Repository navigation
fix(subagents): retire the solo-spawn gate, restore one-task default - #13458
Conversation
GPT 5.6 Review — ✅ no blocking findingsGPT 5.6 completed its review of This comment is updated in place on each push. Review detailsNo findings. False positive or not applicable? A repository writer can comment: |
0ed8103 to
7b769b3
Compare
The gate is removed and delegation policy becomes prompt guidance only, owned by system-specs/modules/subagent.md. This document lands first so the removal is judged against a decision on the base branch; #13458 carries the code. Grounds: local spawn.solo audit from 2026-09-22 to 2026-09-24 showed 33 of 34 one-task calls allowed (16 specialist, 14 named model/agent/crew, 2 other, 1 refused). The gate checks the shape of a reason, not its truth. Co-authored-by: Joe Guo <zejiangg@amazon.com>
Design Review (Fable 5) — ✅ PASSDesign-level review of Design-Verdict: PASS Re-run the [DESIGN-REVIEWED] 56aa777 |
First Principles Review (Fable 5) — 🟡 CONCERNSPremise-level review of First-Principles-Verdict: CONCERNS The escalation-draft check loses its context-free reviewer — exactly the fresh-context spawn this PR's own retained policy licenses — on no measured evidence. Not justified as shipped
What this change shipsInventory (9 items) — 8 justifiedIntent: retire the ineffective solo-spawn runtime gate and make "one task stays in the parent" the prompt-enforced default — a FIX (the RFC on base,
[FIRST-PRINCIPLES-REVIEWED] 56aa777 |
42c4d8b to
969beb0
Compare
Opus 5 Review — ✅ no blocking findingsReviewed Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
969beb0 to
38ec164
Compare
38ec164 to
5f3360d
Compare
|
self-added: yes
|
052e097 to
4230607
Compare
|
|
|
Rebased onto latest main (new head |
05ae016 to
1aeeb13
Compare
1aeeb13 to
dcee9e4
Compare
|
|
|
|
|
|
|
Reviewed and NOT approving this one — holding it for the maintainer. The diff and the measurement are fine; the authorization chain is what stops me. The PR is judged against the Goal frozen in Nothing here says the conclusion is wrong. The 33-of-34 audit is good evidence, the "checks the shape of a reason, not its truth" argument is correct, and keeping What I would want before this lands: a maintainer other than the author signing off on the decision itself, not just on this diff. Two smaller notes for that reviewer — the removal deletes the only runtime brake and leaves prompt text as the sole control, so the Every other readiness-passed PR in this sweep is approved and armed. This is the only one I held. [operator: bolichen97] |
|
|
…efault in the prompts Users report frequent single-child spawn_run calls. The solo-spawn gate (#11710, widened in #12203) validates reason shape, not value, so the model always has a passing token. A local spawn.solo audit (2026-09-22 to 2026-09-24) showed 33 of 34 one-task calls allowed. - Gate removed: solo_spawn.py, the MCP tool half and the api_spawn roster check/audit are gone. The parent_work_supported receipt check moves to parent_work.py, unchanged. - Compat: solo_reason / solo_details are no longer advertised. Validation still accepts and ignores them so old skills are not refused as unknown fields. - prompt.md / prompt-orchestrator.md: a concrete one-task default; the numbered dispatch -> END YOUR TURN -> wait-for-batch steps are back. - Member protocol (context.py): the front-desk rule does focused work inline; escalation drafts are self-reviewed. - subagent.md: delegation policy is prompt guidance only. - Tests rewritten for the new default. The RFC supersede decision is a separate doc-only PR (#13511) so the First Principles lane reads it off the base branch first.
|
Rebased: main moved in files this head touches since it went green (absorbed main's commits cleanly, still one commit / 22 files). One conflict in |
|
|
Problem / Motivation
Goal: Restore the one-task subagent spawn as the default, so a single-target
spawn_runno longer needs to argue its way past a gate.Users see frequent
spawn_runcalls that start just one subagent. The solo-spawn gate was meant to stop this, but it let almost every such call through.Local
spawn.soloaudit, 2026-09-22 → 2026-09-24:specialistfresh_context/parent_parallelWhy it matters
A one-task spawn pays startup, context-transfer and quota cost for no parallelism. The user waits longer and the parent loses the context it needs to verify the result.
Not a goal
What changed (motivation → approach → change)
solo_spawn.py, the check inmcp_tools/spawn.pyand the roster check/audit indashboard/handlers/messaging.pyare deletedparent_work_supportedcheck stays indashboard/handlers/messaging.py, beside its one caller, unchangedmcp_tools/spawn.pywrites one allowed-onlyspawn.solorecord for a one-taskspawn_runor one-entryspawn_sub_agentscall, so the one-task rate stays measurablevalidation.pystill acceptssolo_reason/solo_details, unadvertised and ignoredprompt.md/prompt-orchestrator.mdstate "one task: do it yourself"context.pykeeps focused work in the DM thread; escalation drafts are self-checkedThe gate checks the shape of a reason, not whether it is true. So the model always has a passing token, and the gate adds friction without stopping anything.
The real cause is the prompt. It says "Parent + one child can be two workstreams" and lists five solo reasons, which the model reads as a menu. It also lacks the concrete "one task stays in the parent" default and the "dispatch, then END YOUR TURN" steps. The member protocol sends every substantial item and every escalation draft through a one-task spawn.
This PR removes the gate and fixes the prompts. Spawning is for 2+ independent tasks, bulk output, an independent review that must not see this session, an explicit user request, or a different agent/model/crew. The numbered dispatch → END YOUR TURN → wait-for-batch steps are restored, and the assignment-contract and bounded-parent-work rules are kept. The tool descriptions drop the "ENFORCED … solo_reason" text.
Wording is updated to match in the bundled skill files (
frontend-design-workflow/SKILL.md,image-authoring/SKILL.md,ops-mission-control/sops/dispatch.md,pod-e2e/SKILL.md), the docs (docs/architecture/mcp.md,docs/system-specs/modules/subagent.md,docs/subagents.md,docs/session-control.md) and the eval (eval/README.md,eval/scenarios/subagent_policy.json). Thetesting/workflow_memory_scenario.pyfixture drops itssolo_reason.The decision is recorded on main in
docs/request-for-change/rfc-solo-spawn-gate.md(#13511). Its status moves toin-progressin #13705, the value the First Principles lane reads as a decision; it becomessupersededonce this PR merges.Backwards compatibility
Compatible:
solo_reason/solo_detailsare still accepted by validation, so a skill or workflow that sends them keeps working; they are ignored.Tests
test/test_spawn_single_task_gate.py: no reason vocabulary advertised; a lone task spawns with no gate marker; legacy fields accepted but not forwarded; the gateway never refuses or audits a lone spawn; the parent-work receipt is preserved;parent_work_supportedcases ported. It replacestest/test_spawn_solo_gate.py. It also checks that a one-taskspawn_runand a one-entryspawn_sub_agentseach write onespawn.soloallowed record, and a two-task call writes none.test/test_prompt_compact_contract.py: delegation checks for the new default, plus a denylist of the retired vocabulary.Manual verification
N/A — the change is prompt text and a removed check; unit tests pin both the prompt contract and the spawn path.
Related Issues
no linked issue: reported by users in chat, no tracking issue.
Pattern harvest
Rule candidate: review-prompt
Pattern: a runtime gate that validates a free-text reason's shape, not its truth, gives the model a passing token and blocks nothing.
Checklist