fix(bedrock): stop sending reasoning effort none to GPT-6 Astra when /v1/messages disables thinking by kimnamu · Pull Request #42854 · BerriAI/litellm

kimnamu · 2026-09-24T01:43:17Z

Thanks for the /v1/messages thinking translation and for the per-model supports_none_reasoning_effort flags, which this reads

TLDR

Problem this solves:

GPT-6 Astra gets a 400 when /v1/messages disables thinking
litellm sends reasoning effort none, which Astra lacks
Bedrock Sol and Luna rows wrongly say no none

How it solves it:

Skip none when the target's row says unsupported
Bedrock GPT-6 Sol and Luna rows now allow none

	Before (`eb7eeb5419`)	After (`c6c70f68f9`)
`/v1/messages`, thinking disabled, GPT-6 Astra `bedrock/us.`, `bedrock/global.` and `bedrock_mantle/` in us-west-2, plain and streaming	400 from Bedrock	200, no effort sent
same, GPT-6 Sol, Luna and GPT-5.6 Sol on Converse, GPT-6 Sol on Mantle	200, `none` sent	unchanged
`/v1/messages`, thinking enabled or left out, every model above	200	unchanged
chat completions with explicit `reasoning_effort: "none"`, GPT-6 Astra	400 from Bedrock	unchanged
`/v1/responses`, `bedrock/us.openai.gpt-6-sol` or `bedrock/global.openai.gpt-6-luna`, effort `none` with `temperature: 0.2`	400 from litellm	200
same with effort `low`, and GPT-6 Astra with effort `low`	400 from litellm	unchanged

Details

Before mapping disabled thinking to none, the Anthropic adapter now checks the target's supports_none_reasoning_effort with the existing is_explicitly_disabled_factory, so only an explicit false skips it and unflagged targets keep none. This mirrors maybe_drop_disabled_thinking, which already drops disabled thinking for Claude models that always think. The Bedrock GPT-6 Sol and Luna rows from #42746 and #42798 copied Astra's false. They now say true, like OpenAI's gpt-6-sol and gpt-6-luna rows

# what reaches Bedrock, litellm.anthropic_interface.messages.create(..., thinking={"type": "disabled"}), before -> after
bedrock/us.openai.gpt-6-astra          /model/us.openai.gpt-6-astra/converse       reasoning {'effort': 'none'}, 400  ->  no reasoning field, ok
bedrock/global.openai.gpt-6-astra      /model/global.openai.gpt-6-astra/converse   reasoning {'effort': 'none'}, 400  ->  no reasoning field, ok
bedrock_mantle/openai.gpt-6-astra      /openai/v1/responses                        reasoning {'effort': 'none'}, 400  ->  no reasoning field, ok
bedrock/converse/us.openai.gpt-6-sol   /model/us.openai.gpt-6-sol/converse         reasoning {'effort': 'none'}, ok   ->  same
bedrock_mantle/openai.gpt-6-sol        /openai/v1/responses                        reasoning {'effort': 'none'}, ok   ->  same

# raw Bedrock, no litellm, runtime /openai/v1 (us-east-1) and Mantle (Sol and Luna in us-east-1, Astra in us-west-2)
reasoning effort none                      Sol 200  Luna 200  Astra 400 unsupported_value
reasoning effort none + temperature 0.2    Sol 200  Luna 200
reasoning effort low + temperature 0.2     Sol 400  Luna 400  Astra 400

# new test, adapter at eb7eeb5419 with the cost map change kept, then c6c70f68f9
$ pytest -p no:xdist -q tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py -k sends_effort_none_only
FAILED ...[bedrock-gpt-6-astra-global]
FAILED ...[bedrock-gpt-6-astra]
FAILED ...[mantle-gpt-6-astra]
E       AssertionError: assert 'none' == None
3 failed, 4 passed, 222 deselected
7 passed, 222 deselected
# adapter change with eb7eeb5419's cost map: the Sol and Luna cases fail instead, 3 failed, 4 passed

`pytest -n 4`	Before	After
`tests/test_litellm/llms/anthropic/experimental_pass_through`	979 passed	986 passed
`tests/test_litellm/router_utils`	632 passed	632 passed
`tests/test_litellm/llms/bedrock`	1445 passed	1445 passed
`tests/test_litellm/llms/bedrock_mantle`	227 passed	227 passed
`tests/test_litellm/test_model_prices_schema.py`, `test_cost_map_guard.py`	85 passed	85 passed
ruff strict, type discipline, test quality gates, `ci_cd/cost_map_guard.py`		OK
basedpyright on the adapter		no new diagnostics

User Flow

Before: GPT-6 Astra fails on /v1/messages with thinking disabled

POST http://localhost:4000/v1/messages, model gpt-6-astra, "thinking": {"type": "disabled"}
400, "Unsupported value: 'none' is not supported"

After: the same request is answered

Same POST
200, "pong"

Relevant issues

Fixes #42853

Pre-Submission checklist

Screenshots / Proof of Fix

Live Bedrock calls through a local proxy

# LITELLM_LOCAL_MODEL_COST_MAP=True python litellm/proxy/proxy_cli.py --config config.yaml --port 4000
model_list:
  - model_name: gpt-6-sol
    litellm_params: {model: bedrock/converse/us.openai.gpt-6-sol, aws_region_name: us-east-1}
  - model_name: gpt-6-luna
    litellm_params: {model: bedrock/converse/us.openai.gpt-6-luna, aws_region_name: us-east-1}
  - model_name: gpt-6-astra
    litellm_params: {model: bedrock/us.openai.gpt-6-astra, aws_region_name: us-east-1}
  - model_name: gpt-6-astra-global
    litellm_params: {model: bedrock/global.openai.gpt-6-astra, aws_region_name: us-east-1}
  - model_name: gpt-5.6-sol
    litellm_params: {model: bedrock/converse/us.openai.gpt-5.6-sol, aws_region_name: us-east-1}
  - model_name: mantle-gpt-6-sol
    litellm_params: {model: bedrock_mantle/openai.gpt-6-sol, aws_region_name: us-east-1}
  - model_name: mantle-gpt-6-astra
    litellm_params: {model: bedrock_mantle/openai.gpt-6-astra, aws_region_name: us-west-2}
  - model_name: gpt-6-sol-default
    litellm_params: {model: bedrock/us.openai.gpt-6-sol, aws_region_name: us-east-1}
  - model_name: gpt-6-luna-global
    litellm_params: {model: bedrock/global.openai.gpt-6-luna, aws_region_name: us-east-1}
general_settings:
  master_key: os.environ/LITELLM_MASTER_KEY

H=(-H "Authorization: Bearer $LITELLM_MASTER_KEY" -H "Content-Type: application/json")
U=http://localhost:4000
OFF='"thinking":{"type":"disabled"},'
ON='"thinking":{"type":"enabled","budget_tokens":1024},'
msgs() { curl -s -o out.json -w '%{http_code} ' $U/v1/messages "${H[@]}" -d "{\"model\":\"$1\",\"max_tokens\":2048,$2\"messages\":[{\"role\":\"user\",\"content\":\"Say pong.\"}]}"; \
  jq -c 'if .error then (.error.message | .[0:200]) else ([.content[]? | "\(.type) \(.text // "")"] + [.stop_reason]) end' out.json; }
stream_msgs() { curl -sN -o out.txt -w '%{http_code} ' $U/v1/messages "${H[@]}" -d "{\"model\":\"$1\",\"max_tokens\":2048,\"stream\":true,$OFF\"messages\":[{\"role\":\"user\",\"content\":\"Say pong.\"}]}"; \
  if grep -q '^data: ' out.txt; then sed -n 's/^data: //p' out.txt | jq -rj 'if .type=="content_block_delta" then (.delta.text // "") elif .type=="message_stop" then " [message_stop]" else "" end'; else jq -rj '.error.message | .[0:200]' out.txt; fi; echo; }
resp() { curl -s -o out.json -w '%{http_code} ' $U/v1/responses "${H[@]}" -d "{\"model\":\"$1\",\"input\":\"Say pong.\",\"reasoning\":{\"effort\":\"$2\"},\"temperature\":0.2}"; \
  jq -c 'if .error then (.error.message | .[0:150]) else [.output[]? | select(.type=="message") | .content[0].text] end' out.json; }

Before (`eb7eeb5`)

Messages, thinking disabled

Run

$ msgs gpt-6-astra "$OFF"
400 "litellm.BadRequestError: BedrockException - {\"message\":\"The model returned the following errors: {\\\"error\\\":{\\\"code\\\":\\\"unsupported_value\\\",\\\"message\\\":\\\"Unsupported value: 'none' is not supported wit"
$ msgs gpt-6-astra-global "$OFF"
400 "litellm.BadRequestError: BedrockException - {\"message\":\"The model returned the following errors: {\\\"error\\\":{\\\"code\\\":\\\"unsupported_value\\\",\\\"message\\\":\\\"Unsupported value: 'none' is not supported wit"
$ msgs mantle-gpt-6-astra "$OFF"
400 "litellm.BadRequestError: BedrockException - {\"error\":{\"code\":\"unsupported_value\",\"message\":\"Unsupported value: 'none' is not supported with the 'openai.gpt-6-astra' model. Supported values are: 'low',"
$ msgs gpt-6-sol "$OFF"
200 ["text pong","end_turn"]
$ msgs gpt-6-luna "$OFF"
200 ["text pong","end_turn"]
$ msgs gpt-5.6-sol "$OFF"
200 ["text pong","end_turn"]
$ msgs mantle-gpt-6-sol "$OFF"
200 ["text pong","end_turn"]

Messages streaming, thinking disabled

Run

$ stream_msgs gpt-6-astra
400 litellm.BadRequestError: BedrockException - {"message":"The model returned the following errors: {\"error\":{\"code\":\"unsupported_value\",\"message\":\"Unsupported value: 'none' is not supported wit
$ stream_msgs gpt-6-astra-global
400 litellm.BadRequestError: BedrockException - {"message":"The model returned the following errors: {\"error\":{\"code\":\"unsupported_value\",\"message\":\"Unsupported value: 'none' is not supported wit
$ stream_msgs mantle-gpt-6-astra
400 litellm.BadRequestError: BedrockException - {"error":{"code":"unsupported_value","message":"Unsupported value: 'none' is not supported with the 'openai.gpt-6-astra' model. Supported values are: 'low',
$ stream_msgs gpt-6-sol
200 pong [message_stop]

Messages, thinking enabled or left out

Run

$ msgs gpt-6-astra "$ON"
200 ["text pong","end_turn"]
$ msgs gpt-6-astra ""
200 ["text pong","end_turn"]
$ msgs gpt-6-astra-global "$ON"
200 ["text pong","end_turn"]
$ msgs gpt-6-astra-global ""
200 ["text pong","end_turn"]
$ msgs mantle-gpt-6-astra "$ON"
200 ["text pong","end_turn"]
$ msgs mantle-gpt-6-astra ""
200 ["text pong","end_turn"]
$ msgs gpt-6-sol "$ON"
200 ["text pong","end_turn"]
$ msgs gpt-6-sol ""
200 ["text pong","end_turn"]
$ msgs gpt-6-luna "$ON"
200 ["text pong","end_turn"]
$ msgs gpt-6-luna ""
200 ["text pong","end_turn"]
$ msgs gpt-5.6-sol "$ON"
200 ["text pong","end_turn"]
$ msgs gpt-5.6-sol ""
200 ["text pong","end_turn"]
$ msgs mantle-gpt-6-sol "$ON"
200 ["text pong","end_turn"]
$ msgs mantle-gpt-6-sol ""
200 ["text pong","end_turn"]

Responses, effort with temperature 0.2

Run

$ resp gpt-6-sol-default none
400 "litellm.UnsupportedParamsError: us.openai.gpt-6-sol doesn't support temperature=0.2 while reasoning is active. Only temperature=1 is supported unless "
$ resp gpt-6-luna-global none
400 "litellm.UnsupportedParamsError: global.openai.gpt-6-luna doesn't support temperature=0.2 while reasoning is active. Only temperature=1 is supported un"
$ resp gpt-6-sol-default low
400 "litellm.UnsupportedParamsError: us.openai.gpt-6-sol doesn't support temperature=0.2 while reasoning is active. Only temperature=1 is supported unless "
$ resp gpt-6-astra low
400 "litellm.UnsupportedParamsError: us.openai.gpt-6-astra doesn't support temperature=0.2 while reasoning is active. Only temperature=1 is supported unles"

After (`c6c70f6`)

Messages, thinking disabled

Run

$ msgs gpt-6-astra "$OFF"
200 ["text pong","end_turn"]
$ msgs gpt-6-astra-global "$OFF"
200 ["text pong","end_turn"]
$ msgs mantle-gpt-6-astra "$OFF"
200 ["text pong","end_turn"]
$ msgs gpt-6-sol "$OFF"
200 ["text pong","end_turn"]
$ msgs gpt-6-luna "$OFF"
200 ["text pong","end_turn"]
$ msgs gpt-5.6-sol "$OFF"
200 ["text pong","end_turn"]
$ msgs mantle-gpt-6-sol "$OFF"
200 ["text pong","end_turn"]

Messages streaming, thinking disabled

Run

$ stream_msgs gpt-6-astra
200 pong [message_stop]
$ stream_msgs gpt-6-astra-global
200 pong [message_stop]
$ stream_msgs mantle-gpt-6-astra
200 pong [message_stop]
$ stream_msgs gpt-6-sol
200 pong [message_stop]

Messages, thinking enabled or left out

Run

$ msgs gpt-6-astra "$ON"
200 ["text pong","end_turn"]
$ msgs gpt-6-astra ""
200 ["text pong","end_turn"]
$ msgs gpt-6-astra-global "$ON"
200 ["text pong","end_turn"]
$ msgs gpt-6-astra-global ""
200 ["text pong","end_turn"]
$ msgs mantle-gpt-6-astra "$ON"
200 ["text pong","end_turn"]
$ msgs mantle-gpt-6-astra ""
200 ["text pong","end_turn"]
$ msgs gpt-6-sol "$ON"
200 ["text pong","end_turn"]
$ msgs gpt-6-sol ""
200 ["text pong","end_turn"]
$ msgs gpt-6-luna "$ON"
200 ["text pong","end_turn"]
$ msgs gpt-6-luna ""
200 ["redacted_thinking ","text pong","end_turn"]
$ msgs gpt-5.6-sol "$ON"
200 ["text pong","end_turn"]
$ msgs gpt-5.6-sol ""
200 ["text pong","end_turn"]
$ msgs mantle-gpt-6-sol "$ON"
200 ["text pong","end_turn"]
$ msgs mantle-gpt-6-sol ""
200 ["text pong","end_turn"]

Responses, effort with temperature 0.2

Run

$ resp gpt-6-sol-default none
200 ["pong"]
$ resp gpt-6-luna-global none
200 ["pong"]
$ resp gpt-6-sol-default low
400 "litellm.UnsupportedParamsError: us.openai.gpt-6-sol doesn't support temperature=0.2 while reasoning is active. Only temperature=1 is supported unless "
$ resp gpt-6-astra low
400 "litellm.UnsupportedParamsError: us.openai.gpt-6-astra doesn't support temperature=0.2 while reasoning is active. Only temperature=1 is supported unles"

Type

🐛 Bug Fix

Caveats (if any)

Medium

Disabled thinking skips none for 22 other flag-false rows too
- Mostly OpenAI and Azure gpt-5 variants, not run live

Low

Astra still reasons at default when thinking is disabled

Final Attestation

The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

…/v1/messages disables thinking The /v1/messages bridge turns Anthropic thinking {"type": "disabled"} into reasoning_effort "none" for every non-Claude target. Bedrock GPT-6 Astra has no "none" level and answers 400, although its map rows already say supports_none_reasoning_effort: false. Skip "none" when the target's row explicitly disables it, so the request runs at the model's default effort The Bedrock GPT-6 Sol and Luna rows copied Astra's false flag, but both take "none" on Converse and on the runtime and Mantle OpenAI endpoints, as OpenAI's own gpt-6-sol and gpt-6-luna rows say. Set it to true there so they keep getting "none"

greptile-apps · 2026-09-24T01:45:53Z

Confidence Score: 5/5

The PR appears safe to merge, with the compatibility check scoped to translated disabled thinking and backed by targeted model coverage

Summary

This PR prevents Anthropic Messages requests with disabled thinking from forwarding unsupported reasoning_effort: "none" to models that explicitly reject it

Uses model capability metadata rather than hardcoded model names
Corrects GPT-6 Sol and Luna capability rows in both synchronized cost maps
Adds focused regression coverage for Astra, Sol, Luna, Converse aliases, Mantle, and models without an explicit flag

_{Reviews (1) · Last reviewed commit: "fix(bedrock): stop sending reasoning eff..."}

codspeed · 2026-09-24T01:52:36Z

Merging this PR will not alter performance

✅ 31 untouched benchmarks

_{Comparing kimnamu:bedrock_gpt_6_astra_thinking_disabled (5579ce9) with main (7a7d27c)}

codecov · 2026-09-24T01:55:24Z

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

lets-order-some-fries

I built a reproduction for #42853 at main 571ada0b0f before this PR and re-ran it at head c6c70f68f9 (offline; adapter plus Converse param mapping only, no provider calls). Before:

bedrock/us.openai.gpt-6-astra   cost_map_flag=False supports_none()=False adapter->reasoning_effort='none' converse_wire={'effort': 'none'}

after:

bedrock/us.openai.gpt-6-astra   cost_map_flag=False supports_none()=False adapter->reasoning_effort=None converse_wire=None

all 3 of 3 Astra targets. bedrock/converse/us.openai.gpt-5.6-sol (cost_map_flag=None) still gets 'none'.

That looks like the root cause, not a patch over it: the gate reads the capability flag where disabled thinking is translated, litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py:1168-1169, and is_explicitly_disabled_factory rather than the boolean helper is what leaves the unflagged row alone.

Two questions:

azure/gpt-6-astra, azure/us/gpt-6-astra and azure_ai/gpt-6-astra carry supports_none_reasoning_effort: true while openai/gpt-6-astra carries false; at your head I still get reasoning_effort='none' for all three. Do those rows need the same correction? I have no Azure keys, so I cannot say which value is right.
reasoning_effort_from_thinking_budget (litellm/litellm_core_utils/reasoning_effort_utils.py:20-26) floors at "minimal" one branch above the new check, and every gpt-6/gpt-5.6 row across all five providers sets supports_minimal_reasoning_effort: false. With budget_tokens 0/512/1023 I get {'effort': 'minimal'} on the wire. Does minimal need gating too, as VertexGeminiConfig._supports_minimal_thinking_level does?

Tests at your head: the adapter transformation file, 229 passed, 0 failed; test_model_prices_schema.py plus test_cost_map_guard.py, 85 passed, 0 failed. I did not run the wider bedrock/ or router_utils/ suites the PR body quotes.

kimnamu · 2026-09-25T23:37:55Z

Thanks @lets-order-some-fries for rebuilding the repro at both hashes. As far as I can tell neither question needs a change here

Question	What I found	Checked with
Azure and Foundry Astra rows allow `none`	Measured on Foundry in #39827 and #39983, which list `none` as supported there. OpenAI's API rejects it for Astra, hence the `false` on OpenAI's row	Their PR bodies, no Azure keys here either
`minimal` below a 1024 budget	Already handled: `_normalize_reasoning_effort` (`adapters/handler.py:587`) runs after the adapter and lowers it to `low` from `supports_minimal_reasoning_effort`. It skips `none` by design, so that one needs the adapter check	Output below, the 36 live calls here all returned 200

# c6c70f68f9, /v1/messages with thinking {"type": "enabled", "budget_tokens": 512}
# adapter alone, then the handler step, then a live call with the outgoing body captured
bedrock/us.openai.gpt-6-astra           512   adapter='minimal' handler='low'  live=200 sent={'reasoning': {'effort': 'low'}}
bedrock/converse/us.openai.gpt-6-sol    512   adapter='minimal' handler='low'  live=200 sent={'reasoning': {'effort': 'low'}}
bedrock/converse/us.openai.gpt-6-luna   512   adapter='minimal' handler='low'  live=200 sent={'reasoning': {'effort': 'low'}}
bedrock/converse/us.openai.gpt-5.6-sol  512   adapter='minimal' handler='low'  live=200 sent={'reasoning': {'effort': 'low'}}
bedrock_mantle/openai.gpt-6-sol         512   adapter='minimal' handler='low'  live=200 sent={'reasoning': {'effort': 'low'}}
bedrock_mantle/openai.gpt-6-astra       512   adapter='minimal' handler='low'  live=200 sent={'reasoning': {'effort': 'low'}}
normalize_reasoning_effort_value('none', 'bedrock/us.openai.gpt-6-astra') = 'none'
normalize_reasoning_effort_value('none', 'bedrock_mantle/openai.gpt-6-astra') = 'none'

Did you see minimal on a path that skips that handler? Happy to look if so

Budgets 0 and 1023, other providers, and a local proxy run

# same run, budget_tokens 0 and 1023
bedrock/us.openai.gpt-6-astra           0     adapter='minimal' handler='low'  live=200 sent={'reasoning': {'effort': 'low'}}
bedrock/us.openai.gpt-6-astra           1023  adapter='minimal' handler='low'  live=200 sent={'reasoning': {'effort': 'low'}}
bedrock/converse/us.openai.gpt-6-sol    0     adapter='minimal' handler='low'  live=200 sent={'reasoning': {'effort': 'low'}}
bedrock/converse/us.openai.gpt-6-sol    1023  adapter='minimal' handler='low'  live=200 sent={'reasoning': {'effort': 'low'}}
bedrock/converse/us.openai.gpt-6-luna   0     adapter='minimal' handler='low'  live=200 sent={'reasoning': {'effort': 'low'}}
bedrock/converse/us.openai.gpt-6-luna   1023  adapter='minimal' handler='low'  live=200 sent={'reasoning': {'effort': 'low'}}
bedrock/converse/us.openai.gpt-5.6-sol  0     adapter='minimal' handler='low'  live=200 sent={'reasoning': {'effort': 'low'}}
bedrock/converse/us.openai.gpt-5.6-sol  1023  adapter='minimal' handler='low'  live=200 sent={'reasoning': {'effort': 'low'}}
bedrock_mantle/openai.gpt-6-sol         0     adapter='minimal' handler='low'  live=200 sent={'reasoning': {'effort': 'low'}}
bedrock_mantle/openai.gpt-6-sol         1023  adapter='minimal' handler='low'  live=200 sent={'reasoning': {'effort': 'low'}}
bedrock_mantle/openai.gpt-6-astra       0     adapter='minimal' handler='low'  live=200 sent={'reasoning': {'effort': 'low'}}
bedrock_mantle/openai.gpt-6-astra       1023  adapter='minimal' handler='low'  live=200 sent={'reasoning': {'effort': 'low'}}

# offline, the helper that both the chat and the Responses /v1/messages handlers call
normalize_reasoning_effort_value('minimal', 'gpt-6-astra') = 'low'
normalize_reasoning_effort_value('minimal', 'gpt-6-sol') = 'low'
normalize_reasoning_effort_value('minimal', 'azure/gpt-6-astra') = 'low'
normalize_reasoning_effort_value('minimal', 'azure/gpt-6-sol') = 'low'
normalize_reasoning_effort_value('minimal', 'azure_ai/gpt-6-astra') = 'low'
normalize_reasoning_effort_value('minimal', 'bedrock/us.openai.gpt-6-astra') = 'low'
normalize_reasoning_effort_value('minimal', 'bedrock/converse/us.openai.gpt-5.6-sol') = 'low'
normalize_reasoning_effort_value('minimal', 'bedrock_mantle/openai.gpt-6-luna') = 'low'

# proxy at c6c70f68f9 with a subset of the PR body's deployments and no drop_params, H as in the PR body
msgs() { curl -s -o out.json -w '%{http_code} ' $U/v1/messages "${H[@]}" -d "{\"model\":\"$1\",\"max_tokens\":2048,\"thinking\":{\"type\":\"enabled\",\"budget_tokens\":$2},\"messages\":[{\"role\":\"user\",\"content\":\"Say pong.\"}]}"; \
  jq -c 'if .error then (.error.message | .[0:200]) else ([.content[]? | "\(.type) \(.text // "")"] + [.stop_reason]) end' out.json; }
$ for m in gpt-6-astra gpt-6-sol gpt-6-luna gpt-5.6-sol mantle-gpt-6-sol mantle-gpt-6-astra; do for b in 0 512 1023; do echo "\$ msgs $m $b"; msgs $m $b; done; done > out.txt
$ paste -d' ' - - < out.txt | sed 's/^\$ msgs //'
gpt-6-astra 0 200 ["text pong","end_turn"]
gpt-6-astra 512 200 ["text pong","end_turn"]
gpt-6-astra 1023 200 ["text pong","end_turn"]
gpt-6-sol 0 200 ["text pong","end_turn"]
gpt-6-sol 512 200 ["text pong","end_turn"]
gpt-6-sol 1023 200 ["text pong","end_turn"]
gpt-6-luna 0 200 ["text pong","end_turn"]
gpt-6-luna 512 200 ["text pong","end_turn"]
gpt-6-luna 1023 200 ["text pong","end_turn"]
gpt-5.6-sol 0 200 ["text pong","end_turn"]
gpt-5.6-sol 512 200 ["text pong","end_turn"]
gpt-5.6-sol 1023 200 ["text pong","end_turn"]
mantle-gpt-6-sol 0 200 ["text pong","end_turn"]
mantle-gpt-6-sol 512 200 ["text pong","end_turn"]
mantle-gpt-6-sol 1023 200 ["text pong","end_turn"]
mantle-gpt-6-astra 0 200 ["text pong","end_turn"]
mantle-gpt-6-astra 512 200 ["text pong","end_turn"]
mantle-gpt-6-astra 1023 200 ["text pong","end_turn"]

…hinking_disabled

lets-order-some-fries · 2026-09-26T17:57:45Z

My review was wrong about minimal: the repro called the adapter and the Converse param mapping directly, so it never went through adapters/handler.py:587, and /v1/messages does not send it.

Re-run at ebcb314a19 through the public entry points, offline, with the request body captured at the HTTP client, for bedrock/us.openai.gpt-6-astra, bedrock/converse/us.openai.gpt-6-sol and bedrock_mantle/openai.gpt-6-astra. Effort sent:

entry point	`budget_tokens` 0 / 512 / 1023	thinking disabled, both Astra targets
`litellm.anthropic_interface.messages.create` / `acreate`	`low`	not sent
`litellm.adapter_completion` / `aadapter_completion` with `AnthropicAdapter`	`minimal`	not sent

On your question, adapter_completion does skip the handler: it translates and goes straight to completion() (litellm/main.py:7712-7714, async at 7673-7675). It only applies when AnthropicAdapter is registered in litellm.adapters, which is empty by default (litellm/__init__.py:1503), and it sends minimal at main 14f4c34c61 too, so it is outside this PR. Your check sits in the shared adapter, so it covers that path as well: at main both entry points send none to Astra, at ebcb314a19 neither does.

The Azure answer holds: #39827 quotes Foundry's own error listing 'none' as a supported value for Astra, and #39983 shows none with temperature returning 200 on the azure_ai/ route.

This looks right to me. The adapter transformation test file passes at ebcb314a19 (234 passed).

…hinking_disabled

kimnamu requested a review from a team September 24, 2026 01:43

kimnamu requested review from kerry-berri, mateo-berri and ryan-crabbe-berri as code owners September 24, 2026 01:43

devin-ai-integration Bot mentioned this pull request Sep 24, 2026

fix(cost-map): retirement dates, chatgpt reasoning flags, bing pricing, bedrock mantle and mythos, azure gpt-5.6 alias, anthropic batch rates, new nebius, openrouter and xai rows #42951

Merged

5 tasks done

lets-order-some-fries reviewed Sep 25, 2026

View reviewed changes

Merge remote-tracking branch 'origin/main' into bedrock_gpt_6_astra_t…

ebcb314

…hinking_disabled

mateo-berri mentioned this pull request Sep 26, 2026

fix(ci): fail the cost map guard when a Bedrock regional row drifts from its base row #43363

Open

5 of 7 tasks

Merge remote-tracking branch 'origin/main' into bedrock_gpt_6_astra_t…

5579ce9

…hinking_disabled

Aasif-Simpplr mentioned this pull request Oct 6, 2026

fix(bedrock): honor disabled reasoning efforts on native chat route #44856

Open

6 tasks done

Navigation Menu

Repository navigation

fix(bedrock): stop sending reasoning effort none to GPT-6 Astra when /v1/messages disables thinking - #42854

fix(bedrock): stop sending reasoning effort none to GPT-6 Astra when /v1/messages disables thinking#42854
kimnamu wants to merge 3 commits into
BerriAI:mainfrom
kimnamu:bedrock_gpt_6_astra_thinking_disabled

kimnamu commented Sep 24, 2026

Uh oh!

greptile-apps Bot commented Sep 24, 2026

Uh oh!

codspeed Bot commented Sep 24, 2026 •

edited

Loading

Uh oh!

codecov Bot commented Sep 24, 2026 •

edited

Loading

Uh oh!

lets-order-some-fries left a comment

Uh oh!

kimnamu commented Sep 25, 2026

Uh oh!

lets-order-some-fries commented Sep 26, 2026

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

kimnamu commented Sep 24, 2026

TLDR

User Flow

Relevant issues

Pre-Submission checklist

Screenshots / Proof of Fix

Before (eb7eeb5)

Messages, thinking disabled

Messages streaming, thinking disabled

Messages, thinking enabled or left out

Responses, effort with temperature 0.2

After (c6c70f6)

Messages, thinking disabled

Messages streaming, thinking disabled

Messages, thinking enabled or left out

Responses, effort with temperature 0.2

Type

Caveats (if any)

Medium

Low

Final Attestation

Uh oh!

greptile-apps Bot commented Sep 24, 2026

Confidence Score: 5/5

Summary

Uh oh!

codspeed Bot commented Sep 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Merging this PR will not alter performance

Uh oh!

codecov Bot commented Sep 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

lets-order-some-fries left a comment

Choose a reason for hiding this comment

Uh oh!

kimnamu commented Sep 25, 2026

Uh oh!

lets-order-some-fries commented Sep 26, 2026

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Before (`eb7eeb5`)

After (`c6c70f6`)

codspeed Bot commented Sep 24, 2026 •

edited

Loading

codecov Bot commented Sep 24, 2026 •

edited

Loading