Evidence, not issue-count theater
What Capability Doctor knows—and what it does not
The live diagnosis rules can use official source context, but upstream reports never become confirmed incidents or confidence upgrades automatically.
Review state
Collection is not confirmation
- CandidateMetadata matched a bounded capability-failure query. It has no root-cause or checkpoint claim.
- Source verifiedThe source exists and supports a narrowly written method claim. It still does not prove a user's root cause.
- Human reviewedAn environment, ordered checkpoint boundary, source claim, fix, verification, and limitations were checked.
- PublishedThe record passed freshness and indexability gates. P9 currently publishes zero incident verdicts.
Behavior level
A document cannot prove runtime compatibility
-
DECLARED
A first-party source states intended behavior or a supported interface.
-
PARSED
A bounded definition or response passed structural parsing.
-
CONNECTED
A process or endpoint completed a connection or protocol handshake.
-
DISCOVERED
The host discovered or listed the capability.
-
EXPOSED
The capability appeared in the active model session.
-
ACTIVATED
A representative task selected or triggered the capability.
-
EXECUTED
A bounded representative action completed or failed at execution.
-
VERIFIED
A fix was followed by a recorded recheck in the same scoped environment.
Live source register
Bounded official context used by P9
B · DECLARED
current
Official Codex MCP documentation
OpenAI documents that Codex MCP configuration lives in config.toml, can be user- or trusted-project-scoped, is shared by the supported Codex clients, and can be inspected through MCP settings, /mcp, or codex mcp list. OAuth servers can be authenticated with codex mcp login.
This documents intended configuration and inspection paths; it does not prove that a submitted server connected or exposed tools.
B · DECLARED
current
Official Codex skill-building documentation
OpenAI documents the repository, user, admin, and system locations from which Codex loads skills, and advises restarting Codex when a changed skill does not appear.
A discovered or parsed skill is not automatically proven to activate for a representative task.
B · DECLARED
current
Official Codex configuration basics
OpenAI documents user-level and project-level Codex configuration layers and how to open config.toml from supported surfaces.
The existence of a configuration file does not prove the failing surface read the intended layer.
B · DECLARED
current
Official Codex configuration reference
OpenAI provides the current searchable reference for config.toml and requirements.toml fields.
The reference can validate a field name or value shape but cannot establish which process or surface loaded it.
B · DECLARED
current
Official Codex command rules documentation
OpenAI documents rules for controlling commands outside the Codex sandbox and marks the rules feature as experimental.
A rule or approval boundary is not the same as authentication or capability discovery.
B · DECLARED
current
Official Codex Windows sandbox guidance
OpenAI documents Windows sandbox permission warnings and advises restarting Codex or rerunning sandbox setup after correcting the listed permissions.
This guidance applies to the documented Windows sandbox condition, not every permission or execution failure.
B · DECLARED
current
Official Codex developer command reference
OpenAI lists codex mcp and codex plugin as stable management commands and marks codex mcp-server as deprecated in favor of the Codex app server.
The command maturity table does not prove that an installed capability is available in a particular active session.
Publication stop condition
Until a candidate has an ordered first failing checkpoint, scoped environment, source-supported claim, bounded fix, verification, and human review, it remains excluded from public Incident and Compatibility pages.
Diagnose your evidence