Repository navigation

Fix/pagi 125 release review findings - #1

Merged

asdek merged 816 commits into

main-vxcontrolfrom

fix/PAGI-125-release-review-findings

Sep 28, 2026

This pull request is big! We're only showing the most recent 250 commits

Commits on Sep 10, 2026
- merge: keep min_p off OpenAI's own endpoint (PAGI-294)
  
  sirozha
  and
  claude
  committed
  a7381e0
- fix(llms): give the extra body its own field
  
  sirozha
  and
  claude
  committed
  33b8381
- merge: give the extra body its own field (PAGI-293)
  
  sirozha
  and
  claude
  committed
  ed92d8a
- fix(googleai): stop sending thinking to families that do not think
  
  sirozha
  and
  claude
  committed
  fe7ff17
- merge: stop sending thinking to families that do not think (PAGI-292)
  
  sirozha
  and
  claude
  committed
  488bff0
- test(googleai): compare the thinking numbers as numbers
  
  sirozha
  and
  claude
  committed
  7290377
- merge: compare the thinking numbers as numbers (PAGI-295)
  
  sirozha
  and
  claude
  committed
  efecab6
- fix(llms): let an adaptive request without an effort take the vendor default
  
  sirozha
  and
  claude
  committed
  a459111
- merge: let an adaptive request without an effort take the vendor default (PAGI-296)
  
  sirozha
  and
  claude
  committed
  e56a538
- merge: the caller learns which options never reached the vendor (PAGI-133)
  
  sirozha
  and
  claude
  committed
  7e2ee31
- fix(reasoning): give the ollama door its own disable wire
  
  sirozha
  and
  claude
  committed
  7e4748c
- merge: the ollama door answers about its own disable wire (PAGI-133)
  
  sirozha
  and
  claude
  committed
  e76287c
Commits on Sep 11, 2026
- fix(embeddings): refuse a batch the client answered with fewer vectors
  
  sirozha
  and
  claude
  committed
  4136888
- fix(voyageai): answer a vectorless response with an error, not a panic
  
  sirozha
  and
  claude
  committed
  2a15d56
- fix(bedrock): answer a vectorless response with an error, not a panic
  
  sirozha
  and
  claude
  committed
  70f0e12
- fix(llms): drop the partial slice when the embedding answer is short
  
  sirozha
  and
  claude
  committed
  abd8519
- merge: embedding answers that are empty or short stop being panics and partials (PAGI-306)
  
  sirozha
  and
  claude
  committed
  5d001cf
- fix(huggingface): send the caller's output cap, an explicit zero and nothing else
  
  sirozha
  and
  claude
  committed
  2f5e016
- merge: the huggingface door sends the cap it was given and refuses what it cannot send (PAGI-311)
  
  sirozha
  and
  claude
  committed
  4c27cb2
- fix(reasoning): stop calling Nova 2 Sonic a model that takes reasoningConfig
  
  sirozha
  and
  claude
  committed
  162c3fd
- fix(llms): report the sampling an OpenAI reasoning model refuses
  
  sirozha
  and
  claude
  committed
  d926315
- fix(reasoning): classify a fine-tuned model as its base model
  
  sirozha
  and
  claude
  committed
  203ab04
- fix(llms): decode tool call arguments exactly on googleai and ollama
  
  sirozha
  and
  claude
  committed
  ff1a98a
- merge: the catalogue, the hint and the tool-call wire agree with the vendor (PAGI-307)
  
  sirozha
  and
  claude
  committed
  cd6e582
- fix(llms): let a partial answer still answer the truncation question
  
  sirozha
  and
  claude
  committed
  ae8e357
- fix(fake): hand back the collected text when a stream consumer gives up
  
  sirozha
  and
  claude
  committed
  7445699
- merge: an interrupted answer keeps both signals, and the fakes keep the contract (PAGI-310)
  
  sirozha
  and
  claude
  committed
  02859bf
- fix(llms): keep an explicit off out of the adaptive hand-off
  
  sirozha
  and
  claude
  committed
  64226d5
- fix(openai): let a budget ride with tools, and hand the cost back as a number
  
  sirozha
  and
  claude
  committed
  cc3ef5e
- merge: the openai door keeps the caller's off, budget and cost (PAGI-308)
  
  sirozha
  and
  claude
  committed
  bf1b33c
- fix(vertex): take the credentials the caller names again
  
  sirozha
  and
  claude
  committed
  5471dd9
- merge: the vertex door authenticates with the caller's service account again (PAGI-309)
  
  sirozha
  and
  claude
  committed
  b0aa3b5
- fix(llms): read a token budget with one predicate, not two
  
  sirozha
  and
  claude
  committed
  9129336
- fix(llms): let the doors hand the depth to the vendor they were told to
  
  sirozha
  and
  claude
  committed
  bd34934
- fix(llms): say so when a door cannot carry the delegation
  
  sirozha
  and
  claude
  committed
  8c74336
- merge: an adaptive request without an effort reaches the vendor as one (PAGI-303)
  
  sirozha
  and
  claude
  committed
  a13116f
- fix(httprr): strip the gateway's headers from every recording, not one door
  
  sirozha
  and
  claude
  committed
  c0e72f1
- fix(bedrock): rewrite the recordings without the gateway's headers
  
  sirozha
  and
  claude
  committed
  93680bd
- merge: no recording in this repository carries the gateway's headers (PAGI-302)
  
  sirozha
  and
  claude
  committed
  6b84440
- fix(llms): keep the boundaries of a turn's thinking, and refuse an image the vendor cannot read
  
  sirozha
  and
  claude
  committed
  ab2eae3
- fix(bedrock): read the thinking both streams send, and hand over the salvaged call
  
  sirozha
  and
  claude
  committed
  9473715
- merge: a turn's thinking keeps its blocks across the bedrock and anthropic doors (PAGI-304)
  
  sirozha
  and
  claude
  committed
  6dcd807
- fix(llms): stop calling a schema that reached the wire a lost option
  
  sirozha
  and
  claude
  committed
  ce69695
- fix(bedrock): stop calling the caller's own "no tool" a lost tool list
  
  sirozha
  and
  claude
  committed
  23eb31d
- fix(bedrock): say so when a legacy payload has nowhere to put the thinking
  
  sirozha
  and
  claude
  committed
  b5a2882
- fix(bedrock): report the thinking mechanism the converse door picked, on every family
  
  sirozha
  and
  claude
  committed
  2c7d6da
- fix(bedrock): ask what the payload carries, not what the provider is
  
  sirozha
  and
  claude
  committed
  8dcf844
- refactor(llms): give every door one list of the options it never reads
  
  sirozha
  and
  claude
  committed
  7fb2a4c
- fix(anthropic): let the legacy path answer for what it drops
  
  sirozha
  and
  claude
  committed
  0bca786
- fix(bedrock): report the answer limit the legacy payload named for the caller
  
  sirozha
  and
  claude
  committed
  bfc8dec
- fix(llms): tell the caller what became of an effort on a budget model
  
  sirozha
  and
  claude
  committed
  9228f48
- fix(llms): say each lost option once
  
  sirozha
  and
  claude
  committed
  f49402b
- fix(llms): close three silences the parity work left behind
  
  sirozha
  and
  claude
  committed
  16cbeec
- merge: the warnings surface answers for what a door drops (PAGI-305)
  
  sirozha
  and
  claude
  committed
  f32d642
Commits on Sep 13, 2026
- feat(reasoning): hold every block of a response in its order, and read every stored shape
  
  sirozha
  and
  claude
  committed
  6f235af
- fix(anthropic): hand back every thinking block with its own signature, in its place
  
  sirozha
  and
  claude
  committed
  9109db3
- fix(bedrock): keep every reasoning block of a Converse turn, each in its place
  
  sirozha
  and
  claude
  committed
  44115da
- fix(bedrock): read the legacy Claude payload block by block, and put each back in its place
  
  sirozha
  and
  claude
  committed
  9bc70e5
- refactor(reasoning): let the block list carry encrypted thinking, and read the old field into it
  
  sirozha
  and
  claude
  committed
  5311089
- fix(bedrock): count a block's place from the start of the turn on both bedrock legs
  
  sirozha
  and
  claude
  committed
  ef2bb46
- refactor(llms): read a Claude answer in one pass, and keep streamed blocks in one map
  
  sirozha
  and
  claude
  committed
  ae8df68
- docs(bedrock): describe how reasoning blocks keep their signatures and places
  
  sirozha
  and
  claude
  committed
  fff7806
- docs(bedrock): name exactly which lone block keeps the classic fields
  
  sirozha
  and
  claude
  committed
  1a29adc
- merge: every reasoning block keeps its signature and its place (PAGI-304)
  
  sirozha
  and
  claude
  committed
  ee57698
- docs(reasoning): name which lone block FromBlocks keeps in the classic fields
  
  sirozha
  and
  claude
  committed
  e9ae11a
- merge: FromBlocks names which lone block keeps the classic fields (PAGI-304)
  
  sirozha
  and
  claude
  committed
  e3c31eb
- fix(reasoning): let Bedrock turn thinking off on Opus 5 and Sonnet 5
  
  sirozha
  and
  claude
  committed
  2a02688
- fix(openai): give DeepSeek back the reasoning of every assistant turn
  
  sirozha
  and
  claude
  committed
  6f03b03
- fix(openai): drop the temperature an OpenAI model refuses while it thinks
  
  sirozha
  and
  claude
  committed
  b43095c
- fix(googleai): stop sending sampling values nobody asked for
  
  sirozha
  and
  claude
  committed
  35c8901
- fix(openai): keep Kimi's fixed sampling off the wire
  
  sirozha
  and
  claude
  committed
  49a09c7
- fix(bedrock): refuse structured output for the Claude models Bedrock does not serve it for
  
  sirozha
  and
  claude
  committed
  82ad622
- fix(openai): refuse a JSON schema for DeepSeek and GLM before the request
  
  sirozha
  and
  claude
  committed
  61ac151
- fix(ollama): keep the format field off Ollama Cloud and refuse schemas there
  
  sirozha
  and
  claude
  committed
  ebb99f3
- fix(bedrock): lower xhigh and max where Bedrock does not serve them
  
  sirozha
  and
  claude
  committed
  cf4df89
- fix(openai): send Mistral only the effort it documents
  
  sirozha
  and
  claude
  committed
  5d4ad5a
- fix(anthropic): keep budget under max_tokens unless the model really interleaves
  
  sirozha
  and
  claude
  committed
  b89ada1
- fix(googleai): sign a Gemini 3 function call that arrives without a signature
  
  sirozha
  and
  claude
  committed
  dd890fd
- merge: requests the vendor documents as refused stop before the network (PAGI-133)
  
  sirozha
  and
  claude
  committed
  19df2c4
- fix(anthropic): send every tool result of a tool message
  
  sirozha
  and
  claude
  committed
  32c5a67
- fix(bedrock): drop the empty text block from a replayed legacy Claude turn
  
  sirozha
  and
  claude
  committed
  f5aaaa2
- merge: a replayed Claude turn keeps every tool result and no empty text (PAGI-322)
  
  sirozha
  and
  claude
  committed
  489f5e5
- fix(openai): give Kimi and GLM back the reasoning of every assistant turn too
  
  sirozha
  and
  claude
  committed
  1239856
- fix(openai): refuse a JSON schema for GLM in every spelling of its name
  
  sirozha
  and
  claude
  committed
  516af3a
- fix(googleai): keep a configured zero top_p or top_k on the wire
  
  sirozha
  and
  claude
  committed
  b51b009
- fix(ollama): treat a -cloud model on a local server as Ollama Cloud
  
  sirozha
  and
  claude
  committed
  010688e
- docs: correct three statements the pre-network work made false
  
  sirozha
  and
  claude
  committed
  d4d0e29
- merge: the review of the pre-network work leaves no spelling or turn behind (PAGI-133)
  
  sirozha
  and
  claude
  committed
  3bb6954
- fix(bedrock): drop an empty text block that carries a cache marker too
  
  sirozha
  and
  claude
  committed
  e44296f
- merge: an empty text block leaves the legacy payload even with a cache marker (PAGI-322)
  
  sirozha
  and
  claude
  committed
  8a736bf
- fix(openai): let GLM served by Mistral keep the schema Mistral documents for it
  
  sirozha
  and
  claude
  committed
  225a9fc
- test: pin four guards the second review found unguarded
  
  sirozha
  and
  claude
  committed
  c07fcc3
- merge: the second review leaves no spelling or guard behind (PAGI-133, PAGI-322)
  
  sirozha
  and
  claude
  committed
  1860585
- fix(openai): read Mistral's thinking chunks and send them back the way it takes them
  
  sirozha
  and
  claude
  committed
  4dad8a8
- merge: Mistral's thinking chunks read and replayed through the openai door (PAGI-323)
  
  sirozha
  and
  claude
  committed
  1ffd046
- fix(reasoning): count only Mistral's own spellings as served by Mistral
  
  sirozha
  and
  claude
  committed
  39bd660
- test: pin the prefixed spelling of a Mistral model that does not reason
  
  sirozha
  and
  claude
  committed
  5a23bb1
- test: pin joining the text chunks of one Mistral answer
  
  sirozha
  and
  claude
  committed
  3517d05
- merge: only Mistral's own spellings get the thinking-chunk replay (PAGI-326)
  
  sirozha
  and
  claude
  committed
  0a52486
- test: pin the moved cache marker walking past a redacted thinking block
  
  sirozha
  and
  claude
  committed
  f81a209
- merge: the moved cache marker skips redacted thinking under test (PAGI-327)
  
  sirozha
  and
  claude
  committed
  9c70cd5
- fix(reasoning): refuse any host prefix left after mistral/ in ServedByMistral
  
  sirozha
  and
  claude
  committed
  80f8220
- merge: a host prefix after mistral/ is never Mistral's (PAGI-328)
  
  sirozha
  and
  claude
  committed
  2af9445
Commits on Sep 14, 2026
- fix(googleai): give every streamed call without an ID its own ID
  
  sirozha
  and
  claude
  committed
  dc2be39
- fix(reasoning): switch thinking off on DashScope guests with enable_thinking
  
  sirozha
  and
  claude
  committed
  0cf20b5
- fix(reasoning): switch thinking off on GLM served by Mistral by omitting the field
  
  sirozha
  and
  claude
  committed
  2e0721a
- fix(llms): report Qwen's opt-in thinking as off by default in the hint
  
  sirozha
  and
  claude
  committed
  8be8c30
- test(openai): pin the off wire of GLM and DeepSeek behind dashscope/ and mistral/
  
  sirozha
  and
  claude
  committed
  ee34715
- test(llms): pin that opt-in alone does not make the hint report thinking off
  
  sirozha
  and
  claude
  committed
  b7b35a5
- test: pin six guards whose negation no test noticed (PAGI-337)
  
  sirozha
  and
  claude
  committed
  07f62d5
- fix(bedrock): leave an assistant turn that carries nothing off the Converse wire
  
  sirozha
  and
  claude
  committed
  b54d2d9
- fix(openai): leave DeepSeek's penalties out, and its temperature while it thinks
  
  sirozha
  and
  claude
  committed
  962f8d2
- test: pin the exact value every clamp and substitute warning reports (PAGI-337)
  
  sirozha
  and
  claude
  committed
  34a4387
- fix(openai): leave top_k off MiniMax's API, which has no field for it
  
  sirozha
  and
  claude
  committed
  9f2d8d7
- test: read the wire in the tests that only checked for no warning (PAGI-337)
  
  sirozha
  and
  claude
  committed
  9475176
- fix(openai): refuse a JSON schema for MiniMax M-series before the request
  
  sirozha
  and
  claude
  committed
  7032ea0
- test: pin the exact depth a named effort gets under adaptive on googleai (PAGI-337)
  
  sirozha
  and
  claude
  committed
  b9e5acb
- test: the mistral Call refusal must be the truncation error (PAGI-337)
  
  sirozha
  and
  claude
  committed
  b70ab1d
- test(llmtest): fail a door that takes tools and never calls one (PAGI-337)
  
  sirozha
  and
  claude
  committed
  973ff6b
- test(reasoning): check that each door passes its own provider constant (PAGI-337)
  
  sirozha
  and
  claude
  committed
  e1d18de
- test(llms): reach the unclassified openai branch with Mistral's magistral aliases
  
  sirozha
  and
  claude
  committed
  a0baa8a
- fix(reasoning): turn thinking off on Claude 5 behind the openai door
  
  sirozha
  and
  claude
  committed
  af74c13
- fix(openai): send a Claude thinking budget in Anthropic's thinking object
  
  sirozha
  and
  claude
  committed
  6bec5f8
- fix(llms): name the think levels gpt-oss takes on the ollama door in the hint
  
  sirozha
  and
  claude
  committed
  d981b17
- fix(openai): keep JSON mode off the wire for Claude and say so
  
  sirozha
  and
  claude
  committed
  4023aad
- test(openai): check top_p in the Claude sampling test of the openai door
  
  sirozha
  and
  claude
  committed
  f598c2d
Commits on Sep 15, 2026
- fix(bedrock): send gpt-oss the reasoning_effort its Converse request documents
  
  sirozha
  and
  claude
  committed
  9ea67b8
- fix(llms): offer the gpt-oss levels on the Bedrock hint
  
  sirozha
  and
  claude
  committed
  e246028
- test(llmtest): spare a door that reports the tools dropped (PAGI-337)
  
  sirozha
  and
  claude
  committed
  1c592be
- test(openai): hold the qwen3-max budget warning to the wire, not to 2730 (PAGI-337)
  
  sirozha
  and
  claude
  committed
  d1a8ef8
- fix(openai): take DeepSeek's temperature rule only on DeepSeek's own host
  
  sirozha
  and
  claude
  committed
  1d8ad3c
- fix(reasoning): count deepseek-flash as a model that thinks by default
  
  sirozha
  and
  claude
  committed
  3c9a4d3
- fix(openai): leave DeepSeek's top_p out once it is not thinking
  
  sirozha
  and
  claude
  committed
  91acf11
- fix(openai): leave response_format off MiniMax M-series in JSON mode
  
  sirozha
  and
  claude
  committed
  4ba2e12
- fix(reasoning): resolve deepseek-flash as the model deepseek-v4-flash names
  
  sirozha
  and
  claude
  committed
  6a4cf82
- fix(reasoning): refuse gpt-oss off on Bedrock under its GovCloud ID too
  
  sirozha
  and
  claude
  committed
  ab0b660
- fix(openai): give Qwen back the reasoning of every assistant turn too
  
  sirozha
  and
  claude
  committed
  397ba14
- fix(reasoning): keep reasoning_effort off GLM before 5.2 and Kimi K2
  
  sirozha
  and
  claude
  committed
  51b63ac
- fix(reasoning): keep reasoning_effort off MiniMax
  
  sirozha
  and
  claude
  committed
  7a5ee80
- fix(openai): give MiniMax back its reasoning inside think tags on every turn
  
  sirozha
  and
  claude
  committed
  cea0a15
- fix(openai): refuse a thinking budget that no field on the door can carry
  
  sirozha
  and
  claude
  committed
  2f174a8
- test(openai): pin a budget on GLM 5.2+ and Kimi K3 to the adaptive-only Claude wire
  
  sirozha
  and
  claude
  committed
  6f0c1f5
- fix(llms): offer Nova 2 Lite's off on the Bedrock hint
  
  sirozha
  and
  claude
  committed
  454df74
- test(openai): pin the thinking wire of GLM, Kimi and MiniMax on each host
  
  sirozha
  and
  claude
  committed
  1d71dec
- fix(openai): leave a MiniMax turn alone when its content opens with a think block
  
  sirozha
  and
  claude
  committed
  0fc2f42
- test(openai): pin a MiniMax turn without reasoning going back without think tags
  
  sirozha
  and
  claude
  committed
  8d2f933
- test(openai): pin the budget refusal when only one reasoning-object option is set
  
  sirozha
  and
  claude
  committed
  88d9f1c
- test(reasoning): pin that an undated GLM alias is not a pre-5.2 generation
  
  sirozha
  and
  claude
  committed
  c2a4f22
- merge: switch thinking off by the host that serves GLM and DeepSeek (PAGI-332)
  
  sirozha
  and
  claude
  committed
  207aef7
- merge: Claude behind the openai door gets its documented off, budget and JSON wire (PAGI-331)
  
  sirozha
  and
  claude
  committed
  8846f5a
- merge: no reasoning_effort for GLM before 5.2, Kimi K2 and MiniMax (PAGI-345)
  
  sirozha
  and
  claude
  committed
  5ab7350
- merge: leave out what DeepSeek and MiniMax ignore (PAGI-333)
  
  sirozha
  and
  claude
  committed
  230b6e2
- merge: replay reasoning on every turn for Qwen and MiniMax (PAGI-334)
  
  sirozha
  and
  claude
  committed
  deb66d9
- merge: the hint reports Qwen's opt-in thinking as off by default (PAGI-353)
  
  sirozha
  and
  claude
  committed
  cf329b8
- merge: the hint names the think levels gpt-oss takes on the ollama door (PAGI-300)
  
  sirozha
  and
  claude
  committed
  ad234eb
- merge: resolve deepseek-flash as the model deepseek-v4-flash names (PAGI-133)
  
  sirozha
  and
  claude
  committed
  10b2fa7
- merge: every streamed googleai call without an ID gets its own ID (PAGI-335)
  
  sirozha
  and
  claude
  committed
  72f5ee5
- merge: an empty assistant turn stays off the Converse wire (PAGI-336)
  
  sirozha
  and
  claude
  committed
  e237099
- merge: Bedrock open models get the thinking controls AWS documents (PAGI-354)
  
  sirozha
  and
  claude
  committed
  c7a62d1
- test(llms): leave gpt-oss on Bedrock to the Bedrock hint tests
  
  sirozha
  and
  claude
  committed
  fe97a00
- merge: pin the guards that survived mutation and the weak checks (PAGI-337)
  
  sirozha
  and
  claude
  committed
  981983f
- refactor(llms): build the level-only hint of gpt-oss, Nova and ollama in one place
  
  sirozha
  and
  claude
  committed
  965d331
- fix(googleai): surface a failed stream before the empty-stream budget guess
  
  sirozha
  and
  claude
  committed
  6aa6de9
- fix(openai): leave the DashScope thinking budget out of the answer-limit cut
  
  sirozha
  and
  claude
  committed
  7e8acae
- fix(reasoning): let every DashScope GLM and Kimi guest carry a thinking budget
  
  sirozha
  and
  claude
  committed
  f2da7a9
- fix(llms): report an explicit zero seed as a dropped option
  
  sirozha
  and
  claude
  committed
  9e17757
- fix(openai): keep top_k off DeepSeek's own door
  
  sirozha
  and
  claude
  committed
  fc1fed1
- test(openai): mark the DashScope budget table as a long table test
  
  sirozha
  and
  claude
  committed
  327b9fd
Commits on Sep 17, 2026
- merge: the library debt left on the Story (PAGI-247)
  
  sirozha
  and
  claude
  committed
  2306b4e
- fix(openai): leave the fields extra body does not name as the door built them
  
  sirozha
  and
  claude
  committed
  0a9aec0
- fix(openai): merge extra body objects into the objects the door builds
  
  sirozha
  and
  claude
  committed
  61477f9
- test(openai): pin the GPT-5 sampling that survives none and dies at minimal
  
  sirozha
  and
  claude
  committed
  78d5239
- test(googleai): pin that a thinking Gemini 3 call keeps the caller's sampling
  
  sirozha
  and
  claude
  committed
  cd41669
- test(bedrock): pin the top_p floor on a Claude model that takes both params
  
  sirozha
  and
  claude
  committed
  cb9e16f
- test(reasoning): pin where DeepSeek stops waiting to be asked
  
  sirozha
  and
  claude
  committed
  68f51d2
- refactor(llms): drop the vendor-fact comments that tests already pin
  
  sirozha
  and
  claude
  committed
  6c2d6cc
- fix(openai): let an extra body object of another type replace the door's
  
  sirozha
  and
  claude
  committed
  689197b
- merge: an extra body object merges into the object the door builds (PAGI-338)
  
  sirozha
  and
  claude
  committed
  e9ff02b
- fix(reasoning): stop presenting GLM 4.7, 4.7 Flash and 5 as reasoning models on Bedrock
  
  sirozha
  and
  claude
  committed
  0c57ff2
- merge: GLM 4.7, 4.7 Flash and 5 offer no reasoning control on Bedrock (PAGI-354)
  
  sirozha
  and
  claude
  committed
  eca1855
- merge: vendor facts and version boundaries are pinned by tests, not comments (PAGI-133)
  
  sirozha
  and
  claude
  committed
  2d4d83a
- fix(openai): turn adaptive thinking on for Claude 4.6-4.8 when the depth is delegated
  
  sirozha
  and
  claude
  committed
  dc9cfa3
- fix(openai): ask Claude for summarized thinking when the depth is delegated
  
  sirozha
  and
  claude
  committed
  5109954
- test(openai): pin that a Claude request without reasoning goes out without thinking
  
  sirozha
  and
  claude
  committed
  5d719f3
- fix(openai): keep the adaptive thinking object off hosts that resell Claude
  
  sirozha
  and
  claude
  committed
  e3e6509
- fix(openai): report adaptive thinking instead of sending it to Anthropic's compatibility endpoint
  
  sirozha
  and
  claude
  committed
  f09c0cd
- docs(reasoning): state only what AWS documents for the Bedrock non-reasoning list
  
  sirozha
  and
  claude
  committed
  854dec5
- test(bedrock): pin structured output on the us. and global. spellings of Claude Sonnet 4.6
  
  sirozha
  and
  claude
  committed
  1509b88
- merge: delegated adaptive thinking reaches Claude Opus 4.6-4.8 and Sonnet 4.6 on the openai door (PAGI-357)
  
  sirozha
  and
  claude
  committed
  0cc473d
- merge: the Bedrock non-reasoning list states only what AWS documents and Sonnet 4.6 structured output is pinned on every spelling (PAGI-247)
  
  sirozha
  and
  claude
  committed
  d9b9b44
- test(bedrock): pin that an adaptive request on a model without a mechanism leaves Converse untouched
  
  sirozha
  and
  claude
  committed
  99631b2
- docs(bedrock): stop listing Nova 2 Sonic among the models that take reasoningConfig
  
  sirozha
  and
  claude
  committed
  d9514e1
- docs(bedrock): stop presenting GLM 4.7, 4.7 Flash and 5 as reasoning models
  
  sirozha
  and
  claude
  committed
  833b236
- docs(reasoning): drop the vendor sentence tests already pin from NovaClearsInferenceConfigAt
  
  sirozha
  and
  claude
  committed
  eb17e49
- docs(reasoning): describe IsBedrockNonReasoningModel as the list it checks
  
  sirozha
  and
  claude
  committed
  e93843d
- docs(reasoning): drop the NovaEffort comment that restates its switch
  
  sirozha
  and
  claude
  committed
  5a43de8
- docs(reasoning): drop the EffortToolsDisable comment the wire test pins
  
  sirozha
  and
  claude
  committed
  28e64e9
- test(llmtest): drop the tool-probe comments that contradict the checks below them
  
  sirozha
  and
  claude
  committed
  3fc454e
- fix(reasoning): keep the thinking budget off DashScope's glm-5.3
  
  sirozha
  and
  claude
  committed
  5cdd492
- fix(openai): send the DashScope guest budget next to the answer-only max_tokens
  
  sirozha
  and
  claude
  committed
  a7e9bbb
- fix(openai): leave answer room in the DashScope DeepSeek V4 thinking budget
  
  sirozha
  and
  claude
  committed
  bf7897d
- fix(openai): turn thinking on for DashScope's Kimi K2.5 and K2.6 when asked
  
  sirozha
  and
  claude
  committed
  9a22b9b
- fix(openai): give a model sent to DashScope's own host the DashScope guest wire
  
  sirozha
  and
  claude
  committed
  a0fffdf
- fix(openai): keep top_k off the DeepSeek, Kimi and MiniMax models DashScope serves
  
  sirozha
  and
  claude
  committed
  f36aab1
- fix(openai): keep the Claude budget object off hosts that document no thinking object
  
  sirozha
  and
  claude
  committed
  1f12204
- fix(reasoning): refuse turning off a default-on Claude where no thinking object is taken
  
  sirozha
  and
  claude
  committed
  d861f6e
- fix(openai): send Anthropic's thinking object only to hosts outside the public provider list
  
  sirozha
  and
  claude
  committed
  0bd22a3
- test(openai): pin the Claude thinking object route list as an allow-list
  
  sirozha
  and
  claude
  committed
  b8ea044
- fix(openai): hold a client-level json_object to the JSON-mode model rule
  
  sirozha
  and
  claude
  committed
  73a1618
- merge: DashScope guests get the documented budget, off and top_k wire on its own host (PAGI-247)
  
  sirozha
  and
  claude
  committed
  a6994ea
- merge: Anthropic's thinking object stays on the routes that take it for budget and off too (PAGI-357)
  
  sirozha
  and
  claude
  committed
  1a2f600
- merge: a client-level json_object follows the JSON-mode model rule (PAGI-247)
  
  sirozha
  and
  claude
  committed
  1608cb4
- merge: Bedrock docs and reasoning comments state only what the code does (PAGI-247)
  
  sirozha
  and
  claude
  committed
  ba8ca21
- fix(bedrock): refuse structured output on models whose AWS model card does not list it
  
  sirozha
  and
  claude
  committed
  200f1a6
- merge: Bedrock structured output follows the AWS model cards (PAGI-247)
  
  sirozha
  and
  claude
  committed
  e28a81a
- test(chains): pin the model the cassettes record
  
  sirozha
  and
  claude
  committed
  c21b34b
- merge: chains replays the cassettes it records (PAGI-125)
  
  sirozha
  and
  claude
  committed
  4e702f2
- fix(reasoning): the Bedrock non-reasoning verdict reads the inference profile too
  
  sirozha
  and
  claude
  committed
  4ea2c28
- merge: the Bedrock verdict reads region-prefixed ids (PAGI-125)
  
  sirozha
  and
  claude
  committed
  a515a18
Commits on Sep 18, 2026
- feat(reasoning): carry GPT-6 Astra by its model card
  
  sirozha
  and
  claude
  committed
  cf6e06b
- merge: GPT-6 Astra is carried by its model card (PAGI-125)
  
  sirozha
  and
  claude
  committed
  32b9b9d
- fix(reasoning): GPT-6 Astra takes the four effort levels its endpoint accepts
  
  sirozha
  and
  claude
  committed
  fa65c91
- merge: Astra's effort set follows the endpoint, not the card (PAGI-125)
  
  sirozha
  and
  claude
  committed
  7ca2ccb
- fix(reasoning): GLM 5.3 keeps thinking under the spelling Mistral serves it by
  
  sirozha
  and
  claude
  committed
  2edb2f9
- merge: GLM 5.3 keeps thinking under the Mistral spelling (PAGI-125)
  
  sirozha
  and
  claude
  committed
  da82eee
- feat(bedrock): gpt-oss-safeguard takes the reasoning effort OpenAI documents for it
  
  sirozha
  and
  claude
  committed
  405ffe2
- merge: gpt-oss-safeguard carries its documented effort (PAGI-125)
  
  sirozha
  and
  claude
  committed
  c2d68ef
Commits on Sep 21, 2026
- feat: implement metadata indexing and filtering in pgvector
  asdek
  committed
  504595c
Commits on Sep 22, 2026
- fix(pgvector): a failed init rolls back instead of keeping its connection (PAGI-405)
  
  sirozha
  and
  claude
  committed
  da097ac
- fix(reasoning): grok-4.7 always reasons, so a disable is refused before the wire
  
  sirozha
  and
  claude
  committed
  d88b9de
- test(reasoning): pin grok-4.7 spellings where every other family's are pinned
  
  sirozha
  and
  claude
  committed
  d3bf3ec
- feat(reasoning): add support for new GPT-6 generations and update capability checks
  asdek
  committed
  dd63f69
- fix(reasoning): a us-gov model keeps the capabilities of its bare name
  
  sirozha
  and
  claude
  committed
  9a40f2f
- fix(reasoning): gpt-6-sol and gpt-6-luna take function tools only without effort
  
  sirozha
  and
  claude
  committed
  9a0394f
- test(openai): pin max to xhigh for gpt-6-sol and gpt-6-luna
  
  sirozha
  and
  claude
  committed
  e1c0414
- fix(pgvector): a cancelled init rolls back before the commit can close the connection
  
  sirozha
  and
  claude
  committed
  abcf47d
- test(pgvector): explain the statement SimilaritySearch sends, not a copy
  
  sirozha
  and
  claude
  committed
  fdf1f72
- fix(pgvector): a derived index name tells every declaration apart
  
  sirozha
  and
  claude
  committed
  a35a0a3
- fix(pgvector): a derived index name stays valid UTF-8 and the name guard folds case
  
  sirozha
  and
  claude
  committed
  40b6fb1
- test(pgvector): the plan test builds its own tables
  
  sirozha
  and
  claude
  committed
  c981c07
- fix(openai): the none that function tools force keeps the caller's sampling
  
  sirozha
  and
  claude
  committed
  74cce2a
- fix(pgvector): the name guard compares what two declarations define, not their SQL
  
  sirozha
  and
  claude
  committed
  7ab52b8
- test(pgvector): the name guard refuses exclusions that differ under one name
  
  sirozha
  and
  claude
  committed
  cbf020b
- merge: pgvector index names and init cancellation, gpt-6-sol/luna tools rule, sampling under a tools-forced none (PAGI-125)
  
  sirozha
  and
  claude
  committed
  ddc9643
Commits on Sep 23, 2026
Commits on Sep 24, 2026
- feat(ollama): implement structured output fallback for cloud models
  asdek
  committed
  e4b6095
- feat(openai): implement structured output fallback for DeepSeek and Z.ai models
  asdek
  committed
  355dc71

Commits

Commits on Sep 10, 2026

Commits on Sep 11, 2026

Commits on Sep 13, 2026

Commits on Sep 14, 2026

Commits on Sep 15, 2026