feat(acp): honor platform_toolsets.acp config for ACP tool surface

[israellot]

Problem

ACP (editor integration) sessions hard-code the hermes-acp composite toolset in three places (acp_adapter/session.py agent construction, acp_adapter/server.py MCP-registration tool refresh and /tools command). As a result, platform_toolsets in config.yaml — the mechanism every other platform uses — is silently ignored for ACP, and users cannot trim the ACP tool surface (e.g. drop browser automation to save ~10 tool schemas of context per request in VS Code/Zed sessions).

Change

• New _acp_base_toolsets() helper in acp_adapter/session.py: reads platform_toolsets.acp from config; an explicit non-empty list wins, anything else (missing key, non-list, empty) falls back to the hermes-acp composite.
• _expand_acp_enabled_toolsets(None, ...) now resolves through the helper; the three hard-coded ["hermes-acp"] call sites pass None instead.
• Backward compatible: installs without the config key behave exactly as before.

Usage

platform_toolsets:
  acp: [web, terminal, file, code_execution, vision, skills, todo, memory, session_search, delegation]

Testing

• New: test_create_session_honors_platform_toolsets_acp (config list replaces composite, MCP toolsets still appended) and test_acp_base_toolsets_fallback_on_empty_or_missing (5 fallback shapes).
• pytest tests/acp tests/acp_adapter → 327 passed.
• Verified live: an ACP session under a profile with the config key resolves 15 tools with zero browser_*; a profile without it falls back to hermes-acp unchanged.

[tonydwb]

Code Review Summary

Verdict: Comment

This PR adds platform_toolsets.acp config support for ACP tool surface. Small, targeted addition.

Please verify:

• The config key is properly read and applied
• ACP tools work correctly with the new config

---

Reviewed by Hermes Agent

[tonydwb]

Code Review Summary

Verdict: Comment

Changes

ACP adapter: default toolsets changed from ["hermes-acp"] to None when agent has no enabled_toolsets. Allows the ACP session to inherit the agent's full toolset instead of being restricted to hermes-acp.

Assessment

• Correctness: Without this, the ACP session would always get hermes-acp even if the agent had a broader toolset. Now it inherits properly.
• No debug artifacts.

---

Reviewed by Hermes Agent

[teknium1]

Thanks for targeting a real ACP consistency gap: current main still hard-codes hermes-acp in acp_adapter/session.py:619-622.

Problems

• acp_adapter/session.py:140-147 reimplements raw platform_toolsets.acp lookup instead of using the shared resolver. That bypasses the global agent.disabled_toolsets subtraction in hermes_cli/tools_config.py:1904-1912. ACP construction does not pass disabled_toolsets independently, so a toolset disabled globally but listed under platform_toolsets.acp would be exposed.

Suggested changes

• Route ACP through _get_platform_tools(config, "acp"), preserving its hermes-acp fallback and shared explicit-selection/MCP semantics, then retain the ACP-client MCP append path.
• Add a test for an ACP list containing a globally disabled toolset.

Automated hermes-sweeper review.

[teknium1]

Please use the shared _get_platform_tools(config, "acp") resolver rather than read this mapping directly. It applies agent.disabled_toolsets last (hermes_cli/tools_config.py:1904-1912); ACP _make_agent() does not pass that list separately, so this path can re-enable a globally disabled toolset.

[israellot]

Fixed in fbb39d2 — _acp_base_toolsets() now routes explicit platform_toolsets.acp lists through the shared _get_platform_tools(config, "acp", include_default_mcp_servers=False) resolver, so the global agent.disabled_toolsets subtraction (applied last in hermes_cli/tools_config.py) is honored instead of bypassed. The hermes-acp fallback for missing/non-list/empty values is preserved, and config-level MCP servers stay excluded from the base resolution so the existing per-session append path in _expand_acp_enabled_toolsets remains the single place MCP toolsets are added.

Added the regression test you asked for: test_create_session_strips_globally_disabled_toolsets (tests/acp/test_session.py) asserts that a toolset listed under platform_toolsets.acp but also in agent.disabled_toolsets is excluded from the session's enabled_toolsets.

Verified with scripts/run_tests.sh tests/acp -q — 14 files, 309 tests passed, 0 failed.

[israellot]

Addressed the hermes-sweeper review finding in fbb39d2.

Finding → fix

• acp_adapter/session.py:140-147 reimplements the raw platform_toolsets.acp lookup, bypassing the global agent.disabled_toolsets subtraction → _acp_base_toolsets() now resolves explicit lists through the shared _get_platform_tools(config, "acp", include_default_mcp_servers=False) resolver, which applies agent.disabled_toolsets last. Behavior intentionally preserved from the original PR: missing/non-list/empty platform_toolsets.acp values still fall back to the hermes-acp composite, and the ACP-client MCP append path in _expand_acp_enabled_toolsets is untouched (MCP servers are excluded from base resolution so per-session appending remains the single source).

Tests

• New: TestPersistence::test_create_session_strips_globally_disabled_toolsets — config with web in both platform_toolsets.acp and agent.disabled_toolsets → session enabled_toolsets excludes web.
• Updated: test_create_session_honors_platform_toolsets_acp — expected order now matches the resolver's sorted output.

Verification

scripts/run_tests.sh tests/acp -q → 14 files, 309 tests passed, 0 failed.

[GottZ]

This was generated by AI during triage.

Summary

Two PRs address the ACP configuration gap. Both replace the hard-coded ACP toolset with platform_toolsets.acp, but #64045 covers all three ACP resolution paths and preserves shared disabled-toolset and MCP semantics, while #13167 changes only agent construction and platform registration.

Related pull requests

• #13167 [closed] duplicate — (+152/-2) — superseded duplicate: The closed PR replaces the hard-coded enabled_toolsets value during ACP agent construction and adds ACP to the platform registry, but it does not update the MCP-registration or /tools refresh paths and performs raw config normalization rather than shared resolver handling; it remains relevant as the narrower predecessor to #64045.
• #64045 related — (+109/-4) — preferred consolidation target: The diff routes explicit platform_toolsets.acp values through the shared resolver, updates agent construction, MCP registration, and /tools resolution, preserves per-session MCP appending, and tests global disabled-toolset subtraction. The contributor keep_open review on #64045 identified that disabled_toolsets could be bypassed; the current diff explicitly addresses that concern with _get_platform_tools(..., include_default_mcp_servers=False) and a regression test, although no separate recorded verify verdict is available.

Duplicates

#13167 and #64045 implement substantially the same core configuration support, but #64045 supersedes #13167 with broader call-site coverage and shared resolver semantics.

Suggested consolidation

Merge #64045 after normal maintainer validation — its current diff addresses the contributor review’s blocking resolver concern and covers the full ACP tool-surface lifecycle. Keep #13167 closed as superseded by #64045; neither PR has a recorded verify verdict.

Complex graph

flowchart LR
    classDef open fill:#dbeafe,stroke:#1d4ed8,color:#1e3a8a
    classDef merged fill:#dcfce7,stroke:#15803d,color:#14532d
    classDef closed fill:#e5e7eb,stroke:#6b7280,color:#1f2937
    classDef unverified fill:#f3f4f6,stroke:#9ca3af,color:#374151
    classDef best stroke-width:3px,stroke:#b45309
    classDef target stroke-width:3px,stroke:#4338ca
    subgraph Dup13167 ["PRs duplicating each other"]
        P13167["PR #13167 (closed)"]
        P64045["PR #64045 (open)"]
    end
    class P13167 closed
    class P64045 open
    class P64045 target
    click P13167 "https://github.com/NousResearch/hermes-agent/pull/13167"
    click P64045 "https://github.com/NousResearch/hermes-agent/pull/64045"

Loading

Graph: solid arrow = fixes / best fix, dashed arrow = partial or unverified (see edge label); boxed group = PRs duplicating each other; amber border = best fix; indigo border = target; gray node = closed or no verify verdict yet (state tag in the node label).

Cross-PR triage: Reviewed 2 pull requests and 0 issues in this complex. Each diff was read against this issue; Assessment working set: 17 kB of PR diffs, 6 kB of issue/PR text, 4 kB of discussion (5 comments), 1 verify verdict. verdicts reflect diff content, not PR titles. Part of an automated triage batch.