feat: add context-aware Auto effort and speed with Codex 0.162.1

[nocell]

Wirebot can now choose reasoning effort and processing speed for every user turn from the full active conversation. Research and bug investigations use the strongest supported effort at Standard speed; ordinary work uses High/Extra high; simple questions use Low/Medium. Explicit requests for speed select Ultrafast or Fast when the model catalog makes them available.

• Upgrade bundled Codex from 0.159.0 to 0.162.1, regenerate app-server bindings, and handle partial answers.
• Add default-on Auto mode using an isolated, subscription-backed GPT-6.1 Sol Low/Fast classifier, with local routing on failure. Preserve the selected answering model and existing approval settings.
• Expose Auto/manual controls in the web app and in-chat settings. Slack /wirebot config always opens native buttons; /wirebot web opens the browser app.
• Refresh full Slack thread context and keep DM context within the active conversation after new.
• Use model-catalog capabilities for supported effort and speed choices and reset incompatible options when switching models.

Auto speed is scoped to one turn. Manual and scheduled turns restore default effort instead of inheriting an earlier Auto choice. Cancellation covers classifier startup, and Slack preserves arrival order while allowing stop/new commands through.

Validation on 95dc63476e88cd4d7f52424e4efb07599b355e9d:

• Bun 1.4.0: 61 tests passed (394 assertions); TypeScript and Biome checks passed.
• Production compilation passed; the compiled binary reports 0.4.0.
• Upstream CI passed both verification and Docker jobs on the final commit.
• Codex 0.162.1 app-server was exercised with an isolated loopback inference fixture to verify nullable settings, Standard speed and completion payloads. The observed contracts are documented in docs/auto-mode.md.
• Browser fixtures verified both concurrent-edit conflict detection and preservation of a successful Save when an earlier Auto-toggle response arrives late.
• Live subscription inference, Slack delivery, and account-specific Ultrafast entitlement have not been exercised.

Triad Ultra reviewed 4581d4b5adaa13dba41562ea6daa903aae7ab207 → 83ab3af575d9de94d74064b4838b6c82e7b53e34. All seven accepted findings were fixed in 95dc634: attachment ordering, historical skill mentions, metadata lookup failures, concurrent settings responses, queued acknowledgements, classifier eligibility checks, and voice-request context. No findings required a human decision.

Coverage was five complete reviews (Claude Opus 5.5, GPT-6 Astra, Cursor Grok 4.7 Fast, GLM-5.3, GLM-5.3-Flash), with GPT-6 Astra reducing; Kimi K3 was partial because its file-reading tool could not paginate the diff. The isolated fixer could not bind localhost for HTTP tests (EPERM); the complete suite, including those tests, passed independently in the working checkout. This is degraded review coverage, not an unqualified clean review of the final commit.

The fork release v0.4.0 publishes ghcr.io/nocell/wirebot:0.4.0 for linux/amd64 and linux/arm64. Its manifest digest is sha256:b73d4397704de601048f778f4b539c33f687f22add4c907ee00f8a5c192e54c9. The published ARM64 image was pulled by digest and verified to report Wirebot 0.4.0 and Codex 0.162.1. No deployment was performed.