refactor(layout): group ACP agents under acp/ (acp/ + ai-sdk/ + omnigent/)

[Yiminnn]

What

Group the repo's ACP agents under a single top-level acp/ directory, so the layout becomes three agent buckets + infra:

• acp/ — the 11 ACP agents: claude-agent-acp, codex-acp, gemini, mimo, mimo-acp, opencode, openclaw, openhands, pi-acp, mini-swe-acp, mini-swe-code
• ai-sdk/ — the Vercel AI SDK harness group (unchanged)
• omnigent/ — the Databricks Omnigent session-factory agent (unchanged)
• infra stays at the top: contract/, docs/, skills/, .github/, plus root README.md / CONTRIBUTING.md / LICENSE / config files

All agent dirs move via git mv (renames, history preserved): 169 files renamed, 10 files edited for path refs.

Why this is backward-compatible

The benchflow loader already discovers agents recursively: benchflow.agents.manifest.discover_manifests walks root.rglob("manifest.toml") and skips only _DISCOVERY_SKIP_DIRS (build/vendor/VCS dirs) — acp is not in that set. Nesting agents one level deeper under acp/ therefore does not hide them; the loader's own docstring already anticipates nested families (e.g. ai-sdk/acp/manifest.toml alongside acp/<agent>/manifest.toml).

Verified empirically against load_agents_from_dir(): the discovered agent-name set is identical before and after the move, and every manifest's content hash (sha256) is byte-identical (pure git mv, no content change).

• before: ['claude-agent-acp', 'codex-acp', 'gemini', 'mimo', 'openclaw', 'opencode', 'openhands', 'pi-acp']
• after: ['claude-agent-acp', 'codex-acp', 'gemini', 'mimo', 'openclaw', 'opencode', 'openhands', 'pi-acp']

Path refs updated

Every filesystem path reference to a moved dir is repointed <dir>/ -> acp/<dir>/. Agent names in prose and model ids are deliberately left untouched (e.g. "use BenchFlow's native codex-acp" stays — that's an agent name, not a path).

• .github/workflows/ — lint.yaml (working-directory:), and test-mimo-acp.yaml / test-mini-swe-acp.yaml / test-mini-swe-code.yaml (paths: filters + working-directory:)
• contract/ — manifest.py docstring example; test_manifest_launcher_freshness.py _MANIFEST constant (now acp/pi-acp/manifest.toml); and test_manifest.py's repo-manifest glob made recursive-bounded (*/manifest.toml + */*/manifest.toml) so it keeps validating the moved manifests instead of silently pytest.skip-ing under the new depth
• docs — README.md (agent-table links, quickstart cd, pip #subdirectory=, repository-layout block, license table), CONTRIBUTING.md (dev-setup cd + the mini-swe-code/... file paths), docs/adaptation.md (the mini-swe-acp worked-example link)

Tests

• contract/ suite: 20 passed — incl. test_repo_manifests_validate now finding all 9 acp/<name>/manifest.toml (no skip) and the pi-acp launcher-freshness tests reading acp/pi-acp/manifest.toml
• ruff check + ruff format --check clean on contract/
• loader discovery invariant: baseline agent-name set == post-move set; per-manifest content hashes identical

Follow-ups (out of scope here)

• Move the loader to an explicit scan-by-3-paths strategy (acp/, ai-sdk/, omnigent/) and add a deprecation path for the old flat layout.
• Pre-existing, not introduced by this PR: acp/mimo/ and acp/mimo-acp/ both declare name = "mimo", so load_agents_from_dir raises a duplicate-name error over the whole tree — identical behavior before and after this move. Worth de-duping separately.

🤖 Generated with Claude Code

[greptile-apps]

Too many files changed for review. (179 files found, 100 file limit)