openclaw/rank
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Extract flat Scattermind guardrails

Browse Source
This commit is contained in:
OpenClaw Bot
2026-05-26 23:44:20 +02:00
parent adcef9a6f7
commit 35b3e6a47d
3 changed files with 44 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+1 -1
View File
@@ -47,7 +47,7 @@ Ranker's continuation job is narrow:
`POST /api/rank-feedback` accepts a `prioritix-feature-set-v1`-style payload from Scattermind and returns ranked items plus `buildOrder.doFirst / validateNext / defer / park`. It accepts candidate arrays as `features`, `actions`, `nextMoves`, or `candidates` either at the top level or under `featureSet`, and it can consume a nested Concept Map directly, so Scattermind can hand off `conceptMap.nextActions / nextMoves` without renaming them into fake software features. Sectioned Concept Maps may also include `validateNext`, `deferred`, and `parkingLot`; Ranker combines those sections into one build-order pass while preserving `sourceSection` and treating deferred/parked sections as lane hints. Empty wrapper arrays are ignored rather than allowed to shadow a real nested Concept Map action set, and non-empty normalized wrappers are merged with Concept Map validation/deferred/parking sections rather than dropping that context. That keeps partially-normalized Scattermind exports rankable without losing the source lane contract. It also returns a `brief` with source, next-48-hour actions, carried-forward assumptions/constraints/non-goals, and `whatWouldChangeRanking` checks, plus a `handoff` object (`rank-feedback-result-v1`) with source provenance, item trace rows, and contract warnings for missing artifact IDs, source sections, original prompt provenance, or evidence on active items. Keep this contract action-first; do not use it as a reason to add generic dashboard, auth, billing, or workspace layers before the bridge has proof.
Candidate items may include optional 1–10 `rankerHints` (`value`, `effort`, `confidence`, `urgency`, `revenue`, `novelty`, `risk`). Ranker blends those explicit Scattermind hints with text heuristics; `effort` is inverted into feasibility. Hints improve the defended order, but `recommendedLane: "defer"` or `"park"` remains a safety rail so dashboard-swamp items do not get promoted by flashy wording. For clean bridge handoff, Scattermind should send `sourceSection` and `evidenceNeeded` on each active next move. Scattermind can also send `targetAudience`, `constraints`, `assumptions`, and `nonGoals` / `avoid` at the top level, in `featureSet`, inside `conceptMap.context`, or as a structured top-level `context` object with `summary`, `targetAudience`, `constraints`, `nonGoals` / `avoid`, and `assumptions`; Ranker turns that structured context into readable scoring text instead of leaking `[object Object]`. Ranker returns that decision context in `input.decisionContext` and `handoff.decisionContext`, and penalizes candidates that conflict with source non-goals (for example saved workspaces/auth/billing before the continuation proof). If Scattermind sends duplicate candidate IDs, Ranker keeps the first ID, suffixes later duplicates (`preview-2`), and reports the normalization in `handoff.warnings` / `handoff.itemTrace` so downstream build-order references remain addressable.
Candidate items may include optional 1–10 `rankerHints` (`value`, `effort`, `confidence`, `urgency`, `revenue`, `novelty`, `risk`). Ranker blends those explicit Scattermind hints with text heuristics; `effort` is inverted into feasibility. Hints improve the defended order, but `recommendedLane: "defer"` or `"park"` remains a safety rail so dashboard-swamp items do not get promoted by flashy wording. For clean bridge handoff, Scattermind should send `sourceSection` and `evidenceNeeded` on each active next move. Scattermind can also send `targetAudience`, `constraints`, `assumptions`, and `nonGoals` / `avoid` at the top level, in `featureSet`, inside `conceptMap.context`, or as a structured top-level `context` object with `summary`, `targetAudience`, `constraints`, `nonGoals` / `avoid`, and `assumptions`; Ranker turns that structured context into readable scoring text instead of leaking `[object Object]`. If Scattermind only has a flat context string, Ranker now extracts simple guardrails such as `Solo builder`, `Manual proof`, `Avoid ...`, `No ...`, and `Do not ...` into `input.decisionContext` / `handoff.decisionContext` so early bridge exports still protect against dashboard/auth/billing drift. Ranker returns that decision context in `input.decisionContext` and `handoff.decisionContext`, and penalizes candidates that conflict with source non-goals (for example saved workspaces/auth/billing before the continuation proof). If Scattermind sends duplicate candidate IDs, Ranker keeps the first ID, suffixes later duplicates (`preview-2`), and reports the normalization in `handoff.warnings` / `handoff.itemTrace` so downstream build-order references remain addressable.
Lane safety note: explicit Scattermind `defer` / `park` hints are hard rails, not mild suggestions. Source `nonGoals` / `avoid` guardrails are also hard enough to keep conflicting candidates out of Do first / Validate next even when their local scoring hints look attractive; the result will mark the lane source as `source-non-goal` so the handoff can explain that the candidate needs guardrail resolution before active work.