[snimu]

Fixes RES-1321

What

Every machine-injected user-channel message now opens with one bracket grammar: first line [<kind>(: <qualifier>)( <address>)] (kebab-case kind), a blank line, then the payload. Machine data (message ids, endpoint ids, pids, schedules) lives in message.details, never in prose. Detection keys on customType, never text regexes. Refinement notices ([auto-/user-/self-refinement], from #2098) already followed this grammar and are untouched; they are the reference implementation.

Per-kind before/after

Kind	Before (first lines)	After
agent_message	5-line header (Agent-to-agent message received. / Source: / From: / To: / Message id:)	[agent-message from <relationship>:<name>] + blank + body
rlm_child_terminal_notice	RLM child <name> (<uuid>) completed without sending a reply. Last assistant text: ...	[child-exited: no-reply child:<name>] + optional Last assistant text: ...
	RLM child <name> (<uuid>) was cancelled: <reason>	[child-exited: cancelled child:<name>] + optional reason line
rlm_child_failure	RLM child <name> (<uuid>) failed: <error>	[child-failed child:<name>] + error payload
async_bash_completion	Shell message received. / Source: bash / Command completed (pid ..., exit code ...) / Command: ... / standing Inspect the saved BashHandle... hint	[bash-done pid:<pid> exit:<code>] + Command: <cmd>; hint DELETED (duplicated the system prompt)
heartbeat_prompt	raw job.prompt, no marker (masqueraded as the user)	[heartbeat: <schedule> run#<n>] + blank + prompt
goal_context	<goal_context>...</goal_context> XML wrapper	[goal: continuation|budget-limit|objective-updated] + prompt
autonomous_status	Autonomous mode: on. Continuations: ... (unmarked)	[autonomous-status: on|off] + usage payload
autonomous continuation	raw continuation prompt as a plain user message (unmarked)	[autonomous-continuation] + prompt; gate retries: [autonomous-continuation: gate-failed] + failure report (one owner: buildAutonomousGateFailureContinuation, also covers the headless injector)
prime-agent.update_complete	Prime Agent updated to v... (unmarked)	[update-complete] + notice
ipython_state	<ipython_state>...</ipython_state>	[python-state] + blank + body
ipython_state_restored	<ipython_state_restored>...</ipython_state_restored>	[python-state-restored] + blank + body (customType string kept for compat)
compactionSummary prefix	explanatory sentence + <summary>	[compaction-summary] first line; sentence and <summary> delimiters kept
branchSummary prefix	explanatory sentence + <summary>	[branch-summary] first line, same treatment
harness digest (#2098)	explanatory sentence + <harness_state>	[harness-digest] first line; one owner (HARNESS_DIGEST_PREFIX) covers both the standalone message and the compaction-summary rider

No relay path exists for agent messages today, so no relaying grandchild:<y> form was invented.

All remaining built-in customTypes are either filtered out of LLM context by convertToLlm (session_slash_command, session_slash_command_result, compaction_outcome, refinement_outcome) or entry-only (thread_goal_state, rlm_max_depth_state, ipython_sent_agent_message, prime-agent.refinement). A table-driven pin walks every constructor-backed synthetic kind and asserts a grammar-conforming first line. Status/continuation/update notices regenerate rather than being re-parsed from transcripts, so no accept-old shims were added for them.

Detection migration + compat

• clearQueuedAgentMessages now matches queued actions by the delivered custom message's customType (isAgentSessionMessage), with isAgentSessionMessagePrompt retained only as the legacy-text fallback.
• parseAgentSessionMessagePromptId / isAgentSessionMessagePrompt keep parsing the old header format forever (accept-both, emit-new); persisted old-format transcripts still resolve their message ids and are pinned in tests.
• New-format prompts intentionally carry no id in text; id resolution uses customMessage.details.id.
• TUI renderers for every kind read customType/details, not prose (verified; no renderer changes needed beyond test pins).
• Goal prose now says "Python REPL" instead of "ipython" (tool-name references keep ipython; that rename is a separate later PR).

Tests

One pin per variant format string, merged into existing suites: bus/fire-and-forget (agent-message grammar + legacy parse + sanitization), 4482 (heartbeat marked, provider passthrough), async-bash-completion + 2068 (bash-done grammar, hint gone), recursion + 617 (child-exited/child-failed variants), goal suite (goal kinds), 4530 (python-state-restored), 4531 (customType-keyed clearing of new-format queued messages), session-command-messages (per-kind convertToLlm passthrough + compaction/branch bracket lines), compaction suite (harness-digest ordering). Net: +244/-194 lines.

Validated in a Prime sandbox: coding-agent suites for messages/agent-messages/kernel bridges/heartbeat/goal/compaction/interactive rendering/regressions (only the documented environmental failures: kernel-* bridges, recursion asyncio-kernel test, 4428, 4603 — all pass locally with a real kernel) plus root npm run check.

---

Note

Medium Risk
Broad changes to LLM-visible prompt text and queue-clearing logic, but behavior is guarded by customType, legacy parsers, and extensive test pins.

Overview
Machine-injected messages that reach the model as user-channel text now share a single bracket header pattern ([<kind>…], blank line, body) instead of multiline headers, XML wrappers, or raw unmarked prompts.

Agent-to-agent prompts shrink from the old Agent-to-agent message received. block to [agent-message from <relationship>:<name>], with ids and routing metadata kept in details. Heartbeats, async bash completions ([bash-done …], BashHandle hint removed), RLM child notices, goal context, Python kernel state, autonomous status/continuations, post-update notices, and compaction/branch/harness digest prefixes all move to the same grammar. New sanitizeMessageHeaderValue strips characters that could break or spoof headers.

Detection shifts toward customType (e.g. clearQueuedAgentMessages matches the queued custom message, not only legacy text); parseAgentSessionMessagePromptId remains for old persisted transcripts only.

^{Reviewed by Cursor Bugbot for commit 2480f8f. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Unify synthetic user-channel messages under bracket grammar in coding-agent

• Replaces multiline and XML-like wrappers with compact bracket headers (e.g., [bash-done], [python-state]) for machine-injected messages across coding-agent.
• Adds sanitizeMessageHeaderValue in messages.ts to strip whitespace, commas, colons, and closing brackets from header metadata.
• Updates message builders for bash completions, RLM child events, heartbeats, goals, compaction summaries, and autonomous continuations to use the new grammar.
• Behavioral Change: AgentSession.clearQueuedAgentMessages in agent-session.ts now identifies current agent messages by custom message type instead of legacy text predicates, though legacy text parsing is retained for older persisted transcripts.

^{Macroscope summarized 2480f8f.}

[cursor]

Cursor Bugbot has reviewed your changes and found 3 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want reviews to match your repository better? Bugbot Learning can learn team-specific rules from PR activity. A team admin can enable Learning in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit 5945f93. Configure here.}