fix(daemon): lead the collaboration guidance with a tool-precedence rule (#800)

[Poytr1]

What

Adds one leading bullet to the standing collaboration guidance (every session carries it): AgentConnect's MCP tools are the only channel that reaches other agents and humans; runtime built-ins with similar names (e.g. Claude Code's bare SendMessage) do not reach AgentConnect and anything sent through them is lost — never use them for messaging, reporting back, or collaboration. Cost: ~330 chars ≈ ~80 standing tokens per session. Pinned by a contract test that also asserts the bullet LEADS the section (the collision fires exactly when the model is choosing a messaging tool).

Why

Issue #800, measured in the tool-surface A/B (PR #791, 2026-08-09): on the Claude Code runtime a child session's tool list contains both mcp__agentconnect__sendMessage and the runtime's own built-in SendMessage (agent-teams messaging). In 3 of 6 parent-report trials the model's first attempt went to the built-in — a literal name match for the report-back instruction — and 2 of 6 trials lost the parent's answer entirely, silently.

Measured before/after (parent-session scenario, same harness as #791; criteria fixed before looking)

Metric	Before (6 trials)	After (10 trials, 5 per surface arm)
Trials with any built-in SendMessage attempt	3/6	0/10 (verified in scored attempts AND raw ACP events)
Answers lost to the built-in	2/6	0/10
Total success (parent actually woken with the answer)	4/6	10/10
First-attempt correct form	0/6	4/10

The collision is eliminated in this sample and no answer was lost. First-attempt behavior improved but is not perfect — the child still sometimes wraps the correct parent reply in extra (legal) postless calls — so #800's deeper fixes (suppress the built-in tool for daemon-owned sessions; reply-hint the exact pre-filled call; inferred reply) remain worthwhile hardening.

🤖 Generated with Claude Code

[agentconnect-md-test]

Approved. The new precedence rule is placed before the existing collaboration guidance, so it applies through the shared standing-context path for fresh/resumed and Claude/non-Claude sessions. It names both the exact AgentConnect MCP tool and the conflicting bare runtime tool, and the added regression test pins the required content and ordering. git diff --check passes, and I found no blocking behavior issue.

Two non-blocking follow-ups:

• “the ONLY channel that reaches other agents and humans here” is broader than the intended tool-precedence rule and is literally at odds with the later instruction that an ordinary reply reaches people and agents in the current thread. Consider tightening it to “Among messaging tools, only AgentConnect’s MCP tools…” while retaining the leading emphasis.
• The newly added explanatory comments span multiple lines, while this repository’s AGENTS.md asks for one-line comments. They can be condensed without changing the test contract.

I could not execute the focused Vitest case because this isolated checkout has no installed dependencies and Corepack could not reach the npm registry (DNS failure), even after retrying with a writable cache. Static inspection confirms the test exercises the intended prompt block and ordering.

_{sent by review-bot (Codex · gpt-5.6-sol) · open in session}