fix(blocks): return text payloads and correct provider completion reasons

[jpr5]

Text-only blocks fixtures currently end with a tool-call terminal even though no tools were emitted. Nonstreaming OpenAI chat also drops block content and tools entirely. This change derives block-path defaults from the resolved output: text finishes normally, tool-containing blocks retain tool-call terminals, and nonstreaming OpenAI chat returns the authored block payload.

The shared outcome feeds both streaming and nonstreaming paths for OpenAI chat, Claude messages, Gemini, Gemini Interactions, Cohere, Bedrock invoke, and Bedrock Converse. Existing supported finish overrides, legacy fallback for absent/empty blocks, block order, and provider-supported tool IDs remain intact. Nonempty blocks remain authoritative over conflicting legacy fields. Interactions retains its existing ignored finish-override behavior.

Local real red–green proof

Baseline: b1ba430658c0fac66dcf24464e7d44178553e477.
Verified candidate: 5c3d27461a038b7e6aa055f295ab3782f1991966 (tree f70d3873d3d35717437ef35c3b12e4f28eacfa23).

The same frozen probe exercised 96 actual HTTP responses from locally running LLMock: eight protocols including unchanged Ollama controls, both response modes, and text/tool/mixed/legacy/override/conflicting-legacy fixtures. JSON, SSE, NDJSON, and AWS EventStream responses were decoded from the wire; no fake responses or external LLM calls were used.

Commands run before production changes, from the integration checkout at the baseline:

pnpm build
node /Users/jpr5/.local/share/copilotkit/cr/aimock-section1/repro/probe.mjs /Users/jpr5/.local/state/copilotkit/worktrees/aimock/section1-r1-integration /Users/jpr5/.local/share/copilotkit/cr/aimock-section1/repro/red.json
node /Users/jpr5/.local/share/copilotkit/cr/aimock-section1/repro/check.mjs /Users/jpr5/.local/share/copilotkit/cr/aimock-section1/repro/red.json
# RESULT 63 passed; 33 failed; 96 real HTTP cases
# checker exit 1

Commands run independently after integration, from that checkout at the candidate:

pnpm build
node /Users/jpr5/.local/share/copilotkit/cr/aimock-section1/repro/probe.mjs /Users/jpr5/.local/state/copilotkit/worktrees/aimock/section1-r1-integration /Users/jpr5/.local/share/copilotkit/cr/aimock-section1/integrated-value/green.json
node /Users/jpr5/.local/share/copilotkit/cr/aimock-section1/repro/check.mjs /Users/jpr5/.local/share/copilotkit/cr/aimock-section1/integrated-value/green.json
# RESULT 96 passed; 0 failed; 96 real HTTP cases
# checker exit 0
shasum -a 256 -c /Users/jpr5/.local/share/copilotkit/cr/aimock-section1/repro/probe-sha256.txt
# probe.mjs: OK; check.mjs: OK

Build and probe commands exited 0 in both runs. The exact 33 failing cells flipped; all 63 passing controls stayed passing. Probe and checker were byte-identical between runs, verified against these SHA-256 hashes:

probe.mjs  012a6508d5ff6c18154199661d6523b0151e1c26dfaf3f58fbe06352e31f8ca9
check.mjs  968d495a16bb50965ec91bdbb66b865e70930cb19eb174dab6614b781ab0b89e

Actual text-only wire comparisons:

Surface	Before	After
OpenAI nonstream chat	content="", finish_reason="tool_calls"	content="BLOCK_TEXT", finish_reason="stop"
Claude streaming messages	stop_reason="tool_use"	stop_reason="end_turn"
Gemini nonstream generateContent	finishReason="FUNCTION_CALL"	finishReason="STOP"
Cohere streaming chat	finish_reason="TOOL_CALL"	finish_reason="COMPLETE"
Bedrock nonstream invoke	stop_reason="tool_use"	stop_reason="end_turn"

The standalone probe paths above are local evidence artifacts, not repository files. Full responses and cell comparison are retained under ~/.local/share/copilotkit/cr/aimock-section1/{repro,integrated-value}/. Repository-contained HTTP regressions are reproducible from the candidate checkout:

pnpm exec vitest run src/__tests__/fixture-blocks-outcome-openai.test.ts src/__tests__/fixture-blocks-outcome-claude-gemini.test.ts src/__tests__/fixture-blocks-outcome-bedrock.test.ts src/__tests__/fixture-blocks-outcome-cohere-interactions.test.ts
pnpm exec vitest run src/__tests__/fixture-blocks-*.test.ts src/__tests__/blocks-fixture-tolerance.test.ts
# Integrated broader command: 225 tests passed, 20 files passed, no skips

Each new suite was observed failing against its pre-fix implementation before passing after the fix. Controls cover empty text, both mixed orders, trailing empty text, tool arguments/IDs, overrides, legacy fallback, and payload plus terminal together.

Review and validation

All 17 valid review reports were adjudicated with zero mandatory findings. The independent promotion audit recommended zero promotions; no source changes or further confirmation round were required. Separate maintenance items are recorded in the review follow-ups.

Final pre-push checks passed in order, all exit 0:

pnpm run format:check
pnpm run lint
pnpm run typecheck
pnpm run test --maxWorkers=1 --no-file-parallelism
pnpm run build

The full suite passed 6,974 tests in 213 files, with zero skips or unhandled errors. The initial default-concurrency run had four timing failures and two worker RPC timeouts while host load exceeded 46. All four failed cases passed once in isolation with unchanged assertions and timeouts; the full gate then restarted with one worker to reduce contention. No source fixes, timeout changes, or test exclusions were used.

The four commits already separate the shared OpenAI change and provider adaptations by concern; all commit messages pass commitlint. The checked HEAD and tree above remain unchanged and clean. All 30 CI checks passed on 5c3d27461a038b7e6aa055f295ab3782f1991966, including the supported Node test matrix, Python tests, static quality, package exports, and Docker build.

[pkg-pr-new]

Open in StackBlitz

npm i https://pkg.pr.new/@copilotkit/aimock@467

commit: 5c3d274