feat(cli): warn when the agent's run config differs from the request - #324
Conversation
The applier suppressed user-visible warnings for exactly the four controls users most need told about — model, reasoning effort, Fast, Plan — on Codex and Claude, because it keyed the notice off a REJECTION and those calls fail routinely when re-sent per turn. But a rejection is the wrong signal in both directions. Codex ACCEPTS `fast-mode` on a model with no fast speed tier and then omits the option from the state it publishes: nothing throws, the turn runs at normal speed, and the user is told nothing. Meanwhile a rejected value that was already effective changed nothing and was reported. Key the notice off divergence instead: after applying the turn's config, compare each requested selection against the state the agent itself publishes, and warn only where they disagree. An `on`/`off` select and a boolean toggle are the same choice, so neither shape alone is a divergence. Where the agent published no config options at all — or for a sensitive id, which the runtime state deliberately omits — the failed call remains the only signal, so it is still used there. Rejections keep going to debug diagnostics unchanged, and `agent_warning` notices are already deduplicated by message per session, so a stuck divergence costs one notice rather than one per turn. Model: claude-opus-5[1m] Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 484d31ea8e
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| selection.source.kind === 'mode' | ||
| ? runtimeConfigPatch.modeId | ||
| : selection.source.kind === 'model' | ||
| ? runtimeConfigPatch.modelId | ||
| : effectiveConfigOptionValues[selection.source.configId]; |
There was a problem hiding this comment.
When a config-option-backed request resolves but the response retains a different mode or omits the requested model, AgentClient has already stored that authoritative snapshot, but confirmedLegacyModeId overwrites its mode and confirmedLegacyModelId fills its missing model before this comparison. The comparison therefore sees the requested values rather than the agent's answer and emits no agent_warning, silently defeating the new divergence warning for permission modes and models; apply the fallback only for actual legacy protocol paths or compare against the unmodified snapshot.
AGENTS.md reference: apps/cli/AGENTS.md:L277-L285
Useful? React with 👍 / 👎.
1 participant
Stacked on #316 — review that one first; the base flips to
mainonce it merges.Problem
applyAcpSessionRunConfigdecided a user-visibleagent_warningfrom a REJECTION, then suppressed it for exactly the four controls users most need told about — model, reasoning effort, Fast, Plan — on Codex and Claude (suppressKnownRunConfigWarnings && isKnownRunConfigOption(...)). The suppression existed because those calls fail routinely when the turn config is re-sent every turn.A rejection is the wrong signal, in both directions:
applyFastModeChangesetsfastModeEnabledunconditionally — it ACCEPTSfast-modeon a model with no fast speed tier. Nothing throws;createSessionConfigOptionsthen omits the option andresolveFastServiceTierreturnsnull, so the turn runs at normal speed and the user is told nothing. Even without the suppression, no rejection ever existed to report.Fix
Key the notice off divergence: after applying the turn's config, compare each requested selection against the state the agent itself publishes (
runtimeConfigPatch, already computed fromagentClient.getConfigOptions()), and warn only where they disagree.on/offselect and a boolean toggle express the same choice, so neither shape alone counts as a divergence.config.modeId/modelIdoutranks a duplicate config-option entry; losing that precedence contest is not the agent disagreeing, so only the winner is judged.rejectedSelections(debug diagnostics) is unchanged.The notice text now says the agent did not apply part of the config rather than rejected it, which is what the new signal actually means.
Noise is bounded: the composer re-sends the state the agent reported, so an ordinary turn diverges nowhere, and
recordAgentWarningalready deduplicates by message per session — a stuck divergence costs one notice, not one per turn.Invariant change
apps/cli/AGENTS.mddocumented the old behavior ("Codex/Claude mismatches for model, reasoning effort, Fast, or Plan are not promoted to visibleagent_warningnotices"). That paragraph is rewritten to state the divergence rule, why a rejection is not it, and to say plainly that a per-agent suppression list must not come back. The Claude Fablefast=falseno-op exception is preserved — it is skipped before dispatch, so it is judged for neither.Tests
apps/cli/src/session/acp-session-config-applier.test.ts— the two tests that encoded the suppression are replaced by ones for the new rule: a selection accepted but dropped from the agent's state (the Codex Fast shape, with emptyrejectedSelections), a rejected selection that was already effective (silent), and the select/boolean equivalence. The four tests covering redaction, non-publishing agents, and other agent types are unchanged and still pass.pnpm --filter lody test(2467),typecheck,lintall pass.🤖 Generated with Claude Code