Subagent spawn fails near context limit due to truncated tool-call JSON

[giveen]

Summary

When context cache is near full, Late can emit an oversized spawn_subagent tool call. The tool argument JSON appears to be truncated mid-string, and the backend (llama.cpp OpenAI-compatible server) returns HTTP 500 with a parse error. The subagent spawn fails instead of recovering gracefully.

Environment

1. OS: Ubuntu 24.04 LTS
2. Client: Late (planning flow with subagent spawning)
3. Backend: llama.cpp / llama-swap router
4. Endpoint: http://127.0.0.1/v1/chat/completions
5. Context state at failure: near max (n_tokens = 262143, truncated = 1)

Steps to Reproduce

1. Run a long planning session until context cache is nearly full.
2. Trigger a spawn_subagent call with a large multi-paragraph goal payload.
3. Let the model stream tool-call arguments.
4. Observe backend response and Late behavior.

Expected Behavior

1. Late should avoid sending oversized tool-call arguments when context is near limit.
2. If tool-call JSON is truncated or invalid, Late should recover gracefully (retry with compact args or ask for a smaller payload).
3. Error surfaced to user should be actionable.

Actual Behavior

1. Backend returns HTTP 500 due to malformed tool-call arguments JSON.
2. Subagent spawn fails.
3. Delegation step does not complete.

Error Excerpt

LLAMA-CPP ERROR:

prompt eval time = 221.03 ms / 22 tokens
eval time = 88800.20 ms / 1535 tokens
total time = 89021.23 ms / 1557 tokens
slot release: n_tokens = 262143, truncated = 1
got exception: Failed to parse tool call arguments as JSON: parse error ... invalid string: missing closing quote
POST /v1/chat/completions ... 500

Impact

1. Long-running sessions become unreliable right when delegation is needed most.
2. High-context workflows can fail depending on tool-call argument size.
3. Users lose time recovering from failed subagent spawns.

Suggested Fix Direction

1. Add a preflight context-budget check before requests with tools.
2. Enforce a maximum size for spawn_subagent goal payloads.
3. Add automatic retry on malformed tool-call argument errors with a compact re-prompt.
4. Surface backend error body in Late so the exact failure reason is visible.

Temporary Workarounds

1. Start a fresh session before spawning subagents when context is large.
2. Keep subagent goal text short and atomic.
3. Put long implementation details in a file and reference it instead of embedding large payloads.

[mlhher]

#53 for reference