3.1 KiB
Astra publishing architecture
User requests ruthless reorganization; historical work is in Git. You are openai/gpt-6-astra. Read rollout/architecture-decision.md. No delegates, commits, staging, push or live lab actions. Use apply_patch; checkpoint every 20 calls/15min and before compaction/stopping to campaign/rollout/publishing-worker-state.md. Own only: campaign/board.md, campaign/backlog.md, campaign/open-questions.md, campaign/DASHBOARD.md, tools/dashboard.py, tools/DASHBOARD_README.md, tools/displacement.py, scripts/forgejo_campaign.py, README.md, findings/00-strategy.md, guides/method-rules.md. Lead owns other architecture docs/config, controls architect owns tools/campaign.py and contracts schema, gate worker owns verification tooling. Coordinate by reading their rollout checkpoints.
Replace live planning stack with one concise plan and machine-derived status. No legacy appendix, coverage percentage, displacement score, old verified-table parser or Forgejo second source of truth. Preserve scientific findings/evidence. Delete obsolete displacement and Forgejo publishing scripts after inspecting dependents; remove/update references in owned files. Report references outside ownership for lead to fix, do not silently modify another worker's files.
Board + dashboard generated from campaign/contracts/*.json and explicitly selected current evidence pointers under campaign/current.json (you own current.json). Exact schemas come from controls-worker; inspect campaign/README.md/checkpoint once ready. Until available implement clear adaptable interface; do not invent accepted contracts/results. Current gate schema sots-gate/1 includes status passed|failed, profile host|full, source identities, inputs, binary, tests and limitations. Standalone schema sots-standalone/2 status measured|accepted|failed, pairs and provenance. Never infer freshness from current HEAD or existence of old artifacts. Label historical as historical, dirty snapshots by manifest identity; source mismatch stale. Prefer current.json explicit path+sha256 for gate/replay pointers, validate hash/schema/status and fail on stale/bad pointer. Missing current baseline should say not established, not green. Summary should show capabilities/workloads, executed tests/skips, original dependencies, open blockers/surprises, and next action; no LOC or target percentages. Update dashboard docs + test its negative paths (you may own verify/publishing/ for focused tests).
Current priority: rollout baseline; then complete research write-set replacement pilot; then command payload semantics and ship construction; then recorded-command standalone replay. Pilot remains proposed/blocked pending executable criteria and lease/control evidence. Do not claim it implemented. Collapse strategy and earned-method rules into concise current policy: write-set boundaries; complete state inputs; distinct-state/branch coverage; synthetic boundary tests + real differential; honest failure; original-assisted != independent; Astra resolves surprises; source-bound evidence; no silent stale status. Domain-level old detail remains in findings, not giant active board. Replace README with current entry points and operational commands.