feat: add GLM and Kimi document seats

[alanshurafa]

Summary

Adds GLM-5.3-Flash and Kimi K3 as first-class document-pipeline seats, giving co-evolution two free cross-vendor reviewers without changing the Codex/Opus execution and verification path.

The new adapters fail closed around the trust boundaries that matter:

• glm reuses the Claude CLI in safe mode with no tools, a child-only Z.AI environment, provider-variable scrubbing, and a loud WSL rejection.
• kimi uses the official Kimi Code login, runs in a disposable no-tools project, extracts raw Markdown from stream-json with jq, rejects partial/error output, and guards native Windows command-line limits.
• Role resolution, prerequisite checks, model metadata, unknown-agent handling, and the pinned single-process GLM-to-Claude leak regression now cover all four seats.

flowchart TB
    Entry[co-evolve CLI or MCP] --> Guard[Typed seat prerequisites]
    Guard --> GLM[glm seat]
    Guard --> Kimi[kimi seat]
    GLM --> SafeClaude[Claude safe mode / no tools]
    SafeClaude --> ZAI[Z.AI Anthropic-compatible endpoint]
    Kimi --> Temp[Disposable no-tools project]
    Temp --> Stream[Kimi Code stream-json]
    Stream --> Markdown[jq final-assistant extraction]

Loading

What changed

• Extends the shell dispatcher and MCP schema/runtime with glm and kimi.
• Records concrete seat models and rejects unsupported Kimi effort assignments honestly.
• Adds Windows, PowerShell, Git Bash, and macOS launcher/setup guidance with isolated CLAUDE_CONFIG_DIR handling.
• Adds a roster-independent seat-selection procedure to the existing Model Routing documentation.
• Adds hermetic coverage for credential scoping, provider scrubbing, no-tools behavior, partial output, WSL routing, launcher isolation, Kimi authentication, and opt-in live GLM validation.

Verification

• tests/run-all.sh: 36/36 suites passed under native Git Bash.
• MCP: 9/9 tests passed.
• GLM launcher isolation: 2/2 scenarios passed.
• Kimi seat: 8/8 scenarios passed.
• Document seats: 8/8 scenarios passed, including the single-run GLM-to-Claude leak check.
• Shell syntax, PowerShell parsing, git diff --check, and frozen-fixture checks passed.
• Live API tests remained disabled; no credential or model request was used during automated verification.

Manual follow-up

Live smoke tests remain intentionally staged until the human-owned credentials exist: take in ZAI_API_KEY, complete Kimi Code login on each machine, and run the Mac launcher/1Password setup from docs/agent-seats.md.

---

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e1757eed76

ℹ️ About Codex in GitHub

Your team has set up Codex to 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 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Filter Claude aliases out of GLM overrides

When a GLM role receives a Claude-shaped override such as --composer-model opus, best, or claude-*, this guard accepts it and the next line resolves aliases such as opus to claude-opus-4-8. invoke_glm then sends that model ID to the Z.AI endpoint, causing the pass to fail instead of falling back to GLM_MODEL_BASE; this is especially likely with a mixed glm,claude pair where the composer override is intended for the Claude seat. Include the Claude-only patterns in the GLM mismatch guard, as the Codex and Kimi arms already do.

Useful? React with 👍 / 👎.