fix(bedrock): handle `toolResult` attachment co-location per-model via `bedrock_tool_result_colocatable_content` by Hasnaathussain · Pull Request #6098 · pydantic/pydantic-ai

Hasnaathussain · 2026-06-27T19:13:33Z

Problem

On Bedrock, pydantic-ai merges consecutive user turns into one Converse message. When a tool-result turn is followed by a user turn carrying an attachment, they merge into one user message co-locating a toolResult block with the attachment — which some models reject with a misleading ValidationException about a missing tool_result.

The original fix split every such turn with a synthetic assistant message for all models. Live testing against Bedrock (us-east-1) shows the constraint is per-model, so a blanket split is both incomplete (misses models with a stricter rule) and over-broad (needlessly pollutes models with no rule).

Evidence (live Converse, what may share a user message with a `toolResult`)

Family	text	image	document (pdf)	document (csv/txt/docx)	video
Anthropic	✅	✅	✅	❌	❌ (video unsupported)
Amazon Nova	✅	✅	✅	✅	✅
Meta Llama 4	❌	❌	❌	❌	❌
Mistral (pixtral)	❌	❌	❌	❌	❌
Qwen	✅	✅	— (no doc support)	—	—

Anthropic rejects documents and video beside a toolResult, but accepts text and images. (PDF happens to pass, but the pdf-vs-other-document split is undocumented and brittle, so we conservatively treat all documents as non-co-locatable.)
Nova has no constraint at all — the old blanket fix inserted a pointless dummy turn here.
Llama and Mistral reject any content sharing a toolResult turn, even plain text — a broader form of the same bug that the attachment-only guard missed.

Mechanism notes, also verified: a bare split into two consecutive user turns does not work (Bedrock re-merges same-role turns), so an alternating assistant turn is required; the separator must be non-whitespace (. works everywhere); block order within the message is irrelevant. AWS does not document the toolResult+document constraint — it's model behavior — which is why this is driven by empirical per-family testing.

Fix

Add a bedrock_tool_result_colocatable_content model-profile flag listing the content-block kinds a model accepts alongside a toolResult. The merge pass splits the turns (separating them with a synthetic . assistant message) only when the combined content isn't co-locatable for that model; otherwise it merges as before.

Family	flag value	effect
Anthropic	`{'text', 'image'}`	splits out documents + video
Meta, Mistral	`frozenset()`	`toolResult` isolated from all other content
everyone else (default)	all kinds	merges freely, no dummy turn

The default is permissive, so no behavior change for uncovered families.

Tests

Per-family mapping unit tests pinning the exact reshaped sequence (Anthropic split, Nova merged, Llama isolated).
VCR cassettes recorded against live Bedrock proving each family accepts its reshaped history: Anthropic with a text/plain document (the original PDF recording is accepted even without the fix, so it proved nothing — a text document actually reproduces the rejection), Nova co-located with no dummy turn, and Llama with text isolated.
Provider-profile assertions for the new flag values.

Follow-ups (separate PRs)

Live testing surfaced two adjacent, distinct defects that are handled in their own PRs rather than widening this one:

bedrock_supported_media_kinds_in_tool_returns has wrong per-family values (Mistral/Writer accept document but reject image inside tool results; Qwen accepts neither; Nova rejects binary documents).
The "conversation must start with a user message" workaround is provider-specific (Anthropic and Qwen accept a leading assistant turn) and currently applied as a blanket rule.

🤖 Generated with Claude Code

Copilot

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

DouweM · 2026-07-01T18:14:15Z

There's some related logic here:

pydantic-ai/pydantic_ai_slim/pydantic_ai/models/bedrock.py

Lines 1297 to 1303 in e2bdafb

    
           # https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_Message.html 
        
           # "If you include a ContentBlock with a document field, you must also include a ContentBlock with a text field." 
        
           has_document = any('document' in block for block in content) 
        
           has_text = any('text' in block for block in content) 
        
           if has_document and not has_text: 
        
               content.insert(0, {'text': 'See attached document(s).'}) 
        
           return [{'role': 'user', 'content': content}]

Worth checking to see if we can make it more consistent / fix it in one place.

Hasnaathussain · 2026-07-01T18:58:27Z

I have addressed this! I've updated the logic at the end of _map_messages to perform a clean pass over all final processed user messages. Any message that contains a document block but lacks a text block will now have a "See attached document(s)." prefix automatically inserted. This makes the text-field injection completely consistent and robust across user prompts, tool returns, and split/merged user turns. All 138 Bedrock tests pass locally.

coderabbitai · 2026-07-02T12:06:48Z

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: a97d87fb-02d2-4ee3-8446-0ccc165301b0

📥 Commits

Reviewing files that changed from the base of the PR and between 3e2204b and 0e7ffd3.

📒 Files selected for processing (5)

pydantic_ai_slim/pydantic_ai/models/bedrock.py
pydantic_ai_slim/pydantic_ai/providers/bedrock.py
tests/models/test_bedrock.py
tests/profiles/test_resolution_matrix.py
tests/providers/test_bedrock.py

🔗 Linked repositories identified

CodeRabbit considers these linked repositories for cross-repo context during reviews:

pydantic/logfire (manual)
pydantic/pydantic-ai (manual)
pydantic/pydantic (auto-detected)

✅ Files skipped from review due to trivial changes (1)

tests/profiles/test_resolution_matrix.py

🚧 Files skipped from review as they are similar to previous changes (2)

tests/providers/test_bedrock.py
pydantic_ai_slim/pydantic_ai/models/bedrock.py

📝 Walkthrough

Walkthrough

This PR updates Bedrock message mapping so toolResult content is not merged with disallowed attachments in the same user message. A new Bedrock model-profile field defines which content kinds may colocate with toolResult; Anthropic allows text and image, Meta and Mistral allow none, and Nova stays unset. _map_messages now inserts a synthetic assistant . turn when needed. Tests and cassettes cover Anthropic, Nova, and Llama behavior.

Changes

Related issues: #6081
Suggested labels: bug, bedrock

Sequence Diagram(s)

sequenceDiagram
  participant _map_messages
  participant BedrockModelProfile
  participant BedrockMessages
  _map_messages->>BedrockModelProfile: read bedrock_tool_result_colocatable_content
  _map_messages->>_map_messages: inspect following user turn content kinds
  alt toolResult + disallowed content kind
    _map_messages->>BedrockMessages: insert synthetic assistant "."
  else allowed content kind
    _map_messages->>BedrockMessages: merge into prior user message
  end

Poem:
A toolResult met a document,
and Bedrock said, “Not consequent.”
So . stepped in between the two,
and all the turns now made it through.

🚥 Pre-merge checks | ✅ 4

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly matches the main change: per-model handling of Bedrock toolResult co-location via a new profile flag.
Description check	✅ Passed	The description covers the problem, fix, tests, and closing issue, matching the repository template well enough.
Linked Issues check	✅ Passed	The PR implements the issue's core requirement by preventing tool_result from being merged with attachment-bearing user turns.
Out of Scope Changes check	✅ Passed	The added tests, provider profile field, and cassettes all support the Bedrock co-location fix and stay within scope.

✨ Finishing Touches

🧪 Generate unit tests (beta)

Create PR with unit tests

_{Comment @coderabbitai help to get the list of available commands.}

coderabbitai

🧹 Nitpick comments (2)

pydantic_ai_slim/pydantic_ai/models/bedrock.py (2)

1192-1207: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low value

Confirm this block is truly needed given it's marked unreachable.

This normalization pass is marked # pragma: no cover (per the commit, intentionally, since _map_user_prompt and the merge logic already guarantee the document+text invariant per-message before this runs). Keeping unreachable defensive code is a reasonable safety net, but worth a second look: if it's truly dead given current invariants, consider removing it to avoid maintaining unreachable logic; if it's meant as future-proofing against new merge paths, a short comment explaining why would help future readers.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@pydantic_ai_slim/pydantic_ai/models/bedrock.py` around lines 1192 - 1207, The
normalization in bedrock.py that injects a fallback text block for user messages
with documents appears unreachable given the existing _map_user_prompt and merge
logic, so either remove the dead defensive branch from the processed_messages
loop or keep it only if needed for future-proofing and add a short explanatory
comment near the role/content check to justify why the invariant is not
guaranteed elsewhere.

1128-1148: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Duplicate detection logic could be extracted into a helper.

The any(... for k in ('image', 'document', 'video')) / any('toolResult' in block ...) checks are repeated four times here (and similar logic reappears at Line 1196 and Line 1330-1331). Extracting a small helper (e.g. _has_block_kind(content, *keys)) would reduce duplication and improve readability.

♻️ Proposed refactor

+    `@staticmethod`
+    def _has_block_kind(content: list[Any], *keys: str) -> bool:
+        return any(any(k in block for k in keys) for block in content)
+
     for current_message in bedrock_messages:
         if (
             last_message is not None
             and current_message['role'] == last_message['role']
             and current_message['role'] == 'user'
         ):
-            last_has_tool_result = any('toolResult' in block for block in last_message['content'])
-            current_has_attachment = any(
-                any(k in block for k in ('image', 'document', 'video')) for block in current_message['content']
-            )
-            last_has_attachment = any(
-                any(k in block for k in ('image', 'document', 'video')) for block in last_message['content']
-            )
-            current_has_tool_result = any('toolResult' in block for block in current_message['content'])
+            last_has_tool_result = self._has_block_kind(last_message['content'], 'toolResult')
+            current_has_attachment = self._has_block_kind(current_message['content'], 'image', 'document', 'video')
+            last_has_attachment = self._has_block_kind(last_message['content'], 'image', 'document', 'video')
+            current_has_tool_result = self._has_block_kind(current_message['content'], 'toolResult')

As per coding guidelines, "Consolidate methods with duplicated logic using helpers, delegation, or type overloads."

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@pydantic_ai_slim/pydantic_ai/models/bedrock.py` around lines 1128 - 1148, The
duplicate block-type detection in bedrock message processing should be
consolidated into a helper to reduce repetition. Extract the repeated `any(...)`
content checks used for `toolResult` and attachment kinds into a small method or
local helper, such as one used by the message-merging logic in `bedrock.py`, and
reuse it wherever the same pattern appears in the message processing flow. Keep
the merge decision in the existing code path, but delegate the repeated
detection work to the helper so the `processed_messages` assembly stays readable
and easier to maintain.

Source: Coding guidelines

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Nitpick comments:
In `@pydantic_ai_slim/pydantic_ai/models/bedrock.py`:
- Around line 1192-1207: The normalization in bedrock.py that injects a fallback
text block for user messages with documents appears unreachable given the
existing _map_user_prompt and merge logic, so either remove the dead defensive
branch from the processed_messages loop or keep it only if needed for
future-proofing and add a short explanatory comment near the role/content check
to justify why the invariant is not guaranteed elsewhere.
- Around line 1128-1148: The duplicate block-type detection in bedrock message
processing should be consolidated into a helper to reduce repetition. Extract
the repeated `any(...)` content checks used for `toolResult` and attachment
kinds into a small method or local helper, such as one used by the
message-merging logic in `bedrock.py`, and reuse it wherever the same pattern
appears in the message processing flow. Keep the merge decision in the existing
code path, but delegate the repeated detection work to the helper so the
`processed_messages` assembly stays readable and easier to maintain.

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 6b3ba337-4fc2-4c00-9b40-1c03c92432fc

📥 Commits

Reviewing files that changed from the base of the PR and between c3cf7b3 and 1b6d02e.

📒 Files selected for processing (1)

pydantic_ai_slim/pydantic_ai/models/bedrock.py

🔗 Linked repositories identified

CodeRabbit considers these linked repositories for cross-repo context during reviews:

pydantic/logfire (manual)
pydantic/pydantic-ai (manual)
pydantic/pydantic (auto-detected)

…l_result and media attachment turns

…e mapping

… union check

… block

…ssistant message Bedrock rejects a user message that co-locates a `toolResult` block with an attachment block, and it also requires alternating roles, so a plain split into two `user` turns is not enough. Insert a minimal assistant turn between them, using `.` because Anthropic on Bedrock rejects whitespace-only text. - drop the redundant post-merge document-text guard; `_map_user_prompt` already guarantees a text block alongside any document - add a VCR regression test proving Converse accepts the reshaped history - keep a unit test pinning the exact mapped message sequence Refs pydantic#6081

coderabbitai

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

pydantic_ai_slim/pydantic_ai/models/bedrock.py (1)
1003-1035: 🩺 Stability & Availability | 🟠 Major | ⚡ Quick win

Split unsupported tool-return attachments into a separate user turn The else branch still appends sibling_content to the same user message as toolResult, so a tool return with a video (or any other unsupported media kind) can still produce the forbidden toolResult + attachment shape and hit the same Bedrock validation failure. Emit a second user message here instead.
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@pydantic_ai_slim/pydantic_ai/models/bedrock.py` around lines 1003 - 1035, The
Bedrock tool-return assembly in the branch handling is_multi_modal_content still
combines the toolResult payload and unsupported attachment sibling_content into
one user message, which can trigger the same invalid shape. Update this path in
bedrock.py so the toolResult message is appended first, then emit the attachment
content as a separate user turn instead of extending the same user_content list;
keep the logic around _map_file_to_content_block, tool_result_content, and
sibling_content intact.

🧹 Nitpick comments (1)

tests/models/test_bedrock.py (1)
5182-5211: 🎯 Functional Correctness | 🔵 Trivial | 💤 Low value

Solid regression coverage; only document attachment type exercised.

The fix's merge-guard checks image/document/video uniformly, but only document is covered by these new tests. Given the logic is generic and already unit-pinned, this is a minor coverage gap rather than a real risk.

As per coding guidelines, "Provide comprehensive tests covering 100% of code paths, favoring integration tests and real requests."
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@tests/models/test_bedrock.py` around lines 5182 - 5211, Add regression
coverage for the merge-guard in the Bedrock history reshaping logic by
exercising the other attachment types handled alongside `document`, especially
`image` and `video`. Update the Bedrock test suite around
`test_bedrock_tool_result_followed_by_attachment_accepted` and the related
history-mapping helpers so the same tool-result-followed-by-attachment scenario
is verified for all supported attachment variants, not just `document`.
Source: Coding guidelines

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Outside diff comments:
In `@pydantic_ai_slim/pydantic_ai/models/bedrock.py`:
- Around line 1003-1035: The Bedrock tool-return assembly in the branch handling
is_multi_modal_content still combines the toolResult payload and unsupported
attachment sibling_content into one user message, which can trigger the same
invalid shape. Update this path in bedrock.py so the toolResult message is
appended first, then emit the attachment content as a separate user turn instead
of extending the same user_content list; keep the logic around
_map_file_to_content_block, tool_result_content, and sibling_content intact.

---

Nitpick comments:
In `@tests/models/test_bedrock.py`:
- Around line 5182-5211: Add regression coverage for the merge-guard in the
Bedrock history reshaping logic by exercising the other attachment types handled
alongside `document`, especially `image` and `video`. Update the Bedrock test
suite around `test_bedrock_tool_result_followed_by_attachment_accepted` and the
related history-mapping helpers so the same tool-result-followed-by-attachment
scenario is verified for all supported attachment variants, not just `document`.

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: d5230da7-77d0-433f-8320-74ae9fddfbac

📥 Commits

Reviewing files that changed from the base of the PR and between 1b6d02e and 2f1f287.

📒 Files selected for processing (3)

pydantic_ai_slim/pydantic_ai/models/bedrock.py
tests/models/cassettes/test_bedrock/test_bedrock_tool_result_followed_by_attachment_accepted.yaml
tests/models/test_bedrock.py

🔗 Linked repositories identified

CodeRabbit considers these linked repositories for cross-repo context during reviews:

pydantic/logfire (manual)
pydantic/pydantic-ai (manual)
pydantic/pydantic (auto-detected)

…-result-attachments-conflict

The previous fix split every `toolResult`-then-attachment turn with a synthetic assistant message for all models, but live Converse testing shows the constraint is per-model: Anthropic rejects only documents/video beside a `toolResult` (text and images are fine), Llama and Mistral reject any content sharing the turn, and Nova has no constraint at all — so the blanket split needlessly polluted Nova (and every permissive model) with a dummy turn. Introduce a `bedrock_tool_result_colocatable_content` profile flag listing the content-block kinds a model accepts alongside a `toolResult`. The merge pass splits (separating the turns with a synthetic assistant message, since Bedrock re-merges consecutive same-role turns) only when the combined content isn't co-locatable for that model. Default is permissive, preserving behavior for uncovered families. Set to `{text, image}` for Anthropic and `frozenset()` for Meta and Mistral (the latter reproduce the same defect for plain text, a broader form of pydantic#6081). Replace the PDF VCR recording (which Anthropic accepts even without the fix, so it proved nothing) with `text/plain` recordings that actually reproduce the rejection, plus Nova (co-located, no split) and Llama (text isolated) cassettes. Refs pydantic#6081

coderabbitai

🧹 Nitpick comments (1)

tests/models/test_bedrock.py (1)

5169-5227: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Add VCR-exemption rationale to the Nova/Llama unit tests.

test_bedrock_anthropic_tool_result_and_document_alternate explains why it's a unit test rather than a VCR test (cassette matcher isn't sensitive to request body). test_bedrock_nova_tool_result_and_document_merge and test_bedrock_llama_tool_result_isolated_from_text pin the same kind of internal request-payload shape but their docstrings omit this rationale.

As per path instructions, "Each unit test should explain why it is not (or cannot be) a VCR test, especially when it pins internal behavior, pre-request guards, or request payload shape that cassettes would not reliably catch."

Source: Path instructions

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 0d65779c-6227-41f3-a628-f0ae36d23f96

📥 Commits

Reviewing files that changed from the base of the PR and between 2f1f287 and d76fddd.

📒 Files selected for processing (7)

pydantic_ai_slim/pydantic_ai/models/bedrock.py
pydantic_ai_slim/pydantic_ai/providers/bedrock.py
tests/models/cassettes/test_bedrock/test_bedrock_anthropic_tool_result_followed_by_document_accepted.yaml
tests/models/cassettes/test_bedrock/test_bedrock_llama_tool_result_followed_by_text_accepted.yaml
tests/models/cassettes/test_bedrock/test_bedrock_nova_tool_result_followed_by_document_accepted.yaml
tests/models/test_bedrock.py
tests/providers/test_bedrock.py

🔗 Linked repositories identified

CodeRabbit considers these linked repositories for cross-repo context during reviews:

pydantic/logfire (manual)
pydantic/pydantic-ai (manual)
pydantic/pydantic (auto-detected)

✅ Files skipped from review due to trivial changes (2)

tests/models/cassettes/test_bedrock/test_bedrock_anthropic_tool_result_followed_by_document_accepted.yaml
tests/models/cassettes/test_bedrock/test_bedrock_llama_tool_result_followed_by_text_accepted.yaml

🚧 Files skipped from review as they are similar to previous changes (1)

pydantic_ai_slim/pydantic_ai/models/bedrock.py

…result_colocatable_content` The new profile flag appears in the normalized profile snapshots for Anthropic (`{text, image}`), Meta, and Mistral (`frozenset()`) on Bedrock.

…t VCR

DouweM · 2026-07-04T00:04:59Z

    """Default: `False`."""
    bedrock_supported_media_kinds_in_tool_returns: frozenset[str]
    """Default: `frozenset({'image'})`."""
+    bedrock_tool_result_colocatable_content: frozenset[str]


Let's use a Literal union with the actual supported values, instead of str. That also goes for supported_media_kinds....

…eral` unions Address review: use `frozenset[Literal[...]]` instead of `frozenset[str]` for `bedrock_tool_result_colocatable_content` and `bedrock_supported_media_kinds_in_tool_returns`.

…flict Resolve additive conflicts from pydantic#6255 (leading-assistant-message flag): keep both that flag and this PR's `bedrock_tool_result_colocatable_content` in `BedrockModelProfile`, the Anthropic profile, and the resolution-matrix / test-bedrock snapshots.

…flict Resolve conflicts from pydantic#6254 (media-kinds fix): union the two `bedrock_meta_model_profile` definitions (both `bedrock_tool_result_colocatable_content` and `bedrock_supported_media_kinds_in_tool_returns`), and keep both flags' values in the Mistral profile and in the resolution-matrix / provider-test assertions.

Reconcile with #6098 (now in main): keep the tool-return media-deferral logic and its hoisted `colocatable_content`, and drop the duplicated co-location tests that landed via #6098 (they're a subset of this branch's superset).

Copilot AI review requested due to automatic review settings June 27, 2026 19:13

Copilot AI reviewed Jun 27, 2026

github-actions Bot added size: S Small PR (≤100 weighted lines) bug Report that something isn't working, or PR implementing a fix labels Jun 27, 2026

Hasnaathussain force-pushed the fix/6081-bedrock-tool-result-attachments-conflict branch 2 times, most recently from 573a32e to 79b7ab4 Compare July 1, 2026 20:06

coderabbitai Bot approved these changes Jul 2, 2026

View reviewed changes

coderabbitai Bot reviewed Jul 3, 2026

View reviewed changes

adtyavrdhn mentioned this pull request Jul 3, 2026

fix(bedrock): split tool results from attachment user messages #6085

Closed

Hasnaathussain and others added 9 commits July 3, 2026 18:15

fix(bedrock): prevent Converse ValidationException on consecutive too…

a765309

…l_result and media attachment turns

style: run ruff format to resolve CI failure

cd3c017

fix(bedrock): ensure user messages with documents contain a text block

04d661f

fix(bedrock): resolve pyright lint error on Sequence insert in messag…

01346b4

…e mapping

fix(bedrock): add type cast to MessageUnionTypeDef to satisfy Pyright…

d0a3897

… union check

test(bedrock): add test_bedrock_user_message_document_no_text unit test

12153a4

test(bedrock): add pragma no cover to unreachable document validation…

ffce290

… block

style: ruff format bedrock.py

af0ad90

adtyavrdhn force-pushed the fix/6081-bedrock-tool-result-attachments-conflict branch from a376ac7 to 2f1f287 Compare July 3, 2026 12:49

coderabbitai Bot reviewed Jul 3, 2026

View reviewed changes

adtyavrdhn approved these changes Jul 3, 2026

View reviewed changes

DouweM self-assigned this Jul 3, 2026

DouweM added 2 commits July 3, 2026 22:11

Merge remote-tracking branch 'origin/main' into fix/6081-bedrock-tool…

733caa9

…-result-attachments-conflict

github-actions Bot added size: M Medium PR (101-500 weighted lines) and removed size: S Small PR (≤100 weighted lines) labels Jul 3, 2026

DouweM added the douwebot reviews the PR that gets labeled and removes the label at the end label Jul 3, 2026

github-actions Bot removed the douwebot reviews the PR that gets labeled and removes the label at the end label Jul 3, 2026

coderabbitai Bot reviewed Jul 3, 2026

View reviewed changes

test(bedrock): update profile resolution snapshots for `bedrock_tool_…

8eb5a35

…result_colocatable_content` The new profile flag appears in the normalized profile snapshots for Anthropic (`{text, image}`), Meta, and Mistral (`frozenset()`) on Bedrock.

This was referenced Jul 3, 2026

fix(bedrock): correct bedrock_supported_media_kinds_in_tool_returns for Meta, Mistral, and Qwen #6254

Merged

fix(bedrock): only synthesize a leading user message for models that require it #6255

Merged

test(bedrock): note why the Nova/Llama co-location tests are unit, no…

b0617af

…t VCR

DouweM changed the title ~~fix(bedrock): prevent Converse ValidationException on consecutive tool_result and media attachment turns~~ fix(bedrock): handle toolResult attachment co-location per-model via bedrock_tool_result_colocatable_content Jul 3, 2026

DouweM added the douwebot reviews the PR that gets labeled and removes the label at the end label Jul 3, 2026

github-actions Bot removed the douwebot reviews the PR that gets labeled and removes the label at the end label Jul 3, 2026

DouweM reviewed Jul 4, 2026

View reviewed changes

DouweM added 3 commits July 4, 2026 00:12

fix(bedrock): type the tool-result content-kind profile flags as `Lit…

3e2204b

…eral` unions Address review: use `frozenset[Literal[...]]` instead of `frozenset[str]` for `bedrock_tool_result_colocatable_content` and `bedrock_supported_media_kinds_in_tool_returns`.

DouweM enabled auto-merge (squash) July 4, 2026 00:35

DouweM merged commit e3d7980 into pydantic:main Jul 4, 2026
61 checks passed

DouweM mentioned this pull request Jul 4, 2026

fix(bedrock): deliver unsupported tool-return media to models that isolate toolResult turns #6259

Merged

coderabbitai Bot mentioned this pull request Jul 7, 2026

Add Bedrock model profiles for Writer, Z.AI, and Moonshot AI, and refresh LatestBedrockModelNames #6318

Merged

2 of 3 tasks

This was referenced Jul 8, 2026

feat(FallbackModel): attribute failures to models and support (exception, model) fallback_on #6338

Closed

feat: ephemeral per-request context parts for capabilities and hooks #6343

Closed

dsfaccini mentioned this pull request Oct 23, 2026

Add Hypothesis property tests for message history repair, serialization and provider request mapping #11129

Merged

3 tasks done

Navigation Menu

Repository navigation

fix(bedrock): handle `toolResult` attachment co-location per-model via `bedrock_tool_result_colocatable_content` - #6098

fix(bedrock): handle `toolResult` attachment co-location per-model via `bedrock_tool_result_colocatable_content`#6098
DouweM merged 16 commits into
pydantic:mainfrom
Hasnaathussain:fix/6081-bedrock-tool-result-attachments-conflict

Hasnaathussain commented Jun 27, 2026 •

edited by DouweM

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

DouweM commented Jul 1, 2026

Uh oh!

Hasnaathussain commented Jul 1, 2026 •

edited

Loading

Uh oh!

coderabbitai Bot commented Jul 2, 2026 •

edited

Loading

Walkthrough

Changes

Sequence Diagram(s)

Uh oh!

coderabbitai Bot left a comment

Uh oh!

coderabbitai Bot left a comment

Uh oh!

coderabbitai Bot left a comment

Uh oh!

DouweM Jul 4, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

Hasnaathussain commented Jun 27, 2026 • edited by DouweM Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Problem

Evidence (live Converse, what may share a user message with a toolResult)

Fix

Tests

Follow-ups (separate PRs)

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Uh oh!

DouweM commented Jul 1, 2026

Uh oh!

Hasnaathussain commented Jul 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

coderabbitai Bot commented Jul 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Walkthrough

Changes

Sequence Diagram(s)

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

DouweM Jul 4, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Hasnaathussain commented Jun 27, 2026 •

edited by DouweM

Loading

Evidence (live Converse, what may share a user message with a `toolResult`)

Hasnaathussain commented Jul 1, 2026 •

edited

Loading

coderabbitai Bot commented Jul 2, 2026 •

edited

Loading