singularity/singularity-forge
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

feat(sf): teach executor about the escalation field on sf_task_complete

Browse source 
The escalation feature was invisible to agents — the prompt didn't say it
existed, so agents made silent assumptions instead of surfacing genuine
tradeoffs. Now, when phases.mid_execution_escalation is on, execute-task
includes a guidance block showing the escalation payload shape and noting
auto-mode auto-accepts the recommendation by default. When the feature is
off the field is silently dropped, so the guidance is omitted entirely to
avoid misleading the agent.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Mikael Hugo 2026-05-02 21:41:38 +02:00
parent 3895ae2cd3
commit 08859624f8
2 changed files with 32 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

30
src/resources/extensions/sf/auto-prompts.ts
View file

@ -2489,6 +2489,35 @@ export async function buildExecuteTaskPrompt(
}
})();
// ADR-011 P2: when the feature is enabled, teach the executor that it can
// surface non-obvious choices via the `escalation` field on sf_task_complete
// rather than silently picking. Auto-mode auto-accepts the recommendation
// (see phases.escalation_auto_accept), so this is low-cost overhead — but
// it produces an audit trail and a hard constraint for downstream tasks.
// When the feature is off, the field is silently dropped, so we omit the
// guidance entirely to avoid misleading the agent.
const escalationGuidance =
prefs?.preferences?.phases?.mid_execution_escalation === true
? [
"**Surfacing non-obvious choices (optional).** If you hit a decision with material tradeoffs that downstream tasks should respect (e.g. data-loss vs. block-progress, two valid library choices with different long-term cost), include an `escalation` payload in your `sf_task_complete` call:",
"",
"```json",
'"escalation": {',
' "question": "Short, concrete question",',
' "options": [',
' { "id": "a", "label": "Option A", "tradeoffs": "what it costs" },',
' { "id": "b", "label": "Option B", "tradeoffs": "what it costs" }',
" ],",
' "recommendation": "a",',
' "recommendationRationale": "why a wins on this evidence",',
' "continueWithDefault": true',
"}",
"```",
"",
"Provide 2–4 options with concrete tradeoffs. The recommendation must reference one of the option ids. Auto-mode accepts your recommendation and carries it forward as a hard constraint for downstream tasks; the user can review or override later via `/sf escalate`. Set `continueWithDefault: false` only when the choice is severe enough that the loop should pause for human review even in auto-mode (rare).",
].join("\n")
: "";
return loadPrompt("execute-task", {
memoriesSection,
overridesSection,
@ -2512,6 +2541,7 @@ export async function buildExecuteTaskPrompt(
inlinedTemplates,
verificationBudget,
gatesToClose,
escalationGuidance,
skillActivation: buildSkillActivationBlock({
base,
milestoneId: mid,

2
src/resources/extensions/sf/prompts/execute-task.md
View file

@ -95,6 +95,8 @@ All work stays in your working directory: `{{workingDirectory}}`.
**Autonomous execution:** Do not call `ask_user_questions` or `secure_env_collect`. You are running in auto-mode — there is no human available to answer questions. Make reasonable assumptions and document them in the task summary. If a decision genuinely requires human input, note it in the summary and proceed with the best available option.
{{escalationGuidance}}
**You MUST call `sf_task_complete` before finishing. Do not manually write `{{taskSummaryPath}}`.**
When done, say: "Task {{taskId}} complete."