feat(providers/anthropic): surface refusal stop_reason as content-filter#41
Merged
Merged
Conversation
Author
|
Verified the claim in the
So the raw-JSON parsing is the right call today. Suggested follow-ups, not blocking:
Coder Agents generated (on behalf of @jscottmiller). |
Author
|
Merge-order note relative to #40 (the main -> coder_2_33 sync):
Since the refusal mapping is not Coder-specific, it is also a candidate to send upstream to charmbracelet/fantasy rather than carrying it as a fork patch long-term. Coder Agents generated (on behalf of @jscottmiller). |
Author
|
/coder-agents-review |
Map Anthropic's "refusal" stop_reason to FinishReasonContentFilter and capture stop_details (category, explanation) into the finish part's ProviderMetadata via a new RefusalMetadata type, so callers can surface why a turn produced no content. Coder Agents generated.
ethanndickson
force-pushed
the
scott/anthropic-refusal-content-filter
branch
from
July 9, 2026 06:16
654801e to
798e38d
Compare
ethanndickson
marked this pull request as ready for review
July 9, 2026 06:17
Member
|
@codex review |
|
Codex Review: Didn't find any major issues. Bravo. Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
ethanndickson
added a commit
to coder/coder
that referenced
this pull request
Jul 9, 2026
When a provider's real-time safety classifiers block a response (e.g. Anthropic stop_reason "refusal"), the turn previously ended silently: the fantasy SDK mapped the stop reason to unknown and the empty-content finish looked like a normal stop, leaving the UI stuck on "Thinking" with no output. Bump the fantasy fork to pick up refusal -> FinishReasonContentFilter mapping with stop_details carried as RefusalMetadata (fork change 15, coder/fantasy#41). chatloop now turns an empty content-filter finish into a terminal, non-retryable content_filter classified error, using the provider's explanation as the message and the refusal category as the detail, with a generic per-provider fallback message. The UI renders it as "Response blocked". Partial-content refusals (classifier blocks after some output) still persist the partial text and end the turn normally; surfacing those is a separate product decision. Coder Agents generated.
ethanndickson
added a commit
to coder/coder
that referenced
this pull request
Jul 9, 2026
When a provider's real-time safety classifiers block a response (e.g. Anthropic stop_reason "refusal"), the turn previously ended silently: the fantasy SDK mapped the stop reason to unknown and the empty-content finish looked like a normal stop, leaving the UI stuck on "Thinking" with no output. Bump the fantasy fork to pick up refusal -> FinishReasonContentFilter mapping with stop_details carried as RefusalMetadata (fork change 15, coder/fantasy#41). chatloop now turns an empty content-filter finish into a terminal, non-retryable content_filter classified error, using the provider's explanation as the message and the refusal category as the detail, with a generic per-provider fallback message. The UI renders it as "Response blocked". Partial-content refusals (classifier blocks after some output) still persist the partial text and end the turn normally; surfacing those is a separate product decision. Coder Agents generated.
ibetitsmike
added a commit
to coder/coder
that referenced
this pull request
Jul 9, 2026
gpt-5.6 models were unusable with agents: fantasy's Responses allowlist did not include the new family, so `IsResponsesModel` returned false and chatd fell back to the Chat Completions path (no reasoning params, no encrypted reasoning continuity). ## Changes - Bump the `charm.land/fantasy` replace pin to coder/fantasy `6da0c3b10237` (coder_2_33), pulling in: - coder/fantasy#46: route `gpt-5.6-sol`, `gpt-5.6-terra`, `gpt-5.6-luna` through the OpenAI Responses API as reasoning models. - coder/fantasy#41: surface Anthropic refusal stop_reason as content-filter (already on coder_2_33, rides along with the bump). - Update the fork changelog comment in go.mod. ## Verification - Probed all three gpt-5.6 models through the ai-gateway: `/v1/responses` with `reasoning.effort`, `include: ["reasoning.encrypted_content"]`, and `store: false` completes for each. - `go build ./coderd/...` and `go test ./coderd/x/chatd/chatopenai/ ./coderd/x/chatd/chatprovider/` pass against the new pin. > This PR was authored by Mux on Mike's behalf.
mtojek
pushed a commit
to coder/coder
that referenced
this pull request
Jul 13, 2026
Cherry-pick of #27132 (`f84801ee`) onto `release/2.35`. gpt-5.6 models were unusable with agents: fantasy's Responses allowlist did not include the new family, so `IsResponsesModel` returned false and chatd fell back to the Chat Completions path (no reasoning params, no encrypted reasoning continuity). ## Changes - Bump the `charm.land/fantasy` replace pin from `2e3ddbca75dd` to coder/fantasy `6da0c3b10237` (coder_2_33), pulling in: - coder/fantasy#46: route `gpt-5.6-sol`, `gpt-5.6-terra`, `gpt-5.6-luna` through the OpenAI Responses API as reasoning models. - coder/fantasy#41: surface Anthropic refusal stop_reason as content-filter (rides along; sits below #46 on coder_2_33). - Update the fork changelog comment in go.mod. ## Backport safety - The delta over 2.35's pin is exactly the two commits above; #27132 itself touched only go.mod/go.sum, and no coder/coder code references the refusal metadata or content-filter finish reason (verified on main and 2.35). - Finish-part provider metadata is never persisted by chatd, so a rollback to an unpatched 2.35.x binary cannot hit undecodable stored history. - `go build ./coderd/...` and `go test ./coderd/x/chatd/{chatopenai,chatprovider,chatprompt,chatloop}/` pass on this branch; `go mod tidy` is a no-op. > This PR was authored by Mux on Mike's behalf.
ethanndickson
added a commit
to coder/coder
that referenced
this pull request
Jul 13, 2026
When a provider's real-time safety classifiers block a response (e.g. Anthropic stop_reason "refusal"), the turn previously ended silently: the fantasy SDK mapped the stop reason to unknown and the empty-content finish looked like a normal stop, leaving the UI stuck on "Thinking" with no output. Bump the fantasy fork to pick up refusal -> FinishReasonContentFilter mapping with stop_details carried as RefusalMetadata (fork change 15, coder/fantasy#41). chatloop now turns an empty content-filter finish into a terminal, non-retryable content_filter classified error, using the provider's explanation as the message and the refusal category as the detail, with a generic per-provider fallback message. The UI renders it as "Response blocked". Partial-content refusals (classifier blocks after some output) still persist the partial text and end the turn normally; surfacing those is a separate product decision. Coder Agents generated.
ethanndickson
added a commit
to coder/coder
that referenced
this pull request
Jul 14, 2026
When a provider's real-time safety classifiers block a response (e.g. Anthropic stop_reason "refusal"), the turn previously ended silently: the fantasy SDK mapped the stop reason to unknown and the empty-content finish looked like a normal stop, leaving the UI stuck on "Thinking" with no output. Bump the fantasy fork to pick up refusal -> FinishReasonContentFilter mapping with stop_details carried as RefusalMetadata (fork change 15, coder/fantasy#41). chatloop now turns an empty content-filter finish into a terminal, non-retryable content_filter classified error, using the provider's explanation as the message and the refusal category as the detail, with a generic per-provider fallback message. The UI renders it as "Response blocked". Partial-content refusals (classifier blocks after some output) still persist the partial text and end the turn normally; surfacing those is a separate product decision. Coder Agents generated.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Map Anthropic's
refusalstop reason to a content-filter finish reason andpreserve the block details, so downstream callers can tell a safety block apart
from a normal stop and surface why a turn produced no content.
Anthropic's real-time safety classifiers end a response with
stop_reason: "refusal"and astop_detailsobject (category,explanation), and emit no content. PreviouslymapFinishReasonfell throughto
FinishReasonUnknownand the details were dropped.Changes
mapFinishReason:"refusal"->FinishReasonContentFilter.stop_detailsfrom themessage_deltaraw JSON (the Anthropic SDK doesnot model it) and attach it to the finish stream part via a new
RefusalMetadata{Category, Explanation}provider-metadata type(registered, marshal/unmarshal,
GetRefusalMetadatagetter).Testing
go test ./providers/anthropic/(newrefusal_test.gocovers the finish-reasonmapping,
stop_detailsparsing, and metadata round-trip).Notes
Consumed by coder/coder PR for chat "Response blocked" messaging.
Coder Agents generated.