fix: bridge tool-calls in copilot-acp adapter

[MestreY0d4-Uninter]

Summary

This PR fixes a major behavioral gap in the copilot-acp adapter where Hermes would often respond with intent text ("vou executar...") but make zero tool calls.

What changed

• Updated agent/copilot_acp_client.py to:
  • include available tool schemas (tools) and tool_choice hints in the ACP prompt bridge
  • remove the strict anti-tool JSON instruction and replace it with explicit tool-call emission guidance
  • parse structured tool calls returned in text (<tool_call>{...}</tool_call> and JSON-shaped fallbacks)
  • map parsed calls back into Hermes/OpenAI-compatible tool_calls with proper finish_reason="tool_calls"
  • add conservative heuristic fallback for ACP prose-only responses (initially terminal requests; plus explicit vision request URL fallback path)

Why

In copilot-acp mode, Hermes could claim action but fail to execute tools. This breaks core tool-use guarantees.

Validation

Executed automated ACP battery against enabled CLI toolsets with --provider copilot-acp and claude-sonnet-4.6.

• Passed: 10 / 11
• Failing: vision

Notes on vision failure

Failure appears to be provider/runtime capability mismatch in ACP path (tool listed at UI level but not actually exposed/usable in runtime path), not parser regression.

Evidence artifacts (local run):

• /tmp/acp_toolsets_progress.log
• /tmp/acp_toolsets_report.json

Impact

• Restores practical tool execution for copilot-acp path across terminal/file/code_execution/delegation/skills/memory/session_search/todo/web toolsets.
• Keeps behavior bounded via conservative fallback rules.

Follow-ups (recommended)

• Add focused tests for ACP tool-call text parsing and fallback synthesis in tests/.
• Investigate ACP vision runtime availability contract.

[robert-hoffmann]

ah nice, this was annoying, especially since copilot is one of the cheapest SOTA options

[MestreY0d4-Uninter]

ah nice, this was annoying, especially since copilot is one of the cheapest SOTA options
I've made some more progress with Github Copilot. I'll publish it soon.

[robert-hoffmann]

@MestreY0d4-Uninter yeah the annoying thing was tool calls in acp were broken, and the fine grained pat token : is broken too (i think)

https://github.com/github/copilot-cli

https://github.com/settings/personal-access-tokens/new

[MestreY0d4-Uninter]

Thanks — I kept this PR intentionally narrow around the ACP tool-call bridging fix.

The additional ACP work I’ve been doing locally is already split into separate follow-up branches (session/context persistence, integration, and delegation reasoning/defaults), and I’ll submit those separately so this PR stays a straightforward bugfix.

[MestreY0d4-Uninter]

I’ve now pushed and opened the follow-up work as draft PRs so it doesn’t get lost to upstream drift while this one is pending.

They are intentionally marked as drafts and each one explicitly depends on #4536:

• [DRAFT] fix(agent): preserve copilot-acp session context across requests #5335 — ACP session/context preservation
• [DRAFT] fix(agent): route copilot-acp through non-streaming request flow #5336 — run_agent.py ACP non-streaming integration
• [DRAFT] fix(tools): support delegation reasoning overrides for subagents #5337 — delegation reasoning overrides for subagents

All three were rebased onto current main and revalidated before opening.

[robert-hoffmann]

Thanks for sharing, like that i can integrate and test while waiting for them to update their code, which does seem kinda slow

[robert-hoffmann]

I sent them a request on discord to see if they can accelerate this fix

Btw the merge status seems to have gone red 🥺

[teknium1]

Merged via PR #5460. Your commit was cherry-picked onto current main with your authorship preserved in git log. Stripped the heuristic fallback system and Portuguese-specific patterns (adapter should be language-agnostic), and fixed a duplicate tool call extraction bug. Thanks for the contribution!