Release v1.0.3 - Discovery-yield update · DanRedelien/codebase-optimization-kit

1.0.3 - 2026-05-29

Discovery-yield update focused on surfacing real optimization work for cheap models (for example Composer 2.5) without diluting context.

Added value lanes for the optimization categories the kit was missing: correctness-edge-case, performance-efficiency, resource-lifecycle, concurrency-state-safety, error-handling-recovery, and reinvented-capability.
Wired the previously orphaned performance-auditor role to performance-efficiency.
Added agents plan --focused: one uncapped single-lane task per zone lane (value lanes forced, hotspot-seeded) so weak models go deep on one concern.
Added configurable budgets in policies/audit-criteria.json policy_limits: max_assigned_lanes_per_zone, max_zones_per_agent, max_agent_slots, max_context_tokens, and max_focused_tasks.
Added worked finding examples (finding-performance.json, finding-correctness.json, finding-reinvented.json, finding-resource.json) and referenced them from AGENT.md.
Added a per-lane look-for playbook, a severity bar, and an inlined valid example finding to generated discovery prompts.

Reordered audit lane priority so value lanes run before audit-only lanes; security-risk stays first but only attaches when security signals exist.
Signal-gated resource-lifecycle, concurrency-state-safety, and error-handling-recovery, and capped lanes assigned per zone in broad mode so a larger catalog does not enlarge any single agent's load.
Seeded each task's required_reads with the zone's largest source files and entrypoints ahead of manifests, configs, and tests.
Folded finer audit concerns (contract-boundary, test-gap-hotspot, data-integrity, build-dev-efficiency, observability, configuration-footgun) into existing lane look-fors instead of new lanes.
Made a TODO/FIXME a finding only when it marks a concrete defect, risk, or measurable inefficiency; intentional placeholders, including dormant_planned_code, are out of scope.

Fixed generated task allowed_writes to point at the task-local state/task-findings/TASK-XXX.jsonl output instead of state/findings.jsonl, removing a contradiction with the discovery prompt.
Fixed focused mode dropping whole optimization categories by bypassing the broad-mode per-zone lane cap.
Fixed the dead-code lane never being assigned to any zone: it is now a candidate on source zones, so focused mode runs a dedicated dead-code discovery pass (its strict evidence gates are unchanged).
Added focused-mode regression coverage to scripts/qa.py (single-lane tasks, task-local allowed_writes, uncapped lanes including dead-code and value lanes).