[hpiclaranet]

Summary

Four ReAct stream-parser fixes in output_parser/cot_output_parser.py (+ the shared is_final_action() predicate and its three call sites in strategies/ReAct.py), each with a regression test that fails against the current code:

1. Stale prefix state across JSON blobs — after a completed JSON blob is flushed, last_character kept the closing }/], so a prefix right after a blob ({...}Action:{...}, {...}FinalAnswer:) was leaked as literal thought text (and }FinalAnswer: additionally blocked live answer streaming). Fix: reset last_character to a delimiter in the flush block and accept \t/\r as prefix delimiters.
2. Spurious "Action parse failed" flags — every non-action JSON blob in THINKING was flagged, so innocent JSON (citations [1], markdown links, {}, arbitrary objects) killed the tool round with a misleading error. Fix: only flag action-shaped blobs (emitted right after Action: or carrying action-ish keys); everything else is plain thought text.
3. FinalAnswer/final_answer not recognized as final — only the spaced "final answer" was treated as final, so the compact/underscore JSON action names fell into "there is not a tool named ..." and wasted a round. Fix: shared is_final_action() predicate (lower + strip _/spaces) used at the round-finality check and both assistant-history formatters; the SDK is_final() spellings remain covered.
4. Partial prefixes dropped at EOF — a stream cut mid-prefix (e.g. ...FinalAns) silently dropped the buffered characters. Fix: flush all three PrefixMatcher caches as text in the current state at stream end.

Also: malformed (non-dict/str) action_input (e.g. a list) is coerced to a JSON string so it never raises out of the parser (strategy-side handling is a documented follow-up).

Testing

• Full plugin suite: 96 passed, 15 skipped (baseline 76 passed) — 17 new tests across test_cot_output_parser.py and test_react.py, all verified to fail against the pre-fix code (stashed-source run: 17 failed) and pass with the fix.
• New coverage: citations/links/{}/arbitrary JSON not flagged; arbitrary JSON around a real action (real action wins); Action:-prefixed invalid JSON still flagged; }Action:/}FinalAnswer: adjacency; tab/CR before prefixes; truncated prefix at EOF; list action_input coercion; FinalAnswer/final_answer/final answer variants through the full strategy (tool not invoked, final answer streamed).
• ruff check on the touched files: zero new findings (10 pre-existing E402 from the tests' sys.path pattern, unchanged).
• Regressions fix(cot_agent): recover wrapped action JSON and surface ReAct parse failures #3703/fix(cot_agent): ReAct parser edge cases from #3705 (adjacent JSON, function envelopes, split think tags) #3706 preserved: tool_call wrappers, OpenAI function envelopes, adjacent roots, split think tags — all existing tests green.

Notes

• Manifest bumped to 0.0.50 (0.0.49 is claimed by the open draft feat(anthropic): add Claude Opus 5.5 #3919 dependencies-refresh batch; no collision).
• Scope: only the parser + the is_final_action predicate/call sites + tests + manifest version. No strategy-loop, budget, truncation, prompt, or YAML changes.

Release Notes

• Fixed ReAct stream parsing: tool calls directly after a JSON blob are no longer dropped or leaked as thought text (}Action: / }FinalAnswer:), and live final-answer streaming is no longer blocked.
• Innocent JSON in model thoughts (citations, links, {}) no longer triggers a spurious "Action parse failed" error.
• FinalAnswer and final_answer JSON action names are now recognized as final answers.
• A stream cut mid-prefix no longer loses the trailing characters.
• Malformed tool inputs (e.g. arrays) no longer crash the parser (they are surfaced as a tool invoke error instead).

Change Type

• Documentation / non-plugin change
• Non-LLM plugin (tools, extensions, datasource, etc.)
• Agent strategy plugin