BYOK custom providers: CLI forces greedy sampling (temperature=0) causing reasoning-model degeneration and silent hangs on context overflow

[bmazzarol-bunnings]

Describe the bug

For BYOK / custom OpenAI-compatible providers (e.g. a small thinking model such as qwen-27b served by vLLM), CLI 1.0.81 and later send temperature: 0, top_p: 0.95, frequency_penalty: 0, presence_penalty: 0 in every chat-completion request. CLI 1.0.80 sends no sampling parameters at all, letting the server use its own defaults.

With temperature: 0 (greedy decoding) forced on a thinking model, the model degenerates into long, repetitive, pure-reasoning outputs. Those oversized reasoning outputs (10.6k–25.5k tokens in a single call, verified in session logs) push the request past the provider's hard max_prompt_tokens limit (131,072), triggering compaction + truncation — after which the client either hangs with no error and no further model call (22 minutes until the user killed the process) or ends the turn with a contentless reasoning-only finish_reason: stop response and goes idle with no user-visible signal.

Secondary symptom observed in the same sessions: during compaction / auto-mode utility calls, the CLI sends the internal model id gpt-5.4-nano instead of the configured BYOK model name to the custom endpoint, producing a silent 404.

This is a CLI regression, not a model problem: the same model and the same server run fine on 1.0.80.

Affected version

1.0.81 – 1.0.87 (regression introduced in 1.0.81; verified by wire capture of 1.0.80/1.0.81/1.0.82/1.0.86/1.0.87 — 1.0.80 is healthy)

Steps to reproduce the behavior

1. Configure a custom OpenAI-compatible provider (BYOK):

   COPILOT_PROVIDER_TYPE=openai
   COPILOT_PROVIDER_BASE_URL=<vLLM endpoint>/v1
   COPILOT_PROVIDER_API_KEY=<any>
   COPILOT_PROVIDER_MAX_PROMPT_TOKENS=131072
   COPILOT_PROVIDER_MAX_OUTPUT_TOKENS=8192
   COPILOT_MODEL=qwen-27b
   

2. Run a sustained multi-step task with CLI ≥ 1.0.81 (small tasks like "create fib.py with memoization, run it, verify output" complete fine — degeneration manifests under sustained load, i.e. over many hours / compactions).
3. Observe: repetitive pure-reasoning output; context crossing the 131,072 limit; then either a silent hang (no error, no further model call, for 22+ minutes) or a contentless reasoning-only turn ending the turn with no signal.

Minimal wire-level reproduction of the sampling regression (no long task needed): capture the request bodies of 1.0.80 vs 1.0.81 sent to the same endpoint for the same task. 1.0.80 sends no sampling parameters; 1.0.81 sends temperature: 0, top_p: 0.95, frequency_penalty: 0, presence_penalty: 0.

Expected behavior

• Custom / BYOK providers are not pinned to sampling parameters (1.0.80 behavior: none sent, server defaults apply) — or sampling parameters are configurable via provider settings. temperature: 0 is harmful for thinking models.
• Requests never exceed max_prompt_tokens: the compaction decision accounts for max_output_tokens headroom, so the CLI compacts before overflow rather than truncating after.
• After a truncation/compaction, the CLI issues the next model call or surfaces an error — it never sits silent.
• Contentless reasoning-only turns (large reasoning, 0 content, 0 tool calls, finish_reason: stop) are detected and surfaced or retried with adjusted parameters.
• Utility/compaction calls use the configured BYOK model name, not an internal id (no 404 for gpt-5.4-nano).

Additional context

Wire-level request diff (socket capture, identical task, identical model)

Task: "create fib.py with memoization, run it, verify output" — non-interactive mode, BYOK qwen-27b via an OpenAI-compatible vLLM endpoint.

1.0.80 request body (top-level keys):

model, messages, tools, stream, stream_options

No sampling parameters — server defaults apply.

1.0.87 request body (top-level keys):

model, messages, temperature: 0, top_p: 0.95, frequency_penalty: 0, presence_penalty: 0, tools, stream, stream_options

Version pinning — all five versions captured running the identical task against the identical endpoint:

CLI version	Top-level request keys (sampling params)
1.0.80	model, messages, tools, stream, stream_options — no sampling params
1.0.81	+ temperature: 0, top_p: 0.95, frequency_penalty: 0, presence_penalty: 0 (+ snippy: {enabled: false})
1.0.82	same as 1.0.81 (incl. snippy stub)
1.0.86	same as 1.0.81 (no snippy)
1.0.87	same as 1.0.81 (no snippy)

The regression was introduced in 1.0.81 — the same release that added the interrupted-session restore mechanism (#4673).

Both versions replay assistant reasoning text into the message history (standard for thinking models; not the differentiator).

Session log evidence (2026-09-22, all qwen-27b)

Session	CLI	What happened
26a9ef8c	1.0.80	475 model calls over 22 h, 11 compactions (max 118.5k, never over limit), 0 truncations, 0 hangs. Work completed, PR merged. Max reasoning call 9,754 tokens — productive.
8bdd58ac	1.0.87	User: "you are in a reasoning loop" (02:22)
dd3a60cd	1.0.87	User: "you stuck in a reasoning loop" (04:31); 6 compactions (up to 123.5k — near the 131k limit)
8278f5c4	1.0.87	User: "you are in a reasoning loop" (04:48). Compaction call 404'd once (gpt-5.4-nano does not exist — see below), then an empty 10,662-token pure-reasoning turn (finish_reason: stop, 0 content, 0 tool calls) at 04:58; client went silent for 25 min until user kill.
bf54a1ec	1.0.87	25,510-token degenerate reasoning call (04:56–05:05) → context 131,805 (> 131,072 limit) → compaction_start + truncation 131,805→122,893 at 05:05:21 → no compaction_complete, no further model call → 22-minute silent hang until user kill (05:27).

Causal chain

1. Forced temperature: 0 → greedy decoding → repetition collapse in the thinking model's reasoning trace.
2. Oversized reasoning outputs (and full reasoning replay in history) exceed the provider's hard prompt limit between compaction-threshold checks (threshold ≈ 113k, limit 131k; a single 25k-token call crosses it).
3. Context overflow → truncation. Client then either:
   • (a) hangs: compaction starts, truncation applied, but no compaction-complete and no next model call (bf54a1ec), or
   • (b) emits a contentless reasoning-only finish_reason: stop turn and goes idle with no user-visible signal (8278f5c4).

Secondary bug: utility model calls 404 for BYOK

During compaction (and for auto-mode utility calls) the CLI sends the internal model id gpt-5.4-nano instead of the BYOK model name to the custom endpoint:

model.turn_failed
  model: gpt-5.4-nano   (modelInfo: {id: "gpt-5.4-nano", name: "qwen-27b"})
  error: 404 The model `gpt-5.4-nano` does not exist.

The turn silently recovers (falls back to truncation), but the failure is invisible to the user. Related: #4445 (auto mode picks unavailable models).

Environment

• Operating system: Ubuntu 26.04 LTS, x86_64, bash
• Model server: vLLM 0.28.0 (OpenAI-compatible), max_model_len 262144
• Related issues: 1.0.81: session restore auto-continues work the user aborted, trapping loop-prone models #4673 (auto-restore mechanism from the same 1.0.81 release — fixed/disabled by default since 1.0.83), Auto mode sometimes picks impossible model #4445 (auto mode model availability)