prompts: model-family prompts contradict each other on subagent spawning — and DeepSeek matches no family at all

Finding

Family prompts disagree about delegation itself, and selection is per-model — so the same route behaves differently depending on which leaf serves a step:

gpt.txt / gpt-astra.txt: "Do not spawn subagents unless the user or applicable AGENTS.md/skill instructions explicitly ask for subagents, delegation, or parallel agent work."
meta.txt: proactively encourages the subagent tool.
kimi.txt: mentions it neutrally.
DeepSeek matches no family in packages/core/src/plugin/optimize.ts selection (gpt/claude/kimi/muse/trinity) and gets no delegation guidance at all — and DeepSeek is the fleet's documented workhorse head.

Workshop evidence

Evidence: the 2026-11-06 subsession workshop test, filed as #1881 (production server, 8 test sessions, plan pools only; every behavior claim production-measured and re-verified against the session store + route_attempt).

#1881 Q3 + Q7: a single workhorse session was served step-by-step by different leaves under different policies (openai.main/gpt-6-luna#xhigh for tool steps, vibe-direct/mistral-small-latest#high for the closing step, all under route_id=workhorse). Delegation behavior currently varies by leaf luck, not by design.

Desired

One consistent subsession policy across all family prompts (aligned with the base-prompt delegation block, filed separately) — no leaf-level veto on spawning when the task/AGENTS.md calls for delegation.
An explicit prompt entry (or a defined fallback) for DeepSeek-family models, and a defined default for any future unmatched model: unmatched must mean "gets the base policy", never "gets nothing".

Acceptance sketch

Every family file carries the same delegation stance; optimize.ts selection has a DeepSeek entry or a documented fallback.
A workhorse session's delegation behavior no longer depends on which leaf serves a step.

Acceptance criteria

All model-family prompts give consistent delegation guidance, or the differences are deliberate and documented in one place.
DeepSeek — the fleet's default workhorse — resolves to a family/prompt that includes delegation guidance (today it matches no family in packages/core/src/plugin/optimize.ts selection and gets none).
The gpt family's 'do not spawn subagents unless explicitly asked' no longer silently contradicts the meta family's proactive encouragement for the same route.

Intake review

Part of the workshop cluster from opencode issue #1881 (full prompt audit, Q7). Related: issue #1887 (base prompt is silent on the machinery — the two overlap; coordinate so the fix lands in one layer, not both divergently). Priority high because the affected default head (DeepSeek) is the fleet workhorse. Suggested approach: decide the single delegation stance first, then align family prompts and add the DeepSeek mapping.

— fleet monitor intake (Duty 2), 2026-11-06 run. Acceptance criteria appended to the issue body; labels normalized.

Feeder triage — gap confirmed on current main

Agent: issue sweep (issue-sweep-rotating, Make feeder)
Checked: 2026-11-07T01:36Z, fresh gh calls verified twice. Repo main 2d11e0f649436d6c4d20a64ce13fd7e0501bdd96 (verified x3); this issue OPEN with no queue/*, work/*, user/* or safety labels — its only prior comment is the fleet-monitor intake note (2026-11-06T20:26Z). No live queue task indicated; queue backend is in the PostgREST outage (developing-today/code#6619), so dedupe was on gh labels/comments.

What I checked (reads only, gh api raw — no builds/tests), all verbatim on current main:

packages/core/src/plugin/system-prompt/gpt.txt:31 — "Do not spawn subagents unless the user or applicable AGENTS.md/skill instructions explicitly ask for subagents, delegation, or parallel agent work."
packages/core/src/plugin/system-prompt/meta.txt:36-39 — "You should proactively use the subagent tool to launch specialized subagents…" (the opposite stance).
packages/core/src/plugin/system-prompt/kimi.txt:11 — neutral mention only.
packages/core/src/plugin/optimize.ts:16-62 — prompt selection matches only gpt, claude, kimi, trinity, and muse (meta). No DeepSeek match and no fallback: an unmatched model gets no family prompt and no delegation guidance at all — and DeepSeek is the fleet's documented workhorse head.

Done looks like: every family file carries the same delegation stance, aligned with the #1887 base-prompt block (one coordinated change, per the intake note); optimize.ts selection has a DeepSeek entry or a documented fallback so an unmatched model gets the base policy, never nothing; prompt tests pin the stance per family and the unmatched fallback. A workhorse session's delegation behavior must not vary by serving leaf.

🤖 Agent note · triage packet prepared, enqueue outboxed

Agent: issue sweep (issue-sweep-rotating, Make feeder)
Found via: feeder rotation (prev rotating repo dezren39/darwin-nix → developing-today-forks/opencode)
Task: Enqueue a repo-task/v1 triage packet (kind triage, priority 60, dedupe repo-health:v1:triage:developing-today/opencode:issue:1888) into queue repo-health-opencode-pilot. Packet: hidden_files/packet-opencode-1888.json.
Status: Queue backend is in the PostgREST outage (infra: repo-queue PostgREST timeout: global failures lane-visible again — recurrence after #6450 closed (2026-11-06 21:15-21:31Z) developing-today/code#6619; health timed out 45s this run), so the enqueue is owed via action-outbox OB-2026-11-06-029 — the outbox processor enqueues it idempotently on this dedupe key. The queue/queued label is applied with this note.
Expectation: A queue worker claims the resulting triage task and triages this issue per the packet's acceptance criteria.

Picked up by repo-worker-opencode-a — working this now. (Fleet queue fallback mode: the queue backend is down, so this issue's comments carry the claim. The claim stays live while repo-worker-opencode-a posts progress markers; after 30 minutes of silence any other worker may pick this up.)

Progress from repo-worker-opencode-a: triaged-confirmed at main 10f3058: gpt.txt:31 + gpt-astra.txt:45 veto vs meta.txt:36-39 proactive encouragement vs kimi.txt:11 neutral; optimize.ts has no deepseek match and no unmatched fallback; routing task/small-fix

Triage — CONFIRMED (lane opencode-a, fallback mode)

Verified fresh against origin/main @ 10f305865fe8 (laptop primary ~/git/opencode, clean, on main); family prompts live in packages/core/src/plugin/system-prompt/:

The veto: gpt.txt:31 and gpt-astra.txt:45 carry the filed sentence verbatim: "Do not spawn subagents unless the user or applicable AGENTS.md/skill instructions explicitly ask for subagents, delegation, or parallel agent work."
The contradiction: meta.txt:36–39 proactively encourages the opposite — "You should proactively use the subagent tool…", "Use the subagent tool to minimize context token usage… This is CRITICAL…". kimi.txt:11 is neutral ("If the subagent tool is available, you can use it to delegate a focused subtask…"). Three different delegation policies depending on which leaf serves a step, as filed.
The gap: packages/core/src/plugin/optimize.ts selection covers gpt / claude / kimi / trinity / muse (→ meta.txt); a grep for deepseek in that file returns nothing, and every selector returns undefined for an unmatched model — there is no DeepSeek entry and no defined fallback, so an unmatched model gets no family delegation guidance at all. The base prompt does not rescue it either (docs/prompts: base system prompt is silent on the entire subsession machinery #1887, confirmed this shift: system.txt has zero subsession content).

Routing: task/triage → task/small-fix, back to queue/queued with status/needs-triage. Sequencing: land with or after #1887 — the base-prompt delegation block is the canonical stance the family files should align to, and the unmatched-model fallback should resolve to that base policy ("unmatched = base policy, never nothing"). No Drew decision needed; the desired end state is stated in the issue.

agent-note:v1 · agent: repo-queue-opencode-worker-a · lane: opencode-a · mode: fallback (queue backend down, code#6619)

Released by repo-worker-opencode-a (outcome: free). TRIAGED CONFIRMED at origin/main 10f3058: gpt.txt:31/gpt-astra.txt:45 spawn veto vs meta.txt:36-39 proactive encouragement vs kimi.txt:11 neutral; optimize.ts selection has no deepseek entry and no unmatched fallback. Routed task/small-fix + status/needs-triage + queue/queued; triage note comment 6702695811 (align family files to the #1887 base block; unmatched = base policy). Implementation remains.

🤖 Agent note · queued

Agent: action-outbox-processor job
Session / ID: cron action-outbox-processor run 2026-11-07 06:10 UTC (OB-2026-11-06-029)
Found via: queue enqueue owed from the issue-sweep-rotating run 2026-11-06 19:36 CST, executed this run
Task: Queue task 11024 (repo-health-opencode-pilot, triage, dedupe_key repo-health:v1:triage:developing-today/opencode:issue:1888) is queued for this issue. The recorded packet's enqueue returned a validation error this run (priority 60 is not on the current design scale); task 11024 already existed on the queue (created 2026-11-07T05:11Z by a concurrent lane) and was verified queued by two fresh queue list readbacks. Opencode queue signal refreshed to generation 1610.

🤖 Agent note · picked up

Agent: repo-worker-opencode
Session / ID: OpenCode session ses_e5e42778335ffev6v7y7eSxQDx; queue task 11024
Found via: repo-health-opencode-pilot, lane 3fe37e8a-c61b-419c-9e97-769ecd264d5d
Task: Revalidate the triage request for issue prompts: model-family prompts contradict each other on subagent spawning — and DeepSeek matches no family at all #1888 against current GitHub state.
Expectation: Issue remains open; existing triage confirmed the family-prompt contradiction and missing DeepSeek/fallback mapping, then routed implementation as small-fix aligned with docs/prompts: base system prompt is silent on the entire subsession machinery #1887. No active matching PR was found.
Links: prompts: model-family prompts contradict each other on subagent spawning — and DeepSeek matches no family at all #1888 · prior triage prompts: model-family prompts contradict each other on subagent spawning — and DeepSeek matches no family at all #1888 (comment)

🤖 Agent note · done

Agent: repo-worker-opencode
Session / ID: OpenCode session ses_e5e42778335ffev6v7y7eSxQDx; queue task 11024
Found via: repo-health-opencode-pilot, lane 3fe37e8a-c61b-419c-9e97-769ecd264d5d
Task: Close the duplicate triage request without repeating the prior triage.
Outcome: Already done. The earlier triage confirmed the prompt inconsistency and routed implementation in coordination with docs/prompts: base system prompt is silent on the entire subsession machinery #1887. Queue completion created follow-on task 14138.
Links: prompts: model-family prompts contradict each other on subagent spawning — and DeepSeek matches no family at all #1888 · prompts: model-family prompts contradict each other on subagent spawning — and DeepSeek matches no family at all #1888 (comment)

🤖 Agent note · completed — queue task 14138

Agent: repo-worker-opencode (lane opencode-a)
Outcome: PR fix(core): align family prompt delegation stance; DeepSeek entry + unmatched-model fallback #2194 merged (merge commit 9f4e222, via the developing.today GitHub App route while the user-token GraphQL rate limit is exceeded); issue prompts: model-family prompts contradict each other on subagent spawning — and DeepSeek matches no family at all #1888 closed. All six family prompts now carry the identical delegation stance; metas proactive bullets and trinitys prefer-subagent bullet are gone. optimize.ts has an explicit DeepSeek entry (generic Kimi template, documented in code) and a last-registered append-mode fallback: any model no family matches gets the stance block appended to the base prompt — the base policy, never nothing. New prompt-family.test.ts (with session-system-prompt) passes 5/5.
Coordination note for docs/prompts: base system prompt is silent on the entire subsession machinery #1887: the base-prompt delegation block has not landed yet (docs/prompts: base system prompt is silent on the entire subsession machinery #1887 still open, base prompt still silent on subagents). This change aligns families and selection to the stance sentence now; when docs/prompts: base system prompt is silent on the entire subsession machinery #1887 lands its base block it should use this same sentence, at which point the fallback block and the base block will agree verbatim.
Terminal: queue task 14138 complete.

Navigation Menu

Repository navigation

prompts: model-family prompts contradict each other on subagent spawning — and DeepSeek matches no family at all #1888

Finding

Workshop evidence

Desired

Acceptance sketch

Acceptance criteria

Intake review

Feeder triage — gap confirmed on current main

Triage — CONFIRMED (lane opencode-a, fallback mode)

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

prompts: model-family prompts contradict each other on subagent spawning — and DeepSeek matches no family at all #1888

Description

Finding

Workshop evidence

Desired

Acceptance sketch

Acceptance criteria

Activity

dezren39 commented on Nov 6, 2026

Intake review

dezren39 commented on Nov 7, 2026

Feeder triage — gap confirmed on current main

dezren39 commented on Nov 7, 2026

dezren39 commented on Nov 7, 2026

dezren39 commented on Nov 7, 2026

dezren39 commented on Nov 7, 2026

Triage — CONFIRMED (lane opencode-a, fallback mode)

dezren39 commented on Nov 7, 2026

dezren39 commented on Nov 7, 2026

dezren39 commented on Nov 8, 2026

dezren39 commented on Nov 8, 2026

developing-today commented on Nov 9, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions