Documentation menu
On this page

0041 — Reasoning-only delegation: delegate_analysis (v83-F7)

Status

Accepted (planned v83).

Question

Every delegation in skep spawns a worker: worktree, contract, sandbox, G10. That is the right cost for work that touches files and exactly the wrong cost for "read these three runs and tell me which approach is better" — analysis needs a context window, not a worktree. Hermes's delegate_task covers this shape; skep had no equivalent.

Decision

delegate_analysis(tasks[1..3], context?) — a carded mutation that runs each task as ONE read-only Queen turn in its own fresh chat (source='analysis'):

  • Read tools only, by construction. The analyst turn runs on the /btw machinery (read_only=True): mutations refuse, never card. Since delegate_analysis is itself a mutation, an analyst can never nest a delegation — no recursion, no fan-out explosion, without any bespoke guard (I5: no new permission logic).
  • The transcript is the record (I8). Each analyst's full conversation persists as an ordinary chat: list_chats shows it, search_chats finds it, get_chat_messages replays it. No shadow run records.
  • The Queen synthesizes. Answers return to the proposing chat; composition stays with the Queen exactly as ADR 0025 keeps batch results (I3: workers/analysts never talk to each other).
  • Cap 3, its own resource class. The ADR 0025 worker cap and this cap are separate dials; raising either is a policy question for real field demand, not a code default.
  • Always carded. There is no repo, hence no project auto-dispatch posture to inherit; the card shows the task prompts verbatim. An auto-allow lane can be added later behind an explicit operator grant if card fatigue shows up in the field — amend this ADR then.

Consequences

  • "Have two analysts compare these designs" costs two provider conversations and zero worktrees.
  • Analyst turns inherit the turn loop's own bounds (tool-round cap, loop nudges, provider retry) — no new budget machinery.
  • Sequential execution for now (≤3 bounded turns); parallel provider streams are a latency optimization the field hasn't asked for.