Bundled openai-docs Skill should respect explicit offline instructions
Nobody has claimed this yet.
- Dominant language
- Rust
- Stars
- 125k
- Forks
- 19.4k
- PR merge metrics
- PR metrics pending
Description
Environment and source
- Codex desktop:
26.901.20858(build7658) - Codex CLI:
0.153.0-alpha.5 - macOS:
26.6.2(25G83) - Source path:
codex-rs/skills/src/assets/samples/openai-docs/SKILL.md - Current source commit inspected:
6af345407d9c2a568da9d01b6c4b81a9e61495c0
The installed .system/openai-docs/SKILL.md is byte-for-byte identical to this
source path at commit a5082373f18119dc5d3eb993267c97f37880935d
(SHA-256 7cb8fa1b2a0c635b5c61ffe1da7b8594a7ea0fce5b71e8d523e2025d88b2a05e).
Current main changed the bundled model-migration reference but retains the
online-first rule described below. The deprecated openai/skills copies and
the Cursor repository copy are different bytes and are not used as provenance
for this report.
Reproduction
- Start a Codex task that explicitly prohibits network access and asks for
local implementation, analysis, or testing involving OpenAI documentation. - Invoke the bundled
openai-docsSkill. - Read its required first action and exception.
The Skill requires official online search and page retrieval before local or
repository inspection. Its only stated exception is a broad, manual-first Codex
setup or orientation request. It does not define a path for an explicit
offline-only request.
This is a contract-level reproduction. No additional model run was used to
claim measured runtime impact.
Expected behavior
When active instructions explicitly disallow network access, the Skill should:
- avoid search, fetch, and requests to relax the offline constraint;
- use supplied or local material with source identity and As-of;
- avoid presenting cached material as current;
- mark only a claim or action that needs current information as
UNKNOWNor
HOLD; and - continue independent authorized offline work.
When network access is allowed, the existing official-domain docs-first order
should remain unchanged.
Proposed minimal change
In SKILL.md:
- Scope the existing “First substantive action” to cases where network access
is allowed. - Add one explicit offline-boundary paragraph with the behavior above.
- Rename the current “Only exception” to a manual-first exception that applies
when network access is allowed. - Scope the official-domain search boundary to network-allowed work.
The prepared source diff changes one file: 5 insertions and 3 deletions.
Static validation cases
| Case | Expected result |
|---|---|
| Explicit offline + supplied API snapshot | Use the snapshot with source/As-of; no network request or permission question |
| Explicit offline + current price/latest model required | Mark that current claim UNKNOWN; do not present cached data as current |
| Offline implementation + one current-spec dependency | Hold only the dependent claim/action and continue independent offline work |
| Network-authorized current API question | Preserve official docs-first search and page retrieval |
| Unexpected network failure without an offline instruction | Do not misclassify it as an explicit offline request; existing failure handling still applies |
Validation
- Skill
quick_validate.py— PASS git diff --check— PASS- Manual static evaluation of the five cases above — consistent with the
proposed branch text just fmtandjust test -p codex-skills— not run because the isolated host
has nojustor Rust toolchain; no dependency was installed
Issue #24239 concerns a missing default Docs MCP configuration and is related to
the same Skill but does not cover explicit offline instructions.
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start with codex-rs/skills/src/assets/samples/openai-docs/SKILL.md and read the existing first-action, exception, and official-domain search instructions. Use quick_validate.py and the five static cases in the issue to check the wording. Done means explicit offline requests avoid network guidance while network-authorized work keeps the docs-first order.
Written by the indexing model from the issue text.
Assessment
- Domain
- ai-infra-agents, documentation
- Issue type
- Documentation
- Difficulty
- 2/5
- Estimated time
- 1-3 hours
- Activity status
- Active
- Clarity
- Clearly specified
- Newbie friendliness
- 88/100