Repository navigation
Model "auto" always runs with reasoning effort disabled (reasoningEffort: null), and rejects any attempt to configure one #4560
Description
Describe the bug
When the model is set to auto, every request is sent without any reasoning effort:
- Selecting
autosets the session'sreasoningEfforttonull(visible in the session'sevents.jsonlassession.model_change → "reasoningEffort": null). - The auto router (
auto_v2) picks a model per prompt (session.auto_mode_resolved) but never assigns a reasoning effort to it — the request goes out with none, confirmed byreasoning_effort = NULLandreasoning_tokens = 0in~/.copilot/session-store.db(assistant_usage_events). - Trying to set one explicitly is rejected:
copilot --model auto --effort xhighfails withError: Model "auto" does not support reasoning effort configuration (requested: "xhigh").
So with auto there is no way at all to run with reasoning enabled. The degradation is silent — no error, no indication in the UI — and the impact on agentic work is severe: on a large agentic prompt (spec generation with mandatory file reads), the routed model (gpt-5.3-codex) ends its turn without making a single tool call, replying that it "could not perform the investigation". The exact same prompt, on the exact same routed model, works flawlessly when a reasoning effort is set (see "Same model, only effort differs" below).
Affected version
GitHub Copilot CLI 1.0.80.
Steps to reproduce the behavior
# 1) auto → no reasoning (silent)
copilot --model auto --allow-all-tools -p "Reply with exactly: ok"
# 2) control: explicit model → reasoning works
copilot --model gpt-5.4 --allow-all-tools -p "Reply with exactly: ok"
# 3) auto + explicit effort → hard error
copilot --model auto --effort xhigh --allow-all-tools -p "Reply with exactly: ok"Then inspect what was actually sent (replace the session ids printed in the Resume line):
sqlite3 -readonly ~/.copilot/session-store.db \
"SELECT session_id, model, reasoning_effort, reasoning_tokens
FROM assistant_usage_events ORDER BY id DESC LIMIT 5;"Observed results:
| Run | Routed/used model | reasoning_effort |
reasoning_tokens |
|---|---|---|---|
1 — --model auto |
gpt-5.4-mini (via auto_v2) |
NULL | 0 |
2 — --model gpt-5.4 |
gpt-5.4 |
xhigh |
21 |
3 — --model auto --effort xhigh |
— | — | Error: Model "auto" does not support reasoning effort configuration (requested: "xhigh"). |
The same happens in interactive mode: switching to auto logs
session.model_change ... "previousReasoningEffort": "xhigh", "reasoningEffort": null
in ~/.copilot/session-state/<session-id>/events.jsonl, and every subsequent session.auto_mode_resolved runs the chosen model with no effort.
Same model, only effort differs — night and day
Real-world case that led us here (a large spec-generation prompt, ~19.6k input tokens, requiring file reads and skill loading):
auto→ routed togpt-5.3-codex,reasoning_effort = NULL: the model ends the turn after 158 output tokens, 0 reasoning tokens, 0 tool calls, claiming it could not perform the investigation. Reproduced consistently across several sessions.- Same prompt, same
gpt-5.3-codex, withreasoning_effort = xhigh: normal agentic behavior — tool calls, thousands of reasoning tokens, task completed.
The only variable is the reasoning effort. auto mode currently guarantees the degraded path.
Expected behavior
auto should never silently disable reasoning. Either:
- the auto router assigns an appropriate reasoning effort for the routed model on each request (clamping to what that model supports), and/or
- a user-provided effort (
--effort, settingseffortLevel, session config) is honored underautoand clamped per routed model — instead of being cleared on switch and rejected on the CLI flag.
Additional context
- This looks like an overcorrection of Auto model execution failure due to reasoning level #2870: there, switching to
autokept the previous effort and caused loud400errors on models that don't support it; the fix apparently clears the effort entirely, trading a loud error for a silent, permanent degradation. Auto model execution failure due to reasoning level #4459 and Auto mode sometimes picks impossible model #4445 show the invalid-combination side of this is still not fully solved either — both symptoms share the same root: the auto router does not manage effort per routed model. - The ACP server mode (
copilot --acp) is affected identically, with no workaround at all:session/set_config_optionwithmodel=autoclears the effort and removes the reasoning-effort config option from the session, so ACP clients cannot restore it.
Activity
douglasjunior commentedon Aug 25, 2026
Follow-up with a controlled experiment that isolates the failure mode precisely and explains why
autosometimes appears to work.I replayed the exact same real-world agentic prompt (the ~19.6k-token spec-generation prompt mentioned in the issue — it requires reading project files before producing an artifact) several times on v1.0.80, same machine, same day, changing only model/effort/surface:
copilot -pwith--model gpt-5.3-codex --effort nonenonecopilot -pwith--model gpt-5.3-codex(effort from settings)xhighcopilot -pwith--model autogpt-5.4-miniNULLreasoning_tokens = 0on every requestcopilot --acp(ACP server) with model set toauto["gpt-5.3-codex"]→ codexNULLTwo conclusions:
The failure mode is "reasoning-first model + no reasoning effort", and
autoalways produces the second half of that combination.gpt-5.3-codex --effort nonereproduces the failure deterministically with noautoinvolved at all; the same model with an effort set handles the same prompt fine. Sinceautoalways runs its routed model withreasoningEffort: null, every time the router picksgpt-5.3-codexthe collapse is guaranteed.Whether
auto"works" is luck of the router's candidate pool, not of the surface or the prompt. Rows 3 and 4 are the same prompt, minutes apart: in non-interactive mode the pool includedgpt-5.4-mini(which tolerates the no-reasoning regime) and the task succeeded; in the ACP server the pool contained onlygpt-5.3-codex(session.auto_mode_resolved → "availableModels": ["gpt-5.3-codex"]), soautofailed deterministically — reproduced across 6+ sessions on different days.One more data point on the silent part: even when the no-effort routed model limps through, the output is measurably degraded. A planning task that normally triggers a thorough investigation completed with only 2 tool calls and
reasoning_tokens = 0across 7 API rounds — protocol-wise a "success", with nothing in the UI indicating that reasoning was disabled the whole time.So beyond avoiding invalid model+effort combinations (#2870, #4459, #4445), the router needs to assign a valid reasoning effort to the routed model — or honor the user's configured effort, clamped to what that model supports. Clearing it entirely turns
autointo "always the worst version of whatever model gets picked".patniko commentedon Oct 26, 2026
Thanks for the report. We've shipped fixes in this area since this was filed (most recently in Copilot CLI v1.0.81), and we believe this may now be resolved. Could you retest on the latest release (Copilot CLI v1.0.103) and let us know whether you still see it? If we don't hear back in a couple of weeks, we may close this issue — you can always reopen it.
Triaged as part of a Copilot SDK backlog review, with help from GitHub Copilot.