## Why
Executor skill discovery runs before the remote skills catalog is
available. For a remote environment, each `ExecutorFileSystem` operation
becomes an exec-server RPC.
Previously, every discovered `SKILL.md` independently resolved its
plugin namespace by walking its ancestors and probing both supported
manifest locations. In the common `plugin/skills/<skill>/SKILL.md`
layout, that repeats 8 RPCs per skill even though every skill under the
plugin root uses the same namespace. These lookups happen while skills
are parsed, so their cost grows linearly with the skill count and adds
directly to first-turn latency.
A selected capability root can also contain standalone skills, multiple
sibling plugins, nested plugins, or symlinked directories. The
optimization therefore needs to retain the nearest-ancestor namespace
for each skill rather than assuming the selected root represents exactly
one plugin.
## What changed
- record plugin-root candidates from directory entries already returned
during skill discovery
- prune candidates that are not ancestors of any discovered `SKILL.md`
before reading manifests
- resolve each relevant plugin root once, with one fallback lookup per
canonical traversal root for symlinked directories
- select the nearest cached plugin namespace for each discovered skill
- avoid namespace lookup entirely when the root contains no skills
No additional directory traversal is required. Namespace work now scales
with the number of plugin roots that contain discovered skills, rather
than the total number of skills or unrelated sibling plugins. Standalone
and nested-plugin names keep their previous behavior.
## Benchmarks
I used a temporary counting `ExecutorFileSystem` around the real local
filesystem. Each filesystem operation was counted as one remote RPC and
given 1 ms of injected latency. Each variant ran three times; times
below are medians.
### One plugin with 100 skills
| Operation | Before | After | Delta |
| --- | ---: | ---: | ---: |
| `get_metadata` | 1,002 | 303 | -699 |
| `read_file` | 200 | 101 | -99 |
| `read_directory` | 102 | 102 | 0 |
| **Total filesystem RPCs** | **1,304** | **506** | **-798 (-61.2%)** |
| **Median load time** | **2.890 s** | **0.997 s** | **2.90× faster** |
The namespace-specific work drops from 800 RPCs to 2 in this layout.
### Multiple plugins under one selected root
These runs compare the correct pre-optimization implementation with the
final nearest-plugin-root cache. The total plugin skill count stays at
100 while the number of plugin roots changes.
| Layout | Before RPCs | After RPCs | Reduction | Before | After |
Speedup |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| 2 plugins × 50 skills | 1,312 | 530 | 59.6% | 1,819 ms | 711 ms |
2.56× |
| 10 plugins × 10 skills | 1,344 | 578 | 57.0% | 1,850 ms | 778 ms |
2.38× |
| 50 plugins × 2 skills | 1,504 | 818 | 45.6% | 2,094 ms | 1,086 ms |
1.93× |
| 10 plugins × 10 skills + 10 standalone skills | 1,596 | 630 | 60.5% |
2,209 ms | 860 ms | 2.57× |
The remaining cost grows with the number of relevant plugin manifests.
Each relevant manifest is read once instead of once per skill, while
sibling plugins with no discovered skills are not read. Absolute latency
savings depend on the executor's real RPC latency.
## Tests
- `just test -p codex-core-skills` (109 passed across the library and
integration-test binaries)
- one integration test covers standalone, outer-plugin, nested-plugin,
and unused sibling-plugin layouts, and asserts the exact set of
manifests read
## Why
Config layer provenance describes how effective configuration was
assembled, so it belongs with the config loader rather than in
app-server's serialized API types.
## What changed
- Moved `ConfigLayerSource`, `ConfigLayerMetadata`, and `ConfigLayer`
ownership into `codex-config`.
- Kept app-server's wire payloads unchanged and added explicit
conversions at the app boundary.
- Removed lower-level app-server-protocol dependencies from config
consumers.
## Stack
This is PR 3 of 6, stacked on [PR
#29721](https://github.com/openai/codex/pull/29721). Review only the
delta from `codex/split-auth-domain-types`. Next: [PR
#29723](https://github.com/openai/codex/pull/29723).
## Validation
- `codex-config` coverage passed.
- App-server config-manager and config RPC coverage passed.
## Why
After #28918, selected skill roots are `PathUri`, but the executor skill
provider still converts them to the app-server host's `AbsolutePathBuf`.
A foreign Windows root therefore cannot be discovered by a Unix host,
and the inverse has the same problem.
This PR keeps executor skill discovery and reads on the filesystem that
owns the selected root while reusing the existing skill rules.
## What changed
- Generalize the existing skill traversal to operate on `PathUri`
through `ExecutorFileSystem`, preserving its depth, directory, symlink,
and sibling-metadata concurrency behavior.
- Add a small environment skill loader that reuses the shared discovery,
frontmatter validation, dependency parsing, product policy, and
prompt-visibility rules.
- Keep the environment id and entrypoint `PathUri` in the skill catalog,
then route `skills.read` back through the same environment filesystem.
- Preserve the executor's path convention when deriving catalog handles,
including literal backslashes in POSIX filenames.
- Resolve plugin namespaces from nearby manifests through URI-native
filesystem reads.
- Cover foreign Windows roots, executor-owned reads, namespaces,
metadata, policy, and path identity.
```text
selected root (PathUri)
|
v
shared discovery over ExecutorFileSystem
|
v
environment-bound catalog entry --skills.read--> same ExecutorFileSystem
```
No second filesystem abstraction or duplicate traversal implementation
is introduced.
## Stack
1. #29614 — add lexical `PathUri` containment.
2. #29620 — share URI-native manifest path resolution.
3. #28918 — keep selected plugin roots and resources URI-native.
4. **This PR** — load executor skills without host path conversion.
5. #29628 — resolve executor MCP working directories without host path
conversion.
## Summary
This switches skill discovery to the simpler same-connection scalar
request shape.
After reading a skills directory, discovery now starts the existing
`fs/getMetadata` calls for all visible entries in that directory before
awaiting the results. There is no JSON-RPC batch frame and no new
filesystem API; remote filesystems use the existing request-id
multiplexing on the same exec-server connection.
This is the scoped alternative to the batch-frame approach in #29074 /
#29075.
## What changed
- Collect visible directory entries before processing them.
- Run their existing `fs.get_metadata(...)` calls with `join_all`.
- Process the results in the original directory order, so skill
discovery behavior stays the same.
## Benchmarks
Fresh local benchmark against generated skill trees over a real
exec-server remote filesystem. The benchmark calls the actual
`load_skills_from_roots` path, so this includes directory reads,
metadata stats, `SKILL.md` reads, and parsing.
Times are p50 milliseconds from 5 samples after 1 warmup, using warmed
runs.
| Scenario | Legacy `main` | Batch frame stack (#29074 / #29075) |
Same-connection scalar stack |
| --- | ---: | ---: | ---: |
| 100 flat skills | 377.4 | 389.0 | 378.6 |
| 500 flat skills | 1983.2 | 1856.6 | 1757.5 |
Takeaway: for the actual skill discovery path, same-connection scalar is
tied with legacy at 100 skills and best at 500 skills. The batch-frame
stack does not show enough win here to justify the extra protocol/API
surface.
Benchmark command:
- `just test -p codex-exec-server benchmark_remote_skill_discovery
--run-ignored ignored-only --no-capture`
Checked locally with:
- `just test -p codex-core-skills`
- `just bazel-lock-update`
- `just bazel-lock-check`
## Why
Skill descriptions are used in model-visible lists: the default
available-skills catalog that supports implicit selection, and the
on-demand `skills.list` tool response used to discover orchestrator
skills. A single overlong description should not consume a
disproportionate share of either list.
Enforcing the 1024-character limit while loading or migrating skills is
the wrong boundary: it rejects otherwise-valid skills and discards
metadata that non-model consumers and full skill reads may need. Skill
metadata and `SKILL.md` content should remain intact; the cap belongs at
model-visible list rendering boundaries.
## What changed
- Preserve full `description` and `metadata.short-description` values
when loading skills.
- Preserve full external-agent command descriptions during
`source-command-*` migration instead of skipping commands solely because
their descriptions exceed 1024 characters.
- Preserve full normalized orchestrator descriptions in the underlying
skills catalog.
- Cap each description at 1024 Unicode characters when rendering the
default available-skills context in `codex-core-skills` and
`codex-skills-extension`.
- Apply the same cap when serializing descriptions in the model-visible
`skills.list` response.
- Render truncated descriptions as 1021 original characters plus `...`.
- Leave explicit `$skill` injection, `skills.read`, underlying metadata,
and on-disk `SKILL.md` files unchanged and full-fidelity.
## Implicit skill selection
Codex injects a bounded catalog containing each implicitly allowed
skill's name, description, and source locator, together with
instructions to use a skill when the task clearly matches its
description. The model makes that semantic choice; after selecting a
skill, it reads the full `SKILL.md` from its filesystem or provider
resource. Explicit `$skill` mentions remain a separate path that injects
the full skill instructions. For orchestrator skills, `skills.list`
provides bounded discovery metadata before `skills.read` returns the
full selected resource.
## Test plan
- `just test -p codex-core-skills`
- `just test -p codex-skills-extension`
- `just test -p codex-external-agent-migration`
The focused regressions verify that overlong metadata is preserved at
load and migration boundaries while default available-skills rendering
and `skills.list` output produce the 1021-character prefix plus `...`.
## Summary
- Preserve raw plugin skill-root snapshots in the matching loaded-plugin
cache entry, keyed by the effective plugin root identity including
namespace.
- Pass those snapshots through `SkillsLoadInput` as an optional preload,
so session startup reuses plugin parsing while ordinary skill loads pass
`None`.
- Keep plugin skill loading cohesive: the existing loaders accept the
optional snapshots directly, and uncached or marketplace-detail paths do
not create a cache.
## Why
Plugin discovery already parses plugin skills to determine available
capabilities. Cold session startup then scanned and parsed the same
roots again while building the skills snapshot.
This solves the same duplicate-work problem as #28623 while keeping
ownership narrow: `PluginsManager` creates and owns
`PluginSkillSnapshots` only for its loaded-plugin cache entry;
`SkillsService` consumes an optional clone. Entry replacement or
clearing naturally drops the snapshots, with no separate generation,
capacity policy, or watcher coupling.
## Validation
- `cargo clippy -p codex-core-skills --all-targets -- -D warnings`
- `just test -p codex-core-plugins
skills_service_reuses_skills_parsed_during_plugin_load`
- `just test -p codex-core-skills
namespaces_plugin_skills_using_provided_namespace`
- `just fmt`
## What changed
- retain the parsed plugin manifest namespace on loaded plugins
- carry that namespace through `PluginSkillRoot` and `SkillRoot`
- use the provided namespace when qualifying plugin skill names
- include the namespace in the skills cache key
## Why
Plugin loading has already parsed `plugin.json`, but skill parsing
currently walks every `SKILL.md` ancestor and probes/reads the manifest
again to reconstruct the same namespace. Passing the parsed namespace
removes those repeated filesystem calls, which are particularly costly
on remote filesystems.
Context:
https://openai.slack.com/archives/C0ARA9GF5D4/p1781639496496439?thread_ts=1781202444.891669&cid=C0ARA9GF5D4
## Impact
Plugin skill names remain unchanged. A regression test uses a
deliberately different on-disk manifest name to verify that plugin roots
use the provided parsed namespace.
## Validation
- `just test -p codex-core-skills -p codex-core-plugins -p codex-plugin
-p codex-utils-plugins` (352 passed)
- `just fix -p codex-core-skills -p codex-core-plugins -p codex-plugin
-p codex-utils-plugins`
- `just fmt`
## Why
The community marketplace audit found many skill frontmatter parse
failures where values were intended as prose, but were not valid YAML.
Common examples include unquoted scalar values with `: `, such as
`description: Build for AWS: ECS` or `argument-hint: <duration: e.g.
7d>`, and flow-looking values such as `tags: [next,@supabase/ssr]`.
`serde_yaml` does not expose a permissive mode for this. The parser
fails before unknown frontmatter fields can be ignored, so a
compatibility repair has to happen before retrying YAML parsing.
## What changed
Skill frontmatter loading still uses `serde_yaml` as the primary parser.
If that parse fails, the loader performs a line-oriented repair of
scalar frontmatter field values, then retries parsing.
The fallback now:
- applies to any frontmatter mapping field, not just `description` /
`short-description`
- quotes unquoted scalar values that contain a YAML colon separator such
as `: `
- quotes invalid flow-looking scalar values that start with `[`, `{`,
`@`, or backtick
- preserves already quoted values
- skips `|` / `>` block scalar bodies so multiline descriptions are not
rewritten
- returns the original YAML error if the repaired frontmatter still
cannot parse
## Examples
This previously failed because the second `: ` was parsed as YAML
structure:
```yaml
description: AWS deployment patterns: ECS Fargate, Lambda, and S3
```
The fallback now parses it as if it had been written explicitly as:
```yaml
description: 'AWS deployment patterns: ECS Fargate, Lambda, and S3'
```
The same repair now applies to ignored frontmatter fields that still
need to be valid YAML for the parser to get through the document:
```yaml
argument-hint: <duration: e.g. 7d, 2w>
tags: [next,@supabase/ssr]
```
Valid YAML multiline descriptions continue to work through normal
parsing without repair:
```yaml
description: |-
Build for AWS: ECS
and Lambda
```
## Validation
- Added loader coverage for unquoted `description` values containing `:
`.
- Added loader coverage for unquoted `metadata.short-description` values
containing `: ` and an apostrophe.
- Added loader coverage for unrecognized frontmatter fields that need
quoting, including `argument-hint` and `tags`.
- Added block-scalar coverage to ensure multiline description bodies are
preserved while other fields are repaired.
- `just test -p codex-core-skills` (106 passed)
- `just fix -p codex-core-skills`
## Why
`PathUri::from_abs_path` can fail for absolute paths that do not have a
normal `file:` URI representation, forcing filesystem call sites to
handle a conversion error even though the original path can be preserved
losslessly.
## What
Make `from_abs_path` infallible and migrate its callers. Unrepresentable
paths use `file:///%00/bad/path/<base64>`, encoding Unix bytes or
Windows UTF-16LE; `to_abs_path` validates and decodes that fallback. The
leading encoded null reserves a namespace that cannot collide with a
real Unix or Windows path, and fallback URIs remain opaque to lexical
path operations.
## Validation
Added path-URI coverage for Unix null and non-UTF-8 paths, Windows
device/verbatim and non-Unicode paths, serialization, malformed
fallbacks, opaque lexical operations, invalid native payloads, and
literal `/bad/path` collision resistance.
## Why
We're moving exec-server to use PathUri for its internal path
representations.
## What
Move `ExecutorFileSystem` APIs to use `PathUri` instead of
`AbsolutePathBuf`. Future changes will convert higher-level parts of
exec-server.
## Why
CCA is moving toward a split runtime where the orchestrator may not have
a filesystem, while executors can expose preinstalled plugins and
skills. A thread therefore needs to select capabilities without asking
app-server or core to interpret executor-owned paths through the
orchestrator's filesystem.
The longer-term model is broader than executor skills:
- A plugin is a bundle of skills, MCP servers, connectors/apps, and
hooks.
- A plugin root can be local, executor-owned, or hosted by a backend.
- Components inside one plugin can use different access and execution
mechanisms. A skill may be read from a filesystem or through backend
tools; an HTTP MCP server can run without an executor; a stdio MCP
server or hook needs an execution environment.
- Core should carry generic extension initialization data. The extension
that owns a component should discover it, expose it to the model, and
invoke it through the appropriate runtime.
This PR establishes that architecture through one complete vertical:
selecting a root on an executor, discovering the skills beneath it,
exposing those skills to the model, and reading an explicitly invoked
`SKILL.md` through the same executor.
## Contract
`thread/start` gains an experimental `selectedCapabilityRoots` field:
```json
{
"selectedCapabilityRoots": [
{
"id": "deploy-plugin@1",
"location": {
"type": "environment",
"environmentId": "workspace",
"path": "/opt/codex/plugins/deploy"
}
}
]
}
```
The root is intentionally not classified as a "plugin" or "skill" in the
API. It can point at a standalone skill, a directory containing several
skills, or a plugin containing skills and other components. This PR only
teaches the skills extension how to consume it; later extensions can
resolve MCP, connector, and hook components from the same selection.
The platform-supplied `id` is stable selection identity. The location
says which runtime owns the root and gives that runtime an opaque path.
App-server does not inspect or canonicalize the path.
## What changed
### Generic thread extension initialization
App-server converts selected roots into `ExtensionDataInit`. Core
carries that generic initialization value until the final thread ID is
known, then creates thread-scoped `ExtensionData` before lifecycle
contributors run.
This keeps `Session` and core independent of the capability-selection
contract. The initialization value is consumed during construction; it
is not retained as another long-lived `Session` field.
### Executor-backed skills
The skills extension now owns an `ExecutorSkillProvider` that:
- resolves the selected environment through `EnvironmentManager`
- discovers, canonicalizes, and reads skills through that environment's
`ExecutorFileSystem`
- contributes the bounded selected-skill catalog as stable developer
context
- reads an explicitly invoked skill body through the authority that
listed it
- warns when an environment or root is unavailable
- never falls back to the orchestrator filesystem for an executor-owned
root
Skill catalog and instruction fragments have hard byte bounds, which
also bound them below the 10K-token per-item context limit. If a
selected executor skill has the same name as a legacy local skill, the
executor selection owns that invocation and the local body is not
injected a second time.
Existing local and bundled skill loading remains in place. Omitting
`selectedCapabilityRoots` therefore preserves current local-only
behavior.
## Current semantics
- Only environment-owned locations are represented in this first
contract.
- Roots are resolved by the destination extension, not by app-server or
core.
- An unavailable executor or invalid root produces a warning and no
capabilities from that root; it does not trigger a local-filesystem
fallback.
- Selection applies to a newly started active thread.
- MCP servers, connectors, and hooks beneath a selected plugin root are
not activated yet.
- Selection is not yet persisted or inherited across resume, fork, or
subagent creation. Existing local capabilities continue to behave as
they do today in those flows.
## Planned vertical follow-ups
1. **Hosted HTTP MCP:** add an extension-backed HTTP MCP source that
works without an executor, then replace the special-purpose MCP plugins
loader with that implementation.
2. **Executor MCP:** register and execute stdio MCP servers through the
environment that owns the selected plugin root.
3. **Backend skills:** add a hosted skill source whose catalog and
bodies are accessed through extension tools rather than a filesystem.
4. **Connectors and hooks:** activate those components through their
owning extensions, using the same selected-root boundary and
component-specific runtime.
5. **Durable selection:** define the desired-selection lifecycle,
persist it, and make resume, fork, and subagent inheritance explicit
rather than accidental.
6. **Local convergence:** incrementally route existing local plugin,
skill, and MCP loading through the same extension model while preserving
current local behavior.
Each follow-up remains reviewable as an end-to-end capability. The
platform selects roots, generic thread extension data carries the
selection, and the owning extension resolves and operates its component.
## Verification
Coverage added for:
- app-server end-to-end discovery and explicit invocation of a skill
inside an executor-selected plugin root
- exclusive invocation when a selected executor skill collides with a
local skill name
- executor filesystem authority for discovery, canonicalization, and
reads
- thread extension initialization before lifecycle contributors run
- stable executor catalog context, explicit invocation, context
rebuilding, hidden skills, and preserved host/remote catalog behavior
Targeted protocol, core-skills, skills-extension, core lifecycle, and
app-server executor-skill tests were run during development.
## Summary
- Validate skill base name length before plugin namespacing.
- Bound the composed `plugin:skill` qualified name to 128 characters.
- Keep plugin skill runtime names in the existing `plugin:skill` form.
- Add regression tests for the max qualified-name boundary and rejection
path.
## Root Cause
Plugin skills are represented as `plugin_name:skill_name`, but the
loader previously applied the 64-character skill name limit after adding
the plugin namespace. Moving that check to the base name fixes valid
plugin skills with longer namespaces, and the separate 128-character
qualified-name limit keeps model-visible skill names bounded.
## Validation
- `just fmt`
- `just test -p codex-core-skills plugin_skill_name_length_limit`
- `git diff --check`
## Summary
PR 3 of 5 in the cloud-managed config client stack.
Adds enterprise-managed cloud config as a first-class config layer
source. The layer metadata is preserved through config loading,
diagnostics, debug output, hook attribution, and app-server protocol
surfaces.
## Details
- Enterprise-managed config becomes a normal config layer source with
backend-supplied `id` and display `name` attached for provenance.
- These layers are designed to behave like non-file managed config: they
can surface syntax/type diagnostics by layer name even though there is
no physical config file.
- Relative path settings are resolved from a stored config base so
cloud-delivered config remains consistent with existing MDM-delivered
config semantics.
- Hook attribution distinguishes config-delivered hooks from
requirements-delivered hooks via `HookSource::CloudManagedConfig`.
- This remains pull-based and snapshot-oriented; the PR adds layer
identity/diagnostics, not dynamic reload behavior.
## Validation
Validated through the targeted stack checks after rebasing onto current
`main`:
- Rust crate tests for
config/hooks/cloud-config/backend-client/app-server-protocol
- Filtered `codex-core` and `codex-app-server` `cloud_config_bundle`
tests
- Python generated-file contract test
- `cargo shear --deny-warnings`
- Targeted `argument-comment-lint` for config/hooks
## Why
Skill metadata accepted a `permissions` block and stored the result on
`SkillMetadata`, but that data was never consumed by runtime behavior.
Leaving the dead parsing path in place makes it look like skills can
widen or otherwise influence execution permissions when, in practice,
declared skill permissions are ignored.
This change removes that misleading surface area so the skill metadata
model matches what the system actually uses.
## What changed
- removed `permission_profile` and `managed_network_override` from
`core-skills::SkillMetadata`
- stopped parsing `permissions` from skill metadata in
`core-skills/src/loader.rs`
- deleted the loader tests that only exercised the removed permissions
parsing path
- cleaned up dependent `SkillMetadata` constructors in tests and TUI
code that were only carrying `None` for those fields
## Testing
- `cargo test -p codex-core-skills`
- `cargo test -p codex-tui
submission_prefers_selected_duplicate_skill_path`
- `just argument-comment-lint`
## Summary
- move skill loading and management into codex-core-skills
- leave codex-core with the thin integration layer and shared wiring
## Testing
- CI
---------
Co-authored-by: Codex <noreply@openai.com>
jifandGitHub
Adam Perry @ OpenAIandGitHub
charlesgong-openaiandGitHub
xl-openaiandGitHub
Matthew ZengandGitHub
joeflorencio-openaiandGitHub
alexsong-oaiandGitHub
pakrym-oaiandGitHub
Michael BolinandGitHub
