[TINGyu123644]

Summary

Fixes #45200.

CotAgentOutputParser.parse_action assumed the model's JSON payload is always a single action-shaped object. Four shapes violate that, and each one aborts the agent run with an uncaught exception (the error is not caught in CotAgentRunner.run):

• a JSON list with more than one item (e.g. a direct answer that happens to be JSON data) or a non-object list ([] / [1]) raised AttributeError: 'list'/'int' object has no attribute 'items';
• a non-string action name ({"action": ["search"], ...}) raised a pydantic ValidationError from AgentScratchpadUnit.Action - as did a non-dict/str action_input ([1, 2], 5);
• an extra field after the action ({"action": "search", "action_input": {...}, "reason": "need data"}) silently changed the tool: the last non-input key won, so the agent looked for a tool named need data.

parse_action now:

• runs the first action-shaped item of a list (the prompt allows exactly one action per turn) and keeps a list without one as text, so a JSON direct answer reaches the user intact;
• keeps a payload as text when the action name is not a string or the action input is not a dict/str, instead of raising;
• takes the first non-input key as the action name, so trailing extra keys cannot override the tool;
• serializes non-action payloads with ensure_ascii=False, so non-ASCII text is no longer escaped to backslash-u sequences.

No change to the happy path: a single action-shaped object (or cohere's single-item list) still produces the same AgentScratchpadUnit.Action.

Screenshots

Before	After
run aborts with AttributeError / ValidationError, or calls the wrong tool	malformed payload is treated as text; the first action runs when the payload is a list of actions

Checklist

• This change requires a documentation update, included: Dify Document - no user-facing doc change (internal parser behavior)
• I understand that this PR may be closed in case there was no previous discussion or issues. (This doesn't apply to typos!) - linked issue ReAct agent crashes or picks the wrong tool on malformed action JSON #45200
• I've verified the change and added or updated tests where meaningful regression risk justifies coverage.
• I've updated the documentation accordingly. - not needed
• I ran lint on the touched files - repo ruff config (api/.ruff.toml, line-length 120) plus ruff format --check: clean. (The two unsorted-imports findings ruff 0.17.1 reports on these files also reproduce on the untouched upstream versions - a ruff-version isort difference, not from this change.) I could not run the full make lint / make type-check in my environment (the full backend toolchain isn't available); the added tests are self-contained and need no services.

Tests: 9 new cases in api/tests/unit_tests/core/agent/output_parser/test_cot_output_parser.py covering every shape above (RED verified without the fix - the crash cases raise, the wrong-tool case asserts the correct tool, the escape case asserts literal non-ASCII). The pre-existing 65 tests in that file still pass.

From WorkBuddy