fix(acp): accept owner control commands with rendered mention text

[madmada]

Summary

This is #6101 by @jhgaylor rebased onto current main. #6101 no longer merges cleanly: main has since reworked the same area of lib.rs (#6129 and its revert #6311) and the README (thread-scoped sessions, #6732). The fix and the test are his, and so is the commit authorship. I resolved the conflicts, added tests, updated two docs and ran it end to end against a local relay:

• crates/buzz-acp/src/lib.rs: kept control_command_content_matches and its test unchanged. Dropped workflow_attributed_author, which only came along as conflict context. It was removed from main in Revert "fix(acp): gate relay-signed workflow messages on their attributed author" #6311.
• New test owner_control_command_gate_accepts_rendered_mention_for_every_command (separate commit). buzz-acp: accept owner control commands with rendered mention text #6101's test covers the helper only. This one asserts on is_owner_control_command for !shutdown, !cancel and !rotate, so reverting the gate to an exact-content check fails it (checked). It also rejects a p tag for another agent and a missing p tag. A second commit adds the two other label shapes Desktop produces to the matcher table: @Name (<pubkey>) !rotate (Desktop qualifies a label this way when two selected mentions share a display name) and @Agent @Other Agent !rotate.
• docs/welcome-kickoff-silent-failures.md: buzz-acp: accept owner control commands with rendered mention text #6101 prefixed a "Fixed" note to the backlog item but left its diagnosis in present tense. Rewrote it as resolved and kept the still-current "!cancel is not a loop breaker" note (separate commit).
• desktop/src/features/agents/AGENTS.md: the "Channel-only runtime controls" section told contributors not to suggest @Agent !cancel, because the harness required the exact body. It now says the owner can reply in the target thread with @Agent !cancel or !rotate, and keeps the CLI example.
• crates/buzz-acp/README.md: replaced the "body must be exactly !rotate … an inline @Name does not match" paragraph with buzz-acp: accept owner control commands with rendered mention text #6101's mention-tolerant wording. Kept the CLI --reply-to … --mention … --content '!cancel' example from main, since a bare command with a separate p tag still matches and is the way to target a thread from the CLI.

Behaviour (unchanged from #6101): @Agent !rotate, !rotate @Agent and nostr:npub… !rotate from the owner, with the agent's p tag, are consumed as control commands. Because the harness does not know the agent's rendered (possibly multi-word) display name, any prefix starting with @ or nostr: counts as mention text, so @Agent please !rotate also matches. Content that does not start with a mention or continues past the command (please !rotate, !rotate now) is still forwarded as an ordinary message. The owner, p-tag and kind:9 checks are untouched. Under the thread session policy the command still resolves to the thread it is posted in.

@jhgaylor, if you would rather rebase #6101 yourself, say so and I'll close this.

Related issue

Fixes #6014, fixes #6051. Supersedes #6101.

Testing

• cargo test -p buzz-acp: lib 943 passed, integration 9 passed, 0 failed (run with no BUZZ_ACP_* variables in the environment, because two config::tests default assertions read them)

• cargo fmt -p buzz-acp -- --check, cargo clippy -p buzz-acp --all-targets -- -D warnings

• End to end: local relay (scripts/start-isolated-test-relay.sh stack), buzz-acp built from this branch and from main (77729ab), and a stub ACP agent that logs every session/new and session/prompt. Test events were published with the shape Desktop produces (content @Mock Agent !rotate, tags h, p, ["mention", <agent>, "agent-address"], and ["e", <root>, "", "reply"] for thread replies). They were not typed in the Desktop UI.

  Step	main	this branch
  Owner, idle: @Mock Agent !rotate, then a message	forwarded to the agent as a prompt, same session	consumed, not forwarded; next message opens a new session
  Owner, mid-turn: @Mock Agent !rotate	forwarded (steers the running turn)	session/cancel for the running turn; next message opens a new session
  Owner: @Mock Agent !cancel mid-turn	forwarded	session/cancel, the turn stops
  Non-owner: @Mock Agent !rotate	forwarded	forwarded, no rotate
  Owner: @Mock Agent !rotate now	forwarded	forwarded, no rotate
  Owner: bare !rotate with p tag	rotates	rotates
  thread policy: @Mock Agent !rotate replied in thread A	forwarded into thread A's session	thread A gets a new session; thread B keeps its session

  Unchanged from main, for reviewers: a turn cancelled by any control signal (including !cancel and Desktop's Stop) invalidates the session, so the next message after !cancel also opens a new session.

This PR was prepared with an AI coding agent. I reviewed the change and take responsibility for it.

🤖 Generated with Claude Code

[github-actions]

🔐 Codex Security Review

Status: review required for the current range.

The current range is 26ede6dfa2496993aa62ce5781d2112df4c2d009...c8a63de004ae85ea718e95206697e4e2b33441e6.
A new review must complete for this exact range. When manual authorization
is required, a Block organization member must comment exactly
@buzz-security-review c8a63de004ae85ea718e95206697e4e2b33441e6 to authorize a new review.
Any previous review applies only to its recorded range.