; CI will run
the full checks.
Adam Perry @ OpenAI