refactor(agents): extract OpenClaw from BenchFlow core

[kywch]

Closes #1090: #1090

Summary

• Moves OpenClaw shim, manifest, config, and tests to agents PR #72.
• Deletes the 973-line built-in shim and OpenClaw-specific production policy.
• Adds generic lazy manifest loading with local-checkout override support.
• Makes unknown bare agent IDs fail closed; explicit commands remain supported.
• Preserves catalog collisions, diagnostics, listing, and atomic publication.

Agents PR #72 must merge first.

Validation

• Ruff and ty pass.
• BenchFlow agent tests: 82 passed, 3 skipped.
• Paired manifest parity: 12 passed.
• Agents contract/OpenClaw tests: 98 passed.
• Full suite: 5,952 passed; 3 host-sensitive failures reproduced on main.
• Production Python: net -873 LOC.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-03T21:03:36.845766Z	c3e770f	Draft marked ready

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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[bingran-you]

Thermo-nuclear review of exact head c3e770f.

The central ownership change is a net simplification: the 973-line OpenClaw implementation and its feature-specific registry policy leave core, while generic manifest selection/loading remains in focused modules. I found and repaired one cross-feature regression: the new fail-closed launch boundary also tried to resolve the non-agent task-runtime as an ACP launch during default setup. Setup now models agentless modes once and skips both web policy and launch lookup; a regression test exercises real default rollout planes.

Validation on the exact tree: 5,911 tests passed with 87 skipped and 7 deselected; ty and Ruff passed; the paired agents catalog gate passed 28 tests; local list/show resolved OpenClaw from the external manifest. The supplied DeepSeek key was first validated with a direct HTTP 200 call. A complete Docker hello-world run then installed the manifest-owned OpenClaw shim, produced two proxied HTTP 200 model exchanges with provider-reported usage, emitted one completed write tool call, scored reward 1, and recorded no agent or verifier error.

Merge gate: the behavior and structure are acceptable, but agents #71 and #72 must land first, this PR must be refreshed against their immutable merged source, exact-head CI must finish green, and independent human approval is still required before squash merge.

[devin-ai-integration]

Devin Review found 2 potential issues.

2 flags not posted on this PR by your GitHub settings — view them in Devin Review. (Configure)

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c3e770fa05

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".

[bingran-you]

Exact-head follow-up for f7de9af: the local-catalog override now activates before a directly registered built-in can return, and expected remote/core parity overlaps no longer become permanent warnings. I reran the real default resolution/task-runtime focused gate: 21 passed with 3 skipped; ty and Ruff pass, and every exact-head GitHub test, parity, rollout-smoke, and fixture-scenario check is green. The separate reader/writer atomicity limitation is pre-existing on main and remains tracked by #1089; it should not be misrepresented as solved here. Merge still depends on agents #71, #72, and the new streaming fix #73 landing first, followed by immutable-source parity, exact-final-head real E2E, and independent human approval.

[bingran-you]

Exact-head follow-up for 24c5074.

The new generic discovery path correctly accepts an advertised ACP effort option when the registry does not declare one, while preserving registry override precedence. The generalized proxy-alias metadata path keeps existing mapped capabilities and otherwise derives context/output/vision/reasoning metadata from LiteLLM; focused static probes and the regression suite did not expose a contract regression. Exact-head validation passed 108 focused tests, ty check src/, changed-file Ruff check/format, and all current GitHub test, parity, rollout-smoke, and fixture-scenario checks.

A real Docker run then paired this head with agents #73 at 279aa61538a3d494bb6dfc1affa40c5696c001c7: BenchFlow discovered OpenClaw effort dynamically, delivered reasoning_effort=low, recorded a real DeepSeek deepseek-v4-flash HTTP 200 exchange with provider usage, completed one write tool call, scored reward 1, and recorded no agent or verifier error.

No new merge-blocking defect found. Merge still requires agents #71, #72, and #73 to land first so their immutable sources can be pinned, followed by final parity/CI and independent human approval.