BuzzbeeSCD/openchamber
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: distill any oversized goal objective, not just plans

Browse Source 
The distillation moved from PlanView into setSessionGoal — the shared
choke point for every goal source (composer, fork dialog, plans) — and
into the scheduled-tasks goal creation server-side. Any objective over
5000 chars is distilled into completion criteria by the small model; on a
transient failure a head+tail excerpt keeps the intent (top) and the
acceptance criteria (bottom) with a trim marker, plus a toast. The
working agent always received the full prompt in chat — only the auditor
is bound by the limit.
This commit is contained in:
Bohdan Triapitsyn
2026-07-12 16:34:16 +03:00
parent 19b2a3d0d3
commit 8d956f5f9c
14 changed files with 99 additions and 51 deletions
Download Patch File Download Diff File Expand all files Collapse all files
packages/ui/src/lib/smallModel.ts
+14 -13
View File
@@ -56,24 +56,25 @@ export async function summarizeSelectionForNotes(text: string, sessionId?: strin
}
}
// Plan-goal objectives are capped at 5000 chars for the auditor. Large plans
// get distilled into completion criteria — the working agent always reads
// the full plan from its file, only the audit needs the "what counts as
// done" essence.
const PLAN_GOAL_SYSTEM_PROMPT = [
'You distill an implementation plan into the COMPLETION CRITERIA a progress auditor will judge against.',
// Goal objectives are capped at 5000 chars for the auditor. Oversized ones
// (huge plans, pasted specs, long assignments) get distilled into completion
// criteria — the working agent received the full prompt in chat anyway,
// only the audit needs the "what counts as done" essence.
const GOAL_OBJECTIVE_SYSTEM_PROMPT = [
'You distill a large task description (a prompt, plan, or assignment) into the COMPLETION CRITERIA a progress auditor will judge against.',
'Return ONLY the criteria text — no preamble, no headers, no markdown fences.',
'Capture: the end goals, what must exist and work when the plan is fully implemented, and how each major phase is verified. Omit implementation steps and how-to details.',
'Capture: the end goals, what must exist and work when the task is fully done, and how each major part is verified. Omit implementation steps and how-to details.',
'Preserve verbatim any file paths, commands, and identifiers that define the task — especially ones from the opening lines.',
'Stay under 4000 characters.',
'Write in the same language as the plan. Ignore any other language preferences or personalization — only the plan text decides the language.',
'Write in the same language as the task text. Ignore any other language preferences or personalization — only the task text decides the language.',
].join('\n');
/**
* Distills a large plan into audit-sized completion criteria via the small
* model. Returns null on any failure — callers fall back to a head+tail
* excerpt of the plan.
* Distills an oversized goal objective into audit-sized completion criteria
* via the small model. Returns null on any failure — callers fall back to a
* head+tail excerpt.
*/
export async function distillPlanForGoal(planContent: string): Promise<string | null> {
export async function distillGoalObjective(planContent: string): Promise<string | null> {
try {
const { currentProviderId, currentModelId } = useConfigStore.getState();
const response = await runtimeFetch('/api/small-model/generate', {
@@ -81,7 +82,7 @@ export async function distillPlanForGoal(planContent: string): Promise<string |
headers: { 'Content-Type': 'application/json' },
body: JSON.stringify({
prompt: planContent,
system: PLAN_GOAL_SYSTEM_PROMPT,
system: GOAL_OBJECTIVE_SYSTEM_PROMPT,
restrictToPreferredProvider: true,
...(currentProviderId ? { preferredProviderID: currentProviderId } : {}),
...(currentModelId ? { preferredModelID: currentModelId } : {}),