666ef95f81
A goal round reporting complete or blocked used to conclude the turn at the tool result, so the model never spoke after the call and sessions ended on a bare update_goal card. The terminal update now defers one plugin-sourced <goal_complete>/<goal_blocked> instruction onto its result asking for a grounded closing message without further tool calls; the turn then ends through the ordinary no-tool-calls stop. Direct-human mutations stay uninstructed. Wording chosen by A/B sampling on deepseek-v4-pro; one extra request per goal lifecycle. New keyless ACP snapshot goal-wrapup drives the shipped app through create -> round one -> autonomous complete and pins the injection, the same-turn closing message, and the completed turn end.
4.1 KiB
Agent Note: Goal-round wrap-up message
Status: implemented
English | 中文
Problem
An autonomous goal round that reported update_goal complete or blocked concluded the physical turn at the tool result, so the model never spoke after the call. Sessions ended on a bare update_goal card, and internal testers read that as the agent stopping mid-sentence: the model's pre-call text routinely announces a report ("goal achieved, marking complete:") that never arrives, because the standard tool-use expectation is one more assistant message after a tool result and neither the goal-round prompt nor the tool description said the call was terminal. The hard stop came from the goal-tool decision, whose turn-stop clause this note supersedes.
Decision
A goal-round complete or blocked success no longer calls concludeTurn(). Instead the tool defers one wrap-up context onto its own result: a { kind: 'plugin', plugin: 'tool-goal' }-sourced user message carrying a <goal_complete>/<goal_blocked> instruction to write a grounded closing message to the user and call no more tools. The turn then ends through the agent loop's ordinary no-tool-calls stop, so no new loop primitive exists and steering semantics are untouched. Direct-human mutations remain uninstructed exactly as before. The cost is one additional model request per goal lifecycle, not per round.
The instruction wording was selected by A/B sampling on deepseek-v4-pro with a reconstructed goal-round transcript: a structured instruction (outcome, verification, artifacts, next steps) consistently beat a minimal "summarize" one on completeness; adding a session-grounding clause shifted unsupported detail from asserted fact to hedged suggestion; and the no-instruction control produced high-variance closings, including confidently fabricated file-level detail.
Scripting the keyless proof required one snapshot-harness addition: dsh-llm-replay resolves {{fromRequest:<regex>}} placeholders in scripted entries against the live request, because a static sidecar cannot know the randomly minted goal id the model must echo into update_goal.
Verification
tool-goal package tests pin the injected context (source, tag, objective, no-more-tools clause) and the absent concludesTurn for both terminal actions, plus the uninstructed direct-human pause and complete paths, at 100% file coverage. llm-replay unit tests pin the placeholder contract: last-match-wins capture, whole-match fallback, and loud failures for unmatched, invalid, and unterminated patterns. The new keyless ACP snapshot goal-wrapup drives the shipped application through create → round one → autonomous complete and asserts the plugin-sourced wrap-up injection, the same-turn closing assistant message, and the completed turn end in both the durable session log and the ACP stdout stream.
Alternatives considered
- Surface the completion text on the
update_goalUI card — rejected:completecarries no free text today, and adding asummaryargument would route a user-facing report through tool arguments while still cutting off the model's natural post-result message. - Keep
concludeTurn()and add a "one more text-only step" loop primitive — rejected: newagent-loopmachinery for behavior the ordinary stop already provides once nothing concludes the turn. - Instruct inside the tool result content — rejected: the goal tools' canonical output is compact JSON consumed programmatically; a prose instruction block inside it would mix the model-facing contract with the tool's replayable value.
Consequences
Every autonomous goal ends with a user-facing closing message instead of a bare tool card, at the cost of one model request per goal lifecycle. concludeTurn() keeps its loop semantics but loses its only first-party caller outside subagent structured output. Snapshot scenarios can now script values that only exist at run time via {{fromRequest:...}}, which unblocks keyless coverage of any echo-an-id tool flow, goal or otherwise.