Shared status and publisher protocol

Shared skills work

Read https://skills.imerlino.com/api/work-status before resuming skills work. The human view is https://skills.imerlino.com/activity. Completed model responses are not automatically skill wins. Source versions and aliases are not active skill counts.

The laptop publisher checks allowlisted receipts and editorial records every 30 seconds. The site refreshes every 10 seconds. A publisher heartbeat older than three minutes is marked overdue. A recent heartbeat confirms the publisher is connected; inspect the work record timestamp and receipt state to determine whether testing is advancing.

Every environment writes a separate status ID. Never replace another environment's record. Update current/next and state as work changes. Use working, waiting, idle or complete. Keep original skills archived and activate only the tested procedure needed for a project.

Read access requires no credential. Authenticated writers on the user's machines can pipe a UTF-8 JSON record to ssh main python F:/Projects/skill-status/publish.py. That helper reads the existing dedicated publisher token on MAIN; do not copy it into prompts, logs or other hosts.

Record envelope: {"kind":"status","id":"your-environment-id","value":{...}}. Copy the field structure from the status API, set environment to your ID, refresh updated_at in UTC, and supply factual counts and explicit limitations. The write endpoint is POST https://skills.imerlino.com/api/work-status. Successful writes return saved:true and received_at.

Only publish intended review material: status, decisions, source hashes and fictional final outputs. Never publish raw CLI streams, reasoning traces, credentials, local private paths or unreviewed source SOPs. Evidence records use kind=artifacts and an ID, title, summary and outputs array. Original outputs retain SHA-256 hashes.

The laptop bridge lives in work/publish_shared_status.py in this task. once publishes immediately; watch follows changes with a single-instance lock. Set outputs/Skill-Lab/shared-live/activity.json to the current work statement. Create outputs/Skill-Lab/shared-live/STOP to stop the watcher. No scheduled restart or automatic account switching is configured.

Current reviewed local SEO outputs, source hashes, short operator rules and the tested pure freshness calculation are portable at https://skills.imerlino.com/activity/results/local-operations-20260912 (JSON: https://skills.imerlino.com/api/work-status/artifacts/local-operations-20260912). A local pathname in older evidence is not available on every host; use the embedded final answer and reference text, or retrieve a verified archive source when needed.

The laptop watcher supports per-batch files at outputs/Skill-Lab/shared-live/batches/*.json. Each contains a batch object and optional evidence with outputs. These files are reread every 30 seconds. Keep completed, planned and review status distinct; only publish reviewed-safe final outputs.

Readiness and immutable change log

The homepage is the use/shelve desk. Read /api/work-status first, then every environment's usability_artifact through /api/work-status/artifacts/<id>, and /api/work-status/changes. The latter returns the latest 50 append-only decisions plus next_offset; request ?offset=<next_offset> for older entries. This includes decisions, not connection heartbeats. Original source archives remain intact.

Each environment owns a separate record: usability-<environment-id>. Do not overwrite another environment's snapshot or status. Use the existing MAIN publisher, which accepts a JSON envelope or array on stdin; credentials stay on MAIN. Inspect the current snapshot and exact source hash before working, and mark your environment working with the claimed item IDs in current/next. This is advisory coordination, not an atomic task lock. Check other environments before duplicating a test.

Readiness artifact value: schema_version:1, environment, title, updated_at (UTC ISO), summary, outputs:[], and items. Every item needs a stable id, name, kind (skill/command/helper/reference), domain, status (usable/needs-work/shelved/retire-candidate), reason, scope, next_action, exact 64-character source_sha256, and evidence entries with title and an HTTPS skills.imerlino.com activity/evidence URL. Copy an existing record as the schema example. Usable and retire-candidate decisions require evidence. No evidence means shelved, not garbage. A bounded win is not a whole-workflow pass. Cross-environment readiness still needs local dependency verification.

Derive counts from your snapshot's actual item statuses. Keep source_coverage and status counts aligned with your verified evidence ledger; do not carry forward older totals or copy another environment's coverage as your own. Reject duplicate item IDs and unknown states. Linking an older completed comparison is zero new model requests: record its existing-answer count separately so summaries do not double-count work. Evidence or reason corrections can retain the same from/to state in a new change event.

On a meaningful decision change, first publish the evidence artifact, then your updated readiness snapshot, then a separate kind:changes envelope. Its unique id must start with <environment-id>-. The value contains environment, at, title, summary, registry_id:usability-<environment-id>, and a non-empty changes array: item_id, from (old state or null), to, reason, evidence_url. Publish only changed items. Use one stable event ID for retries. Identical retries succeed; changing an existing event returns409. Correct mistakes by appending a new event, preserving history. There is no atomic transaction across snapshot/event writes: keep both receipts, and retry a missing event with the same ID.

Finally update your own status with usability_artifact, truthful current/next, model/run state and UTC timestamps. Set idle when no worker runs. A newer publisher heartbeat is not a new test or review. Keep original failed outputs and do not rerun completed comparisons without a concrete reason. Use signed-in CLI subscriptions only. Published readiness values and event receipts are the shared handoff; laptop paths alone do not make a helper available elsewhere.