refactor(tools): extract supervised batch execution engine

CHANGELOG.md

@@ -17,6 +17,7 @@ GitHub Releases page; `0.8.0` is the new starting line.
 
 - **Parallel streamed tool calls are now correlated safely.** Interleaved argument chunks stay attached to their indexed calls, malformed or truncated call streams stop before tool execution, and failed attempts are not retried after output has already been shown.
 - **Provider compatibility and Z.AI routing are now explicit.** Immutable compatibility profiles keep request-format quirks behind the chat-provider boundary, while independent Z.AI Coding Plan and API login routes use separate credentials, endpoints, model identities, catalog refresh, logout, and usage/rate-limit state. Curated GLM requests now apply exact context/output limits, thinking controls, reasoning replay, and tool-stream support without activating for local or unknown models.
+- **Tool execution is now supervised as a terminal batch.** A private execution engine preserves the Toolset registry and legacy per-call API while centralizing ordered results, deduplication, callbacks, and batch summaries; cancellation is bounded, late work stays owned, and new batches fail closed until timed-out cleanup drains.
 
 ## 0.58.0 (2026-07-11)
 

CONTRIBUTING.md

@@ -48,3 +48,19 @@ If you believe a new runtime dependency is genuinely necessary:
    automatically so reviewers know to look for the justification.
 
 Dev-only dependencies under `[dependency-groups]` are not subject to this policy.
+
+## Tool-execution characterization
+
+Changes to `soul/tool_execution.py`, execution scheduling, deduplication, the reader/writer gate, or
+MCP publication should run the deterministic local characterization harness:
+
+```bash
+uv run python scripts/benchmark_toolset.py --scenario all --runs 5 --output before-or-after.json
+```
+
+For before/after evidence, use the same machine, Python environment, fixture matrix, warm-up count,
+and five-run command at both revisions. Keep the raw JSON and report timing regressions as well as
+improvements; do not discard samples or use `--smoke` for final evidence. Confirm every scenario has
+zero leaked tasks/processes/sessions, all cancellation and recovery flags are true, registry hashes
+are stable, and deterministic decision states are derived by the harness. Treat results as directional
+local engineering evidence, not universal product telemetry.

docs/en/customization/agent-architecture.md

@@ -180,8 +180,9 @@ sequenceDiagram
 ```mermaid
 flowchart LR
   Core[pythinker_core.step]
-  Call[ToolCall]
-  Toolset[PythinkerToolset.handle]
+  Calls[Terminal ToolCall batch]
+  Toolset[PythinkerToolset.handle_batch]
+  Engine[ToolExecutionEngine]
   Parse[Parse JSON arguments]
   Pre[PreToolUse hooks]
   Execute[tool.call(arguments)]
@@ -193,10 +194,13 @@ flowchart LR
   Post[PostToolUse or PostToolUseFailure hooks]
   Telemetry[Telemetry]
   Result[ToolResult]
+  Batch[ToolBatchHandle]
+  Step[StepResult.tool_results]
+  Soul[PythinkerSoul]
   Context[tool_result_to_message -> Context]
   Wire[Wire event stream]
 
-  Core --> Call --> Toolset --> Parse --> Pre --> Execute
+  Core --> Calls --> Toolset --> Engine --> Parse --> Pre --> Execute
   Execute --> Builtin
   Execute --> Plugin
   Execute --> MCP
@@ -208,12 +212,26 @@ flowchart LR
   Builtin --> Result
   Approval --> Result
   Result --> Post --> Telemetry
-  Result --> Core
+  Result --> Batch --> Step --> Soul
   Result --> Wire
-  Result --> Context
+  Soul --> Context
 ```
 
-The toolset is both a registry and an execution boundary. It hides tools from the LLM when needed, validates tool names, parses JSON arguments, triggers hooks, converts exceptions to `ToolRuntimeError`, and returns async `ToolResult` tasks to `pythinker_core.step`. The advertised tool list is filtered by the active execution profile, subagent/root role, plan-mode state, and hard permission profile before each model call; tool-specific execution guards still run even if a hidden tool is somehow called. MCP tools are registered as local wrappers. Wire external tools are sent to the active Wire client as `ToolCallRequest` messages and wait for a client-provided result.
+The toolset is both a registry and an execution boundary. For Pythinker Soul, `pythinker_core.step`
+passes the complete terminal call batch and immutable `ToolBatchContext` to
+`PythinkerToolset.handle_batch`. Its non-exported `ToolExecutionEngine` validates tool names, parses
+JSON arguments, applies duplicate/consecutive-call policy, schedules calls through the reader-writer
+gate, triggers hooks, converts exceptions to `ToolRuntimeError`, and supervises ordered results and
+bounded cancellation through a `ToolBatchHandle`. `StepResult.tool_results()` settles that owned
+batch before Soul appends the assistant and tool messages to `Context`. Core's per-call
+`Toolset.handle()` path remains only as the compatibility fallback for toolsets that do not
+implement `BatchToolset`.
+
+The advertised tool list is filtered by the active execution profile, subagent/root role,
+plan-mode state, and hard permission profile before each model call; tool-specific execution guards
+still run even if a hidden tool is somehow called. MCP tools are registered as local wrappers. Wire
+external tools are sent to the active Wire client as `ToolCallRequest` messages and wait for a
+client-provided result.
 
 ## Subagent graph
 

docs/en/customization/architecture.md

@@ -66,8 +66,10 @@ The end-to-end flow when a session starts and processes a turn:
 4. **Core loop** — `src/pythinker_code/soul/pythinkersoul.py:PythinkerSoul.run` handles user
    input and slash commands, calls the LLM through `pythinker_core.step`, runs tools, gates
    side effects through approvals, injects dynamic reminders, and compacts the context.
-5. **Tool execution** — `src/pythinker_code/soul/toolset.py:PythinkerToolset` loads built-in
-   and MCP tools, injects dependencies, executes calls, and returns structured results.
+5. **Tool execution** — `src/pythinker_code/soul/toolset.py:PythinkerToolset` owns the
+   built-in/MCP registry, visibility, and dependency injection facade. The private
+   `src/pythinker_code/soul/tool_execution.py:ToolExecutionEngine` executes terminal batches,
+   deduplicates calls, orders results, and supervises bounded cancellation.
 6. **Wire and UI** — `src/pythinker_code/soul/run_soul` connects the soul to
    `src/pythinker_code/wire/`; Shell, Print, ACP, Web, and Dashboard frontends consume Wire events.
 
@@ -99,7 +101,8 @@ canonical list lives in `src/pythinker_code/wire/types.py` (`Event` union): `Ste
 | `src/pythinker_code/soul/pythinkersoul.py` | Core loop: user input, slash commands, LLM calls, tool runs, compaction, telemetry spans. | `PythinkerSoul`, `PythinkerSoul.run`, `FLOW_COMMAND_PREFIX` |
 | `src/pythinker_code/soul/agent.py` | `Runtime` and `Agent` construction, system-prompt assembly, AGENTS.md discovery. | `Runtime`, `Agent`, `load_agent`, `load_agents_md`, `BuiltinSystemPromptArgs` |
 | `src/pythinker_code/soul/context.py` | Conversation history, checkpoints, JSONL persistence. | `Context` |
-| `src/pythinker_code/soul/toolset.py` | Loads built-in + MCP tools, injects deps, executes calls. | `PythinkerToolset` |
+| `src/pythinker_code/soul/toolset.py` | Owns built-in + MCP registration, visibility, dependency injection, and the legacy per-call facade. | `PythinkerToolset` |
+| `src/pythinker_code/soul/tool_execution.py` | Private batch execution state machine: preparation, deduplication, reader/writer scheduling, callbacks, ordered results, bounded cancellation, poison, and late-task recovery. | `ToolExecutionEngine` |
 | `src/pythinker_code/soul/slash.py` | Slash-command registry and dispatch. | `registry` |
 | `src/pythinker_code/soul/dynamic_injection.py` (+ `dynamic_injections/`) | Injects budgeted `<system-reminder>` content per step: plan-mode, auto-mode, model-defense, LSP diagnostics. | `DynamicInjectionProvider` |
 | `src/pythinker_code/soul/permission.py` | Per-step permission profiles (`read_only`/`plan`/`ask`/`implement`/`review`/`verify`) and destructiveness classification. | `tool_destructive_reason`, `shell_command_signature` |
@@ -299,7 +302,7 @@ Both frontends build with `tsc -b && vite build` and are synced into the Python
 
 | Path | Purpose | Key entry points and interfaces |
 | --- | --- | --- |
-| `packages/pythinker-core/` | LLM abstraction: message models, streaming chat providers, tool abstractions, and the `generate`/`step` primitives. Independently versioned (1.x). | `generate`, `step`, `Message`, `ContentPart`, `ToolCall`, `ChatProvider`, `Toolset`, `ToolReturnValue`/`ToolOk`/`ToolError`, `CallableTool2`, `DisplayBlock`; contrib providers (`Anthropic`, `GoogleGenAI`, `OpenAIResponses`) and `LinearContext` |
+| `packages/pythinker-core/` | LLM abstraction: message models, streaming chat providers, optional batch/legacy tool dispatch, and the `generate`/`step` primitives. Independently versioned (1.x). | `generate`, `step`, `StepResult`, `Message`, `ContentPart`, `ToolCall`, `ChatProvider`, `Toolset`, `BatchToolset`, `ToolBatchHandle`, `ToolReturnValue`/`ToolOk`/`ToolError`, `CallableTool2`, `DisplayBlock`; contrib providers (`Anthropic`, `GoogleGenAI`, `OpenAIResponses`) and `LinearContext` |
 | `packages/pythinker-host/` | OS abstraction for filesystem + shell across local and SSH backends via a context-var-dispatched `Host` protocol. | `Host`, `HostPath`, `LocalHost`, `HostProcess`, `get_current_host`/`set_current_host` |
 | `packages/pythinker-review/` | Standalone review/security/debug engine and stateful Reviewflow; strict Pydantic schemas with fail-closed evidence validation. State in `.pythinker-review/` and `.pythinker-review-flow/`. See `packages/pythinker-review/AGENTS.md`. | `run_engine`, `ReviewLLM`, `Finding`, `RawFinding`, `ReviewerOutput`, Reviewflow `init`/`map`/`review`/`fix` |
 | `packages/pythinker-code/` | Thin distribution package exposing the `pythinker-code` script. | — |

docs/en/release-notes/changelog.md

@@ -19,6 +19,7 @@ GitHub Releases page; `0.8.0` is the new starting line.
 
 - **Parallel streamed tool calls are now correlated safely.** Interleaved argument chunks stay attached to their indexed calls, malformed or truncated call streams stop before tool execution, and failed attempts are not retried after output has already been shown.
 - **Provider compatibility and Z.AI routing are now explicit.** Immutable compatibility profiles keep request-format quirks behind the chat-provider boundary, while independent Z.AI Coding Plan and API login routes use separate credentials, endpoints, model identities, catalog refresh, logout, and usage/rate-limit state. Curated GLM requests now apply exact context/output limits, thinking controls, reasoning replay, and tool-stream support without activating for local or unknown models.
+- **Tool execution is now supervised as a terminal batch.** A private execution engine preserves the Toolset registry and legacy per-call API while centralizing ordered results, deduplication, callbacks, and batch summaries; cancellation is bounded, late work stays owned, and new batches fail closed until timed-out cleanup drains.
 
 ## 0.58.0 (2026-07-11)