Agent answer triggering

How a comment the nebuko bot itself publishes — an agent’s answer — can re-enter the engine and advance the workflow, while every other bot comment stays inert.

How it works

The bot prefixes every comment it posts with the universal terminal marker (<!-- bot-workflow-terminal -->). Agent answers additionally carry an agent-answer marker (<!-- bot-agent-<role> -->). Historically the engine’s routing guard (run_loop._is_bot_trigger) dropped any bot-classified comment, so the bot’s own agent answer could never trigger the next step.

The fix is a narrow routing exemption in _is_bot_trigger only:

Body carries	Example	Routing
Universal terminal marker only	<!-- bot-workflow-terminal -->	Inert (terminal notice)
Agent-answer marker, nothing inert	<!-- bot-agent-dev -->	Routable (answer part 1)
Agent-answer + continuation marker	<!-- bot-agent-dev --> + <!-- bot-answer-continuation -->	Inert (answer tail)
Agent-answer + AGENT_RUN_CONFIG	—	Inert (stale sentinel echo)
Continuation marker only	<!-- bot-answer-continuation -->	Inert

The predicate lives in src/nebuko/engine/lib/resolve_config/features/parse_trigger/parse_trigger.py as is_routable_agent_answer(), and run_loop._is_bot_trigger composes it:

bot trigger  =  is_bot_comment(body)  and  not is_routable_agent_answer(body)

Only the routing choke point honors the exemption. is_bot_comment() itself and all gate-decision consumers (resolve_config.py) are unchanged, so a bot comment still can never carry a gate decision — the gates stay bot-blind.

How to use

For a bot-published agent answer to advance the workflow:

1. Part 1 of the answer must carry the agent-answer marker (<!-- bot-agent-<role> -->).
2. Part 1 must not contain the <!-- bot-answer-continuation --> token or the literal AGENT_RUN_CONFIG.
3. The CI ingestion workflow (issue-comment GHA pre-filter) must admit bot-authored comments; it was aligned with this change.

If any of these fail, the comment is treated as an ordinary bot notice and the engine does not run.

Loop-safety invariants

• Gates remain bot-blind: is_bot_comment / is_machine_comment semantics are unchanged, so no bot comment can ever carry a gate decision.
• Terminal notices and continuation parts can never route.
• Not protected: a human echoing the bot-agent- marker can trigger a step — the same power as any /nebuko run command. Gate decisions are still rejected for such comments.

Troubleshooting

Symptom: “agent comment published by nebuko bot doesn’t trigger next step.”

Check each layer in order:

1. CI filter — does the issue-comment workflow’s pre-filter drop bot-authored comments before the engine runs? It must admit bodies carrying the agent-answer marker.
2. _is_bot_trigger — confirm the body contains <!-- bot-agent- and neither <!-- bot-answer-continuation --> nor AGENT_RUN_CONFIG; then the guard must return False (not a bot trigger).
3. Routing log — look for the [phase:guards] bot-marker check lines; a warning there means detection crashed and the fallback classified the comment as “not a bot” (posting-safe default).