gh-xj/rw-optimise

rw-optimise: MASE+ Feedback Loop Worker

Cold Start

No optimiser rules fire until 3+ cases produce reflections and quality scorecards. The first 3 cases run with default budget from .agentops/budget.md. This is expected — the feedback loop bootstraps from zero data points. Do not attempt to force adjustments before sufficient signal accumulates.

This worker closes the missing loop: accumulated quality signals and lessons from reflections are consumed to adjust budget thresholds, worker prompts, and guardrails for future cases.

Input

• _reflections/*.md and {domain}/_reflections/*.md — accumulated reflection files
• {domain}/synthesis/quality-scorecard.md — quality scores per domain
• .claude/skills/*/cycle-logs/ — per-invocation quality signals (gitignored)
• .agentops/budget.md — current budget allocation
• .agentops/memory/evolution/ — accumulated evolution memory (budget adjustments, routing lessons)
• .agentops/memory/operational/ — operational patterns (domain context, routing patterns)
• .agentops/worker-eval/ledger.md — per-case worker effectiveness scores
• .agentops/worker-eval/signals.md — signal definitions for scoring

Output

• _evolution/optimiser-log.md — append-only log of all adjustments
• _evolution/prompt-history/{worker}-{version}.md — prompt version snapshots
• Edits to .agentops/budget.md (budget threshold adjustments)
• Edits to .claude/skills/{worker}/SKILL.md (guardrail additions)

Responsibilities

1. Read accumulated signals: Scan all _reflections/ files and domain quality-scorecard.md files. Extract: challenge catch rates, quality scores, error patterns, lessons categorised as recurring.
2. Apply rule-based adjustments (v1): See references/optimiser-rules.md for the full rule set. Each rule has a trigger condition, action, evidence requirement, and oscillation guard.
3. Log every adjustment: Append to _evolution/optimiser-log.md with: date, rule triggered, evidence (which reflection/scorecard), action taken, files modified.
4. Snapshot prompt changes: Before modifying a worker SKILL.md, save the current version to _evolution/prompt-history/{worker}-v{N}.md for rollback capability.
5. Preflight mode (--preflight): When invoked at case-open (via on-case-open hook), read signals and recommend budget adjustments without writing any files. Output recommendations to stdout for dispatcher review.

Guardrails

• Max 1 adjustment per cycle: Prevent oscillation. If multiple rules trigger, apply the highest-priority rule only (priority order defined in references/optimiser-rules.md).
• Human review for worker removal: The optimiser may add guardrails and adjust thresholds, but removing a worker from a chain requires human approval. Log the recommendation to _evolution/optimiser-log.md with tag [HUMAN-REVIEW-REQUIRED].
• Evidence requirement: Every adjustment must cite the specific reflection file and metric that triggered it. Adjustments without evidence are rejected.
• Rollback capability: Prompt history in _evolution/prompt-history/ enables manual rollback if an adjustment degrades quality.

Job-Hunter Signal Priorities

When multiple rules trigger, prioritise signals in this order for the UK SaaS context:

1. Legal accuracy degradation — Any finding that legal accuracy (GDPR/ICO analysis) is below high confidence triggers a budget increase for 05-gdpr-legal and 03-visa-sponsorship cases. Legal errors are the highest-risk failure mode.
2. Platform coverage gaps — If scraping feasibility findings are missing major UK platforms (Reed, Totaljobs, CV-Library) after a scraping case, trigger re-discovery.
3. Challenge catch rate below 20% — If rw-challenge raises zero issues across 3+ cases, the challenge worker is rubber-stamping. Apply a guardrail strengthening adversarial prompts.
4. Cost-model findings missing — If infrastructure decisions lack cost-at-scale modelling, flag for deliberation re-run.

Self-Check (Level 1 Self-Correction)

Before completing, verify:

• [ ] At most 1 adjustment was made this cycle
• [ ] Adjustment is logged in _evolution/optimiser-log.md with evidence citation
• [ ] If a worker prompt was modified, previous version is saved to _evolution/prompt-history/
• [ ] No worker was removed from a chain without [HUMAN-REVIEW-REQUIRED] tag

Max 2 self-correction iterations. If self-check still fails after 2 retries, emit status-recommendation: blocked with a description of what failed.

Reference Files

• references/optimiser-rules.md — v1 rule set with trigger conditions, actions, evidence requirements, and oscillation guards
• references/skill-lifecycle.md — skill lifecycle rules, skill.yaml format, cycle-log format, quality drop trigger