Repository navigation
feat(spawn): refuse a one-task spawn that gives no solo_reason - #11710
Conversation
|
Intent: Make the "one sub-agent for one task is not a reason" rule mechanical: a one-task |
Design Review (Fable 5) — 🟡 CONCERNSDesign-level review of Design-Verdict: CONCERNS Closed vocabulary omits the user-requested background job, so honest models block the chat slot or file false Watch
Suggestions
[DESIGN-REVIEWED] c9af747 |
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: |
First Principles Review (Fable 5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Verify model-pinned skill spawns (prepare-pr retrospective, llm-council members) survive a parent slot pinned to the same model — the gate refuses that match by design. What this change shipsInventory (10 items) — 10 justifiedIntent: stop the agent spawning one sub-agent for one task it could do itself, by making the tool refuse until a defensible reason is given — an ADDITION, implementing the decision recorded in
Checks run: [FIRST-PRINCIPLES-REVIEWED] c9af747 |
Opus 4.8 Review — ✅ no blocking findingsReviewed Review detailsThe discovery pass produced no candidates, and having traced both halves of the solo gate through the diff — the tool-side count refusal in No findings. [OPUS-REVIEWED] c9af747 Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
|
|
|
|
7854070 to
50384fb
Compare
50384fb to
7e169cb
Compare
|
|
|
|
|
|
A one-task spawn_run / one-entry spawn_sub_agents that names no solo_reason and no model/agent/crew other than the caller's own is refused with a question instead of spawned; the reason is a closed vocabulary (bulk_data, fresh_context); a genuinely different model/agent/crew is a reason in itself, checked by the gateway against the parent's own; every outcome is audited as spawn.solo; unknown parent facts fail open; programmatic /api/spawn callers are not gated. This is the product-shape record the First Principles lane requires for PR kirodotdev#11710, which implements the design whole. Status in-progress; the README index row states what is on main today (nothing of the design). Co-authored-by: EC2 Default User <ec2-user@ip-172-31-10-71.ec2.internal>
…otdev#11710) A single spawn_run(task=...) or spawn_sub_agents with one entry is a round-trip with no parallelism gain, and the textual "do it yourself" gate in the tool descriptions kept losing to the base prompt's "delegate to a sub-agent". Make the gate mechanical: a one-task call that names no solo_reason and no model/agent/crew other than the caller's own is refused before any POST, and the result asks whether the caller can do the task itself. A retry needs a reason from a closed vocabulary (bulk_data, fresh_context) or a genuinely different model/agent/crew; the gateway compares the named one against the parent session's own and refuses a match with solo_spawn_unjustified. The reason is echoed in the spawn result and audited as spawn.solo. Batches of 2+ tasks and programmatic /api/spawn clients (SDK, apps) do not pass through the gate. Docs, prompt.md and the pod-e2e skill name the new parameter; existing spawn tests that exercised other behaviour through a lone task now pass a reason. Co-authored-by: EC2 Default User <ec2-user@ip-172-31-10-71.ec2.internal>
…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.
…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.
…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.
…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.
…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.
…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.
…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.
…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.
…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.
…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.
…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.
…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.
…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.
…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.
…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.
…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.
…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.
…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.
…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.
…efault in the prompts (#13458) 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. Co-authored-by: Joe Guo <zejiangg@amazon.com>
Problem / Motivation
Decision recorded in
docs/request-for-change/rfc-solo-spawn-gate.md(merged in #11848, statusin-progress, implementation PR: this one).spawn_run(task=...)with ONE task spawns one sub-agent for one job. That is a round-trip with no parallelism gain. The tool text already says "do it yourself", but text is advice: the base prompt's "delegate to a sub-agent to preserve context" kept winning, and a lone sub-agent kept appearing for a single investigation, fix or review.Why it matters
Every lone sub-agent is a full extra model run for work the parent could do in its own turn. The rule had to be re-taught by hand, and a rule that lives only in prose does not hold.
What changed (motivation → approach → change)
A one-task spawn is now a handshake, not a straight dispatch. The tool refuses the first call and asks: can you do this yourself? The second call must carry a reason from a closed list, or name a model / agent / crew that is really not the caller's own. The refusal names the
solo_reasonparameter and points at its description; it never spells the passing words, which live in the schema alone.tasks=[a, b](2+ tasks)task, nothing elsetask+solo_reason(bulk_data/fresh_context)task+ amodel/agent/crewthat differs from the parenttask+ the parent's OWN agent / model / crewsolo_spawn_unjustifiedtask+ onlykeep=true/api/spawn🟩 added · 🟨 changed · 🟥 removed · 🟦 unchanged
Why a closed list and not a yes/no flag: a
confirm=trueis rubber-stamped by the second call. The list has no "preserve my context" and no "separate investigation", so the caller has to pick a reason it can defend.bulk_datameans the step would flood the caller's context with output.fresh_contextmeans the result would be wrong if the run saw this session's context (a blind review, a clean-slate repro).Why two halves: the MCP tool process is the only place the task COUNT is known (the gateway sees one POST per task), so
mcp_tools/spawn.pyrefuses the one-task call that names nothing. Only the gateway knows the parent session's own agent and, for a dashboard slot, its model, sodashboard/handlers/messaging.pycatches the call that named the parent's own agent / model / crew to slip past.agentis compared against the parent's resolved template, never its member alias, so a member onkirocrew-workercannot namekirocrew-workeras "different". The tool marks such a POST withsolo=true; the SDK and apps never send it, so they are never gated. An unknown parent fact fails open: the roster check never refuses what the tool side let through unless it can prove a match. All three gate outcomes are audited asspawn.solo: denied, allowed on a reason, allowed on a difference (with what differed, e.g.differs=agent).solo_spawn.py(new) holds the vocabulary, the refusal text, the result line and the roster comparison, so the tool description, the validator and the gate cannot drift apart.validation.pyboundssolo_reasonto that vocabulary on bothSPAWN_RUN_SCHEMAandSPAWN_SUB_AGENTS_SCHEMA.spawn_sub_agentsgets the same gate: one entry, noagent_or_mode, no reason → refused.Words the agent reads:
config/prompt.mdstep 1 of Subagent Orchestration names the refusal and the two reasons;docs/subagents.mdgains a "The solo gate" section; the module specdocs/system-specs/modules/subagent.mddescribes the gate where it pinned the old one-task shape; the Dev Fleetpod-e2e/SKILL.mdpassessolo_reason="bulk_data"for its one QA sub-agent and stops calling "not derailing my context" a reason to delegate. Every shipped instruction that produced a bare one-task spawn is brought under the gate:builtin_skills/ops-mission-control/sops/dispatch.md(the unattended investigator launches withfresh_context),builtin_skills/frontend-design-workflow/SKILL.md(the blind usability reviewer is afresh_contextcase and now says so),builtin_skills/image-authoring/SKILL.md(fan out only for 2+ images; one image is authored inline). The offline E2E stub scenariotesting/workflow_memory_scenario.pypasses the same reason for its nested one-task spawn, which exercises private-memory routing, not the gate.Tests
test/test_spawn_solo_gate.py, 53 tests:spawn.solodenied; each reason → spawns with the reason on the wire and in the result line; a named model / agent / crew → passes and the result points at thespawn.soloaudit for the ground;model="auto"(the "no model chosen" sentinel) counts as nothing named and is refused;keepalone → refused; 2+ tasks → never gated; the same forspawn_sub_agentswith one entry.defaultcrew, a member naming itself as crew, a member naming its own resolved template as agent →""; another agent / model / crew → the ground (agent/model/crew); a dashboard slot is found by its effective session key, so a chat linked to a channel thread is compared too; unknown parent, unpinned slot model, a slot owned by another session → fail open and say(parent unknown).solo=truenaming the parent's own agent → 400solo_spawn_unjustified, auditspawn.solodenied; a member naming its own template → the same 400; a reason → allowed and audited; another agent → allowed and audited withdiffers=agent; a POST without thesolomarker → never gated; a reason outside the vocabulary → validation 400.""means not given, and both tools advertise the enforced gate with the same enum.Nine existing spawn test files that drove OTHER behaviour through a lone
task(effort forwarding, the context-group wire contract, cancellation, audit owner, redaction, model forwarding) now passsolo_reason="bulk_data"so they still reach the POST they assert on;test_spawn_reasoning_effort.pyandtest_subagent_context_group_boundary.pydo it through a small_through_solo_gatehelper that adds the reason only where the gate would fire.Manual verification
N/A — unit coverage sufficient: both halves of the gate are pure functions over the call's arguments and the parent's roster, and every row of the matrix above is a test.
Related Issues
no linked issue: the gate came out of a dashboard conversation, not a tracked issue.
Checklist
feat|fix|docs|refactor|perf|test|chore|ci|build|revert: ...)