Best possible solution:

Keep the unified PR open, make one-shot thinking: "off" count as an explicit thinking setting, add focused regression coverage, and attach redacted real behavior proof before merge.

Do we have a high-confidence way to reproduce the issue?

Yes for the review finding by source inspection: PR head resolves the one-shot thinking override, but the explicit-thinking guard excludes that override. I did not run tests or live flows under the read-only review constraint.

Is this the best way to solve the issue?

No, not yet. The unified surface is plausible, but the runtime override path must preserve /think off semantics and the PR needs real behavior proof before merge.

Full review comments:

• [P2] Treat one-shot thinking off as explicit — src/auto-reply/reply/get-reply-directives.ts:543
thinkingLevelOverride now feeds resolvedThinkLevel, but this explicitness check still ignores it. When chat.send, gateway model run --thinking off, or Talk consult config sends a one-shot off override with no session/agent thinking default, resolvedThinkLevelWithDefault becomes off, then the block below sees thinking as not explicit and can restore the model default reasoning level. Include the normalized override in thinkingExplicitlySet so one-shot off behaves like inline /think off.
  Confidence: 0.87

Overall correctness: patch is incorrect
Overall confidence: 0.82

What I checked:

• related PR consolidation: The PR body says it replaces the closed PRs Add realtime consult thinking and fast mode overrides #72237, feat(cli): add thinking override to infer model run #73642, and Allow chat commands to clear thinking and fast defaults #79351 with one combined run-control patch. (009b338b70fe)
• current main behavior: Current main accepts chat.send thinking and injects it into CommandBody, but it does not pass thinkingLevelOverride or fastModeOverride through reply options. (src/gateway/server-methods/chat.ts:1904, a855414bc665)
• PR override path: PR head resolves opts.thinkingLevelOverride into resolvedThinkLevel before directive/session precedence. (src/auto-reply/reply/get-reply-directives.ts:430, 009b338b70fe)
• off override gap: The explicit-thinking check still only considers directives, session state, and agent defaults, so a one-shot thinkingLevelOverride: "off" is not treated like inline /think off. (src/auto-reply/reply/get-reply-directives.ts:541, 009b338b70fe)
• dependency contract checked: @mariozechner/pi-ai@0.73.1 exports completeSimple and SimpleStreamOptions.reasoning?: ThinkingLevel, where ThinkingLevel is minimal|low|medium|high|xhigh; the PR's local mapping for off, adaptive, and max is necessary for that API. (package.json:1703, a855414bc665)
• proof status: The PR body lists formatter, targeted tests, and pnpm build, but no screenshot, terminal output, logs, recording, or linked artifact showing the changed CLI/Talk/voice behavior in a real setup. (009b338b70fe)

Likely related people:

• Peter Steinberger: Current-main blame for the central reply-directive, chat.send, CLI model-run, and voice-call consult paths points to commit 0071f67 in this checkout; the PR timeline also shows Peter/steipete force-pushed the finishing commit on this branch. (role: recent maintainer and adjacent owner; confidence: medium; commits: 0071f67fe6ea, 009b338b70fe; files: src/auto-reply/reply/get-reply-directives.ts, src/gateway/server-methods/chat.ts, src/cli/capability-cli.ts)

Remaining risk / open question:

• The PR has no real behavior proof for the new CLI, Talk, or voice consult paths.
• The correctness finding is source-derived; tests and live scenarios were not run because this was a read-only review.

Codex review notes: model gpt-5.5, reasoning high; reviewed against a855414bc665.