Bug Report: `swarm spawn` `model` parameter is ignored and inherits coordinator model

Bug Report: swarm spawn model parameter is ignored and inherits coordinator model

• Reporter: yamada-masa
• Date: 2026-09-04
• Version: jcode v0.81.6 (4f0f9819a)
• Severity: High (Directly causes self-review failures)

---

Summary

When specifying the model parameter in swarm spawn, the swarm worker inherits the coordinator's model instead. The specified model is completely ignored.

This completely breaks use cases like "spawning a worker with a different model for code review":

• Results in a self-review (reviewing oneself using the same model), making the review meaningless.
• Prevents split-role workflows (e.g., Coder using one model and Reviewer using another, per Constitution Article 7).

---

Reproduction Steps

Experiment 1: Specifying a bare model

{
  "action": "spawn",
  "model": "z-ai/glm-5.2:free",
  "prompt": "(short prompt)",
  "label": "PR82 reviewer"
}

• Actual Model Used: deepseek-v4-flash (same as coordinator)
• Expected: z-ai/glm-5.2:free

Experiment 2: Specifying a route-pinned model

{
  "action": "spawn",
  "model": "openrouter/z-ai/glm-5.2:free",
  "prompt": "(short prompt)",
  "label": "model probe"
}

• Actual Model Used: deepseek-v4-flash (same as coordinator)
• Expected: z-ai/glm-5.2:free

In both cases, the spawned session's journal (~/.jcode/sessions/session_*.journal.jsonl) recorded "model":"deepseek-v4-flash", and the specified z-ai/glm-5.2:free was never used.

---

Expected Behavior

When model (and optionally provider) is passed to swarm spawn, the worker should start with the specified model.

The list_models help text states:

"configure agents.swarm_model with a bare model or route-pinned value"

This implies setting agents.swarm_model pins the worker model, but having spawn-time parameters ignored is clearly an inconsistency.

---

Actual Behavior

1. Ran swarm spawn model="z-ai/glm-5.2:free".
2. Checked the journal of the spawned agent.
3. Found that "model":"deepseek-v4-flash" (inherited from the coordinator) was used instead.

---

Impact

• Self-Review Failures: Intending to review with a different model fails silently, using the same model instead and rendering reviews meaningless.
• Breakdown of Role-Based Model Operations: Impossible to delegate tasks (e.g., Coder = minimax-m3:free / Reviewer = glm-5.2:free).
• Potential Cost Incidents: If using a paid model for the coordinator, all workers might unintentionally run on paid models.

---

Environment

• jcode: v0.81.6 (4f0f9819a)
• settings.json:

{
  "default_provider": "openrouter",
  "model": "openrouter/free"
}

• config.toml [agents]:

[agents]
swarm_spawn_mode = "inline"
swarm_strip_layout = "vertical"
memory_sidecar_enabled = true
memory_rerank_cadence = 3
memory_rerank_votes = 2
memory_rerank_min_agree = 2
memory_embedding_backend = "local"
swarm_max_concurrent_agents = 32

→ swarm_model not set (Could this be the cause? Though spawn parameters being ignored is a separate issue).

• agents.swarm_model is single-value only (multiple not allowed).
• Coordinator model: deepseek-v4-flash (via OpenCode Go? Evidence found in journal under "provider":"openai-compatible:opencode-go").

list_models Header (Relevant Output)

Current coordinator model: deepseek-v4-flash
No agents.swarm_model pin configured (workers inherit the coordinator's model).
(configure agents.swarm_model with a bare model or route-pinned value)

---

Possible Causes (Hypothesis)

1. Spawn API ignores the model parameter: The spawn tool schema might include it, but implementation is missing.
2. Hardcoded coordinator inheritance when agents.swarm_model is unset: Spawn-side models only take effect if swarm_model is configured globally.
3. Model name normalization issue: z-ai/glm-5.2:free fails to resolve in the backend, triggering a fallback.

---

Suggested Fixes

1. Ensure spawn model / provider parameters are respected (Highest priority).
2. Prioritize spawn-parameter models even if agents.swarm_model is unset.
3. Documentation: Either document the limitation that models cannot be overridden at spawn time, or fix the implementation.

---

Workaround (Current)

• Set swarm_model = "z-ai/glm-5.2:free" under [agents] in ~/.jcode/config.toml (unverified, though suggested by list_models help text).
• However, this forces a single fixed model globally, making role-based separation impossible.

---

Attachments

• Reproduction session journal: session_hog_1788536706117_67b74ae9a3cf5916 (Experiment 1)
• Reproduction session journal: session_kitten_1788537002476_f9127c22cd55af65 (Experiment 2)
  (Actual measured values are logged in the "model" field of these files)

Start by tracing the swarm spawn model-selection path and compare it with the agents.swarm_model handling described by list_models. Reproduce both model forms from the issue, then inspect the referenced session journals and confirm that the requested model and provider are recorded instead of deepseek-v4-flash; done means both spawn requests select their specified models.