Rp Subagent Dispatch

rp-subagent-dispatch

Most of a run's token cost is history re-sent on every model call, not tool output. A subagent that is resumed round after round keeps re-sending everything it ever saw: one measured run spent 239.5M tokens in a single resumed Developer whose context reached 919K, out of 522M for the whole run. These rules keep each subagent short.

A fresh subagent every round

• Every Developer or QA dispatch is a new subagent. Never resume or re-message one from an earlier round — not for QA findings, a crash fix, a retry, or "continue".

• Before each re-dispatch, write the handoff:

  Copy

  node skills/wix-replatform/scripts/subagent-handoff.js write --project <dir> \
    --stage <stage> --resource <resources/rp-...> --outcome <outcome> \
    [--findings <QA result file>] [--changed <path,path>] [--notes <one line>]

  outcome is done, failed, qa_findings, budget_exceeded, needs_decision or resume_limit. It writes state/handoffs/<stage>-<round>.json (≤ 8 KB) from durable state only: the QA findings verbatim, a write-log summary (counts, never rows), the completion report status, open user actions and the paths changed last round. It prints { file, round, bytes, budgetRoundsWithoutProgress }.

• The new subagent's prompt is only: the resource path, the project directory and the handoff file. Paste nothing else — no earlier result text, transcript or log. The subagent reads the handoff first; anything it trimmed is still in its source files.

A context budget per subagent

• Each Developer works within about 200K context or 150 model calls, whichever comes first.

• Near the budget it stops starting new work, runs subagent-handoff.js write --reason budget_exceeded (with --changed and --notes for where it stopped), and returns budget_exceeded with the handoff path.

• budget_exceeded is not a failure and not a QA miss. Dispatch a fresh Developer for the same resource with that handoff.

• When the printed budgetRoundsWithoutProgress reaches 3 (three budget rounds in a row with no new settled import progress), tell the operator in one status line that the stage is repeating without progress, add the same note to the next handoff (--notes), and keep going. Never stop the run for it.

• Where the host does not report context size, count your own model calls, and keep every tool output under 20 KB by using the bounded helpers below.

Bounded helpers instead of raw files

• Never open, cat, grep, tail or jq state/attempts/write-attempts.ndjson or logs/import-audit.ndjson — they reach tens of MB and stay in context for the rest of the subagent's life. Use:

  Copy

  node skills/wix-replatform/scripts/write-log-summary.js --project <dir> \
    [--entity <targetEntityType>] [--status <status>] [--limit 20] \
    [--run latest | --since-run <runId>] [--log audit] [--json]

  Default output: counts per entity × status (current status per attempt) and top error classes; --entity/--status add up to --limit matching rows. Output stays ≤ 4 KB.

• Run a generated import with scripts/import-supervise.js (see rp-execute-import): one tool call that launches, polls, resumes and returns complete, needs_decision or resume_limit, instead of one model call per poll.