[Bug Report]: Foreground subagent timeout defaults to 120s despite schema claiming 'no default timeout'

Bug Report

Title: [Bug Report]: Foreground subagent timeout defaults to 120s despite schema claiming "no default timeout"

What version of Kimi Code CLI are you running?

kimi, version 1.44.0

Which platform/provider were you using?

Kimi Code platform (also reproducible with other providers)

What platform is your computer?

Linux (WSL)

What issue are you seeing?

When the root agent calls the Agent tool to spawn a foreground subagent (e.g. subagent_type="explore"), it always passes timeout: 120, causing the subagent to be killed after exactly 120 seconds with:

Agent timed out after 120s.

However, the Agent tool's JSON schema explicitly states:

"Foreground: no default timeout (runs until completion)"

And the timeout field's default in the Pydantic model is null:

# tools/agent/__init__.py
timeout: int | None = Field(
    default=None,
    description="...Foreground: no default timeout (runs until completion)...",
)

This creates a contradiction: the code/schema say "no default timeout", but the model's behavior is to always supply timeout: 120, effectively making 120s the hard default.

Screenshots / Evidence

Agent tool call as rendered in the TUI:

description "排查NEKO浏览器端报错"
subagent_type "explore"
timeout 120          <-- model always supplies this
prompt > 用户在浏览器直接访问 N.E.K.O (http://<IP>:<PORT>/) 时，控制台报错: ...

Agent timed out (120s)

Error returned to root agent:

error: Agent timed out after 120s.
message: Agent timed out after 120s.

Root Cause Analysis (from local source inspection)

1. The Agent tool's timeout parameter has default=None in tools/agent/__init__.py.
2. The generated JSON schema also shows "default": null.
3. Yet the model consistently fills in 120 when calling the tool.

This suggests the model is receiving a tool definition (either via system prompt or server-side schema augmentation) that implies the default/recommended timeout is 120 seconds, overriding the null default from the local schema.

Notably, utils/aiohttp.py sets a _DEFAULT_TIMEOUT of 120s for HTTP requests, but this is unrelated to the Agent tool's task-level timeout parameter.

Impact

• Foreground subagents doing legitimate long-running work (deep codebase exploration, complex planning) are killed after 2 minutes.
• The root agent then re-does the same exploration work itself, wasting tokens and time.
• Users must manually work around this by somehow forcing larger timeouts, but there is no user-facing configuration for foreground agent timeouts.

Expected Behavior

One of the following should be true:

1. If "no default timeout" is the intended behavior: the model should not auto-fill timeout: 120 when the user hasn't specified one. Foreground subagents should run until completion (or until max_steps_per_turn is hit).

2. If a default timeout is intentional: it should match the CHANGELOG claim of "foreground default 10 min" (600s), not 120s. And this default should be user-configurable in ~/.kimi/config.toml (e.g. [agent] foreground_timeout_s = 600).

Suggested Fixes

Option A: Fix the model-facing schema/prompt
Ensure the tool definition presented to the model correctly reflects default: null (or no default), so the model stops auto-filling 120.

Option B: Add a user-configurable default
Add a config option in ~/.kimi/config.toml:

[agent]
foreground_timeout_s = 600  # default for foreground subagents

The Agent tool should fall back to this config value when timeout is not explicitly provided by the model.

Option C: Change the code default
Change default=None to default=600 (or 3600) in tools/agent/__init__.py so that even if the model doesn't specify, the subagent gets a reasonable timeout.

Related References

• CHANGELOG 1.28.0: "Agent tool gains an optional timeout parameter (foreground default 10 min, background default 15 min)"
• Issue #2368: mentions that "The timeout description in the Agent tool schema claims 'Foreground: no default timeout (runs until completion)'"

Additional Context

There is currently no [agent] section in config.toml and no environment variable to control this. The only timeout config is [background].agent_task_timeout_s, which only affects background subagents.

Start with tools/agent/init.py and inspect how the Agent schema and timeout value reach the model and foreground subagent. Compare this with utils/aiohttp.py, the CHANGELOG 1.28.0 entry, and the existing [background].agent_task_timeout_s configuration in config.toml. Done means the chosen timeout behavior, schema description, and any user configuration agree.