emecii opened a new pull request, #1107:
URL: https://github.com/apache/flink-agents/pull/1107
Linked issue: #936
### Purpose of change
Anthropic responses that stop at the configured output-token limit now reach
the shared chat action
as `extra_args["finish_reason"] == "length"`. The action already rejects
that canonical value before
it dispatches incomplete content, so callers receive its clear truncation
error instead of parsing
or consuming a partial response.
Both Anthropic connections read the SDK response stop reason after
converting the content. They map
Anthropic's `max_tokens` value to the existing cross-provider `length` value
and retain every other
reported value unchanged.
### Behavioral Semantics
#### Key decisions
Only `max_tokens` is normalized because it is Anthropic's name for the
shared action's established
token-limit condition. Other reasons remain provider values so that an
unknown future reason is
observable without changing control flow.
#### Interaction decisions
| Anthropic stop reason | Stored `finish_reason` | Shared action behavior |
| --- | --- | --- |
| absent | absent | Existing behavior |
| `max_tokens` | `length` | Rejects incomplete response before dispatch |
| any other supplied value | same value | Existing accepted-reason behavior |
#### Behavioral contracts
- A Java or Python Anthropic response stopped by `max_tokens` exposes the
canonical `length`
metadata that the shared action recognizes.
- A supplied non-token-limit stop reason remains available without being
reclassified.
- A response without a stop reason gains no metadata entry.
#### Failure behavior
This converter change does not catch, retry, or transform provider failures.
When the model returns
`max_tokens`, the existing shared action raises its established truncation
error; direct connection
callers can inspect the response metadata before invoking that action.
### Tests
| Contract | Tests |
| --- | --- |
| `max_tokens` becomes `length`; ordinary reasons remain observable | Java
`AnthropicChatModelConnectionTest` (94 passed); Python
`test_anthropic_response_parsing.py` |
| Shared action rejects `length` before dispatch | Java
`ChatModelActionTest` and `ChatModelActionRetryTest` (36 passed); Python
`test_chat_model_action.py` |
Focused Python run: 107 passed across the Anthropic response parser and
shared-action tests.
Spotless, Ruff format/check, and Apache RAT license checks pass.
Not verified: live Anthropic API calls. Tests use SDK response objects and
local action fixtures, so
they require no credentials or network access.
<details>
<summary>Implementation evidence</summary>
The Python converter stores the key independently of token-usage metrics.
The Java converter reads
the optional SDK stop reason after constructing the response and does the
same. This keeps the
mapping available for responses with or without recorded usage and for both
plain-text and tool-call
conversion paths.
</details>
### API
No public API signature changes. Existing consumers of `extra_args` gain
`finish_reason` on
Anthropic responses; the shared action now classifies Anthropic token-limit
responses consistently
with OpenAI-family responses.
### Documentation
- [ ] `doc-needed`
- [x] `doc-not-needed`
- [ ] `doc-included`
### Was this patch authored or co-authored using generative AI tooling?
- [x] Yes
- [ ] No
Generated-by: Codex CLI 0.153.4 (GPT-6)
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
