Zhichang Yu
3f805a64f1
feat(agent): align Go agent behavior with Python (except retrieval component) ( #16225 )
...
## Summary
Aligns the **Go agent runtime/canvas/components/tools** behavior with
the **Python `agent/` implementation** so the same stored canvas DSL
produces the same execution result on either side. Every component,
tool, and runtime primitive in `internal/agent/` is now driven by the
same semantics as its Python counterpart — variable resolution, template
substitution, control flow, error reporting, retry/cancel, and stream
event shapes.
The **retrieval component is the one explicit exception** in this PR. It
is being reworked in a separate change and is excluded from this
alignment pass; the wrapper slot (`universe_a_wrappers.go →
newRetrievalComponent`) is preserved.
## Scope of alignment
### Components (all aligned with `agent/component/`)
`Begin` · `Message` · `LLM` (incl. ChatTemplateKwargs,
MessageHistoryWindowSize, VisualFiles, Cite, OutputStructure,
JSONOutput, TopP, MaxRetries, DelayAfterError, credentials) · `Agent`
(react + tool artifact capture + `Reset()` interface-assert) · `Switch`
(12/12 operators, Python-equivalent semantics) · `Categorize` · `Invoke`
· `Iteration` · `Loop` (macro-expansion through `workflowx.AddLoopNode`)
· `UserFillUp` (Python-equivalent interrupt/resume via eino
`compose.Interrupt`/`ResumeWithData`) · `FillUp` · `DataOperations` ·
`ListOperations` · `StringTransform` · `VariableAggregator` ·
`VariableAssigner` · `Browser` (full stagehand runtime parity) ·
`DocsGenerator` · `ExcelProcessor`.
### Tools (all aligned with `agent/tools/`)
`Retrieval` (wrapper slot only — logic out of scope) · `MCPToolAdapter`
(streamable-HTTP) · `CodeExec` (sandbox bridge with
`code_exec_contract.go` matching Python contract) · `AkShare` · `ArXiv`
· `Crawler` · `DeepL` · `DuckDuckGo` · `Email` · `ExeSQL` · `GitHub` ·
`Google` · `GoogleScholar` · `Jin10` · `PubMed` · `QWeather` · `SearXNG`
· `Tavily` · `Tushare` · `Wencai` · `Wikipedia` · `YahooFinance` —
uniform `eino tool.InvokableTool` interface, SSRF protection, shared
HTTP client.
### Canvas execution engine (`internal/agent/canvas/`)
Aligned with Python's `agent/canvas.py`:
- **Scheduler** (`scheduler.go`): state pre/post handlers, node lambdas,
per-component timeout resolver (4-level: per-class env → per-class table
→ uniform env → 600s fallback), `legacyNoOpNames`.
- **Loop subgraph** (`loop_subgraph.go`): Python-equivalent
`AddLoopNode` macro expansion + condition translation.
- **Multibranch** (`multibranch.go`): `Switch` / `Categorize` routing
via `compose.NewGraphMultiBranch` — same branch selection semantics as
Python.
- **Parallel subgraph** (`parallel_subgraph.go`): matches Python's
parallel fan-out contract.
- **Interrupt/Resume** (`interrupt_resume.go`): `UserFillUpNodeBody` /
`IsInterruptError` / `ExtractInterruptContexts` — replaces the
deprecated Python sentinel chain with eino's native interrupt API,
preserving the same external behavior.
- **Checkpoint** (`checkpoint_store.go`): `RedisCheckPointStore`
Get/Set/Delete, with business metadata (status / canvas_id /
parent_run_id) on a parallel Redis Hash.
- **RunTracker** (`run_tracker.go`): Start / MarkSucceeded / MarkFailed
/ MarkCancelled / AttachCheckpoint — same lifecycle as the Python run
record.
- **Cancel** (`cancel.go`): Redis pub/sub watch.
- **Stream** (`stream.go`): SSE channel with `messages` / `waiting` /
`errors` / `done` events, same shape as Python's `agent.canvas.RunEvent`
payload.
### DSL bridge (`internal/agent/dsl/`)
- `normalize.go`: v1↔v2 collapsed into a single wire format — Python and
Go consume the same stored JSON.
- `reset.go`: per-run state reset matches Python's `Canvas.reset()`
semantics.
- Testdata mirrors Python's `agent_msg.json` / `all.json` / etc.
### Runtime (`internal/agent/runtime/`)
- `CanvasState` / `NewCanvasState` / `GetVar` / `SetVar` / `ReadVars`:
same `{{cpn_id@param}}` resolution model.
- `ResolveTemplate` (regex fast path + gonja fallback) — Python
Jinja-style semantics.
- `selector.go`, `metrics.go`, `component.go`: shared runtime contracts.
## Out of scope (intentionally)
- **`Retrieval` component logic** — wrapped only; full parity lands in a
follow-up PR.
- **Frontend** — only minor dsl-bridge / canvas UX fixes ride along.
- **CLI / admin / model registry** — orthogonal to agent behavior.
## How alignment is verified
`internal/service/agent_run_e2e_test.go` exercises the **full production
chain** against real Python-shaped DSL fixtures:
```
loadCanvasForUser → versionDAO.GetLatest → decodeCanvasFromDSL →
canvas.Compile → cc.Workflow.Invoke → answer extraction
```
using in-memory SQLite + miniredis (no Docker). Covers:
- `TestRunAgent_RealCanvas_BeginMessage` — happy path, `{{sys.query}}`
resolution
- `TestRunAgent_RealCanvas_WaitForUserResume` — two-run resume cycle
(Python-equivalent)
- `TestRunAgent_RealCanvas_CompileFails` — unknown component name →
sanitized error (Python-equivalent)
- `TestRunAgent_RealCanvas_InvokeFails` — unresolvable template ref
(Python-equivalent)
- `TestRunAgent_RunTracker_AttachCheckpoint_CallSequence` —
Start→AttachCheckpoint→MarkSucceeded lifecycle
`internal/handler/agent_test.go` — SSE streaming parity (`Content-Type:
text/event-stream`, `data: {…}\n\n`, trailing `data: [DONE]\n\n`,
OpenAI-compatible non-stream `choices`).
`internal/agent/canvas/fixture_compile_test.go` + per-component tests
pin the Python-equivalent outputs.
```
go test -count=1 -v -run 'TestRunAgent_RealCanvas|TestRunAgent_RunTracker' ./internal/service/
```
## Design reference
`docs/develop/agent-go-port-design.md` (1329 lines, last cross-checked
2026-06-17) — module layout, per-component / per-tool inventory,
corner-case catalogue, and the actionable backlog (Section 14, including
the retrieval alignment follow-up).
---------
Co-authored-by: Claude <noreply@anthropic.com >
2026-06-22 11:58:29 +08:00
Jack
b363146997
refactor: overhaul task executor with layered architecture and comprehensive test suite ( #15471 )
...
## Summary
Decomposes the monolithic `task_executor.py` (1945 lines) into a 6-layer
architecture with clear separation of concerns. The refactored code is
functionally equivalent to the original, verified through 400 passing
tests and a production-vs-dry-run comparison framework.
## Architecture
```
entry (task_manager)
└─ orchestration (task_handler)
├─ services (chunk_service, embedding_service, dataflow_service, raptor_service, post_processor)
│ └─ utilities (chunk_builder, chunk_post_processor, embedding_utils)
└─ infrastructure (task_context, recording_context, interceptor)
```
Key design decisions:
- **TaskContext** — typed facade over raw task dict, injects rate
limiters + callbacks via composition
- **RecordingContext + Comparator** — enables side-by-side production vs
dry-run execution for safe migration
- **NullRecordingContext** — zero-allocation no-op for production, uses
`__slots__`
- **WriteOperationInterceptor** — FIFO replay of previous runs function
returns for comparison mode
## Migration Strategy
The original `handle_task()` in `task_executor.py` uses a 3-way switch
via `TE_RUN_MODE`:
- `TE_RUN_MODE=0` (default) → runs refactored code
- `TE_RUN_MODE=1` → runs both original + refactored, compares all
intermediate results
- `TE_RUN_MODE=2` → runs original code (fallback)
The comparison mode (`TE_RUN_MODE=1`) records ~40 intermediate values
(chunks, vectors, token counts, func return values) from the production
run and replays them during dry-run, then uses `ContextComparator` to
report mismatches.
## Functional Equivalence Fixes
All divergences between original and refactored code were identified and
fixed:
- Timeout decorators (handle/build_chunks/raptor/embedding)
- NullRecordingContext leak in finally block causing RuntimeError
- MinIO None-binary check with proper FileNotFoundError
- Dataflow dispatch after embedding binding + init_kb
- Memory task missing return after processing
- RAPTOR checkpoint progress reporting
- Tag cache (get_tags_from_cache/set_tags_to_cache) restoration
- dataflow_id correction in _load_dsl
- Language default Chinese, dead code guard removal
- embed_chunks made async with proper thread_pool_exec
- Full GraphRAG default configuration (10 parameters)
- Hardcoded q_768_vec fallback removal in RAPTOR
## Test Changes
- 20 new tests covering table parser manual mode, tag cache, embedding
edge cases, RAPTOR checkpoint, dataflow_id correction, storage binary
None, cancel cleanup, metadata=None boundary
- Unified `make_task_context`/`make_task_dict` factories eliminated 10+
duplicated helpers
- DataflowService tests migrated from internal method mocks to IO
boundary mocks (real orchestration code executes)
- Parametrized duplicate build_chunks post-processor tests
- 7 raptor tests modernized to @pytest.mark.asyncio
- Mock count per test reduced through boundary-level mocking strategy
**Test count: 400 passing, 0 warnings, 0 skips**
## Files Changed
| File | Change |
|------|--------|
| `rag/svr/task_executor.py` | +1 line (NullRecordingContext fix) |
| `rag/svr/task_executor_refactor/task_handler.py` | Orchestration
layer, 8 logic fixes |
| `rag/svr/task_executor_refactor/chunk_service.py` | +timeout +
None-check |
| `rag/svr/task_executor_refactor/embedding_service.py` | sync→async
rewrite |
| `rag/svr/task_executor_refactor/dataflow_service.py` | dataflow_id fix
+ timeout |
| `rag/svr/task_executor_refactor/raptor_service.py` | checkpoint fix +
assert |
| `rag/svr/task_executor_refactor/chunk_post_processor.py` | tag cache
restore |
| `rag/svr/task_executor_refactor/task_context.py` | language default
fix |
| `test/.../conftest.py` | +294 lines shared helpers |
| `test/.../*.py` | 15 test files refactored, 20 new tests |
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com >
2026-06-03 17:18:31 +08:00
wdeveloper16
14c0985182
feat: bump Python minimum from 3.12 to 3.13, drop strenum backport ( #14767 )
...
Closes #14753
## What changed
| File | Change |
|---|---|
| `pyproject.toml` | `requires-python` → `>=3.13,<3.15`; remove
`strenum==0.4.15` |
| `Dockerfile` | `uv python install 3.13`, `uv sync --python 3.13` |
| `.github/workflows/tests.yml` | `uv sync --python 3.13` on both matrix
legs |
| `CLAUDE.md` | dev setup command + requirements note updated |
| `deepdoc/parser/mineru_parser.py` | `from strenum import StrEnum` →
`from enum import StrEnum` |
| `agent/tools/code_exec.py` | same |
`StrEnum` has been in the stdlib since Python 3.11 — the `strenum`
backport package is no longer needed once the floor is 3.13.
## Why uv.lock is not regenerated
`uv lock --python 3.13` fails because:
1. The infiniflow/graspologic fork pins `numpy>=1.26.4,<2.0.0`
2. `tensorflow-cpu>=2.20.0` (the first release with cp313 wheels)
depends on `ml-dtypes>=0.5.1`, which requires `numpy>=2.1.0`
3. These two constraints are irreconcilable on Python 3.13
The lockfile regeneration requires loosening the `numpy` upper bound in
the `infiniflow/graspologic` fork. Once that fork commit is updated and
the SHA in `pyproject.toml:49` is bumped, `uv lock --python 3.13` will
succeed.
## RFC corrections
Two claims in the original RFC (#14753 ) did not hold up under code
review:
- **"graspologic hard-blocks 3.13"** — the infiniflow fork at the pinned
commit has no `<3.13` Python constraint. The blocker is the transitive
`numpy<2.0.0` conflict with tensorflow-cpu's test dependency, not a
direct Python version cap.
- **"free-threading throughput gains for I/O-bound workload"** — Python
3.13 free-threading requires a special `--disable-gil` build and
provides no benefit for async I/O code (the GIL is already released
during I/O). The real motivation is forward compatibility and improved
error messages.
2026-05-15 14:40:53 +08:00
