Skip to content

README.md

Latest commit

 

History

History
285 lines (222 loc) · 28.3 KB

README.md

File metadata and controls

285 lines (222 loc) · 28.3 KB

Plan and Task E2E Example

This example demonstrates an interactive plan→review→execute workflow using the ECS-based LLM Agent framework. It features a robust state machine, review-gated planning, artifact persistence, recovery semantics, and framework-native auto compaction for both the main agent and spawned subagents.

Overview

The workflow follows a structured lifecycle:

  1. Draft Interview: The agent interviews the user to build a draft plan (DRAFT_INTERVIEW).
  2. Draft Reviews: The draft must be approved by both an Advisor (DRAFT_ADVISOR_REVIEW) and a QA subagent (DRAFT_QA_REVIEW).
  3. Write Plan: Once QA approves the draft, the system automatically transitions to WRITE_PLAN and triggers a dedicated plan_writer subagent (equipped with the writing-plans skill) to convert the approved draft into a structured workflow_plan.md. No manual command is needed.
  4. Plan QA Review: When the plan writer finishes, the system automatically transitions to PLAN_QA_REVIEW where QA reviews the final plan document.
  5. Execution: Once finalized, the plan is decomposed into a task queue and executed.

Architecture

  • Built-in Tools — The main agent has read_file, write_file, edit_file, bash, and glob tools pre-installed via BuiltinToolsSkill, workspace-bound to the example directory. edit_file uses a hash-anchored interface: supply op, pos, optional end, and content, with pos/end in "N#HASH" format obtained from a prior read_file call. The main agent has unrestricted access to all these tools.
  • Subagent Tool Permissions — advisor, qa, and plan_qa review subagents inherit only read_file and glob (read-only) via InheritancePolicy(inherit_tools=["read_file", "glob"], inherit_permissions=True). They cannot write files or run shell commands. plan_writer inherits read_file, write_file, edit_file, and glob since it must produce the final plan document.
  • Two Distinct QA Subagents — Draft QA (qa) and Plan QA (plan_qa) are registered as separate subagents with separate system prompts and separate state machine transition paths:
    • qa uses QA_SYSTEM_PROMPT (draft review lens) and routes DelegationCompletedEvent → controller.handle_qa_review() → DRAFT_QA_REVIEW verdict.
    • plan_qa uses PLAN_QA_REVIEW_SYSTEM_PROMPT (final plan review lens) and routes DelegationCompletedEvent → controller.handle_plan_qa_review() → PLAN_QA_REVIEW verdict.
    • The planner system prompt calls subagent(category="qa", ...) for draft review and subagent(category="plan_qa", ...) for plan review.
  • Review Prompts via File Path — When invoking advisor, qa, plan_qa, or plan_writer subagents for review, the prompt passes the artifact file path (e.g. scratchbook/<workflow_id>/plan/draft.md) rather than embedding the file content inline. The subagent reads the file itself using read_file, avoiding prompt token bloat.
  • Auto Compaction — build_plan_task_world(...) installs CompactionConfigComponent(threshold_tokens=300_000, compaction_method="predrop_then_compact") by default, ConversationArchiveComponent, and CompactionSystem at priority -30 so compaction runs before workflow prompt rendering and before reasoning. SystemPromptRenderSystem then injects the current summary into the effective system prompt as <chat_history_summary>...</chat_history_summary> XML.
  • Subagent Compaction Inheritance — Child worlds created by SubagentSystem inherit the parent CompactionConfigComponent, receive their own ConversationArchiveComponent, and register CompactionSystem at the same priority. Long-running review and task subagents therefore compact independently without requiring plan-and-task-specific special cases.
  • Workflow Reset Safety — /plan:start, /plan:resume, and /task:start <workflow_id> clear stale CurrentCompactionSummaryComponent, reset archived summaries, and invalidate RenderedSystemPromptComponent before restoring or switching workflow state. This prevents an old summary from leaking into a newly loaded workflow phase.
  • Log Truncation — Structured log fields last_user_prompt and user-normalization prompt_text are truncated to 200 characters to keep logs readable without losing signal. System-prompt render logs still report prompt_length, but the rendered prompt text itself is not truncated in this example.
  • ECS Core: Uses SystemPromptRenderSystem, UserPromptNormalizationSystem, ReasoningSystem, and ToolExecutionSystem.
  • Prompt Configuration: The planner entity declares SystemPromptConfigSpec with DRAFT_INTERVIEW_SYSTEM_PROMPT, and SystemPromptRenderSystem bridges the rendered value into LLMComponent.system_prompt before reasoning.
  • Workflow DSL: Uses install_workflow and WorkflowStateSystem (priority -25) to manage the phase graph and automatic prompt-profile selection via ${_workflow_state_prompt}.
  • State Machine: Explicit phase transitions managed by WorkflowStateMachine.
  • Artifacts: Durable persistence of plans, state, and execution evidence via PlanTaskScratchbookAdapter. Main-agent tool results are currently kept inline in ECS conversation/tool-result state rather than being written through ToolResultsSink.
  • Controller: PlanController manages the high-level workflow logic and review gates.
  • Subagent Reviews: Advisor, QA, and Plan QA review steps are wired as ECS subagents via SubagentRegistryComponent. The planner invokes them with subagent(category="advisor", ...), subagent(category="qa", ...), and subagent(category="plan_qa", ...) respectively. Verdicts are automatically extracted from subagent results via DelegationCompletedEvent subscription, routed to the correct controller method based on the subagent name.
  • Plan Writer Subagent: The WRITE_PLAN phase is executed by a dedicated plan_writer subagent registered in SubagentRegistryComponent. It is pre-loaded with the writing-plans skill (discovered from .claude/skills/writing_plans/SKILL.md) and inherits read_file, write_file, edit_file, and glob tools. When it completes, handle_write_plan_completed() transitions the state to PLAN_QA_REVIEW.
  • Task Execution: TaskExec handles plan loading, dependency resolution, and subagent dispatch.
  • Slash Commands: Dispatched via ECS TriggerSpec script handlers on UserPromptConfigComponent. Commands appear as transformed messages in conversation history.
  • System Execution Order: UserInputSystem runs at priority -15 (before UserPromptNormalizationSystem at -10). This ensures the user's message is already in ConversationComponent when script handlers fire, so slash commands like /task:start are matched in the same tick they are entered.

Supported Commands

The interactive runtime supports eleven slash commands:

  • /plan:start <description>: Initialize a new workflow with a draft description.
  • /plan:resume <workflow_id>: Restore a previously-started workflow from disk by its workflow ID (e.g. creative-writing-assistant-with-llm-workflow). Marks any in-flight subagents as stale and resumes from the persisted phase.
  • /plan:status: Show the current workflow phase, status, and review verdicts.
  • /plan:finalize: Finalize the plan and transition to task execution (requires all three approved reviews).
  • /plan:write: Transition from DRAFT_QA_REVIEW to WRITE_PLAN phase to produce workflow_plan.md. Optional — this transition now happens automatically when QA approves the draft, but can still be invoked manually.
  • /plan:qa_review <approved|revise|blocked> [notes]: Record a QA verdict on the final plan document.
  • /task:start <workflow_id>: Start execution of a specific task. If no workflow is active in the current session, providing a workflow_id auto-loads the persisted state from scratchbook (equivalent to /plan:resume <workflow_id> followed by starting task execution). Accepts phases PLAN_FINALIZED, TASK_READY, TASK_RUNNING, and TASK_BLOCKED.
  • /task:status: Show the status of the current task and subagent sessions.
  • /task:resume: Resume a blocked or replanned task.
  • /task:replan <reason>: Request a replan for the current task.
  • /task:abort: Abort the current task and transition to a terminal state.

Artifact Layout

All workflow data is persisted in scratchbook/<workflow_id>/:

  • plan/: Contains draft.md (working draft, included as draft_plan artifact) and workflow_plan.md (the single living plan file, edited in-place).
  • state/: Contains runtime_state.json, events.jsonl, and task_queue.json.
  • memory/: Contains knowledge.jsonl for cross-task context.
  • evidence/: Directory for task execution artifacts.
  • review/: Contains JSON verdicts from Advisor and QA reviews.

Main-agent tool call results are not currently persisted as separate canonical records by this example. They remain inline in ECS tool result state and conversation tool messages while durable workflow artifacts continue to live under scratchbook/<workflow_id>/.

Usage

Interactive Mode

Run the entry point to start an interactive session.

LLM_API_KEY=your-api-key uv run python examples/e2e/plan_and_task/main.py

Multi-line input

The prompt supports multi-line messages. Press Enter to start a new line; submit with a blank line (press Enter on an empty line):

You> /plan:start 我想开发一个辅助写作软件，
... 支持长篇小说和剧本创作，
... 需要多 Agent 协作完成各章节生成。
...
         ↑ blank line submits

Single-line commands work as before — just type and press Enter, then Enter again on the empty continuation line:

You> /plan:status
...

exit or quit typed as the first line (followed by Enter + blank line) terminates the session. Ctrl+D (EOF) also exits cleanly.

Automation Mode (piped input)

Automate interactions by piping commands. In pipe mode each \n\n (double newline) acts as a submit boundary:

printf '/plan:start Build demo\n\n/plan:status\n\nexit\n\n' | uv run python examples/e2e/plan_and_task/main.py

Recovery / Restart

The workflow can be restarted at any time. On startup, no workflow ID is resolved and no scratchbook folder is created. Instead:

  1. Call /plan:start <original description> — the LLM re-derives the same slug from the same description (or uses slug_from_description() as fallback).
  2. State is restored from scratchbook/<workflow_id>/state/runtime_state.json.
  3. Any in-flight subagents are marked stale and the machine transitions to TASK_BLOCKED for safe resumption.

Note: Use the same description text (or the same slug) as the original /plan:start call so the derived workflow ID matches the existing scratchbook directory.

Mid-flight State Reconciliation

When /plan:resume <workflow_id> is called, the system reads the persisted state and automatically reconciles any in-progress phases so the workflow can continue without manual intervention:

Resumed phase Condition Automatic action
DRAFT_QA_REVIEW review_verdicts contains an approved verdict for this phase Transitions to WRITE_PLAN, injects a write-plan trigger message to start the plan_writer subagent
WRITE_PLAN (any — plan_writer was mid-flight) Injects a write-plan trigger message to restart the plan_writer subagent
PLAN_QA_REVIEW review_verdicts contains an approved verdict for this phase Transitions to PLAN_FINALIZED
All other phases — No automatic action; resumes normally

This means after a process restart you can call /plan:resume <workflow_id> and, if QA had already approved the draft before the restart, the plan_writer will be triggered automatically — no need to manually issue /plan:write.

Testing

Integration tests

Run the integration suite to verify command parsing, state machine logic, artifact persistence, and credential-gated CLI coverage:

uv run pytest tests/integration/test_plan_and_task_flow.py -v
  • uv run pytest tests/integration/test_plan_and_task_flow.py -k "subagent" — verifies subagent component wiring
  • uv run pytest tests/integration/test_plan_and_task_flow.py -k "compaction or stale_compaction_state" -v — verifies main-agent auto compaction, subagent inheritance, and stale-summary reset on workflow switch/resume

Specific test filters

uv run pytest tests/integration/test_plan_and_task_flow.py -k "commands"
uv run pytest tests/integration/test_plan_and_task_flow.py -k "artifacts"

Real-LLM acceptance tests

Requires LLM_API_KEY. Verifies the controller and task execution with a real model:

LLM_API_KEY="$LLM_API_KEY" \
LLM_BASE_URL=https://dashscope.aliyuncs.com/api/v2/apps/protocols/compatible-mode/v1 \
LLM_MODEL=qwen3.5-flash \
uv run pytest tests/live/test_plan_and_task_flow_live.py -v

To exercise the new compaction path against an Anthropic-compatible endpoint:

LLM_API_FORMAT=anthropic_messages \
LLM_BASE_URL=https://cc2.caaa.tech \
LLM_API_KEY="$LLM_API_KEY" \
LLM_MODEL=glm-5.1 \
uv run pytest tests/live/test_plan_and_task_flow_live.py::test_anthropic_plan_task_auto_compaction_summarizes_context -v

Environment Variables

  • LLM_API_KEY: API key for the chosen provider.
  • LLM_BASE_URL: API base URL (defaults to DashScope Responses API: https://dashscope.aliyuncs.com/api/v2/apps/protocols/compatible-mode/v1).
  • LLM_MODEL: Model ID (defaults to qwen3.5-flash).
  • LLM_API_FORMAT: Provider/format selector. Accepted values:
    • openai_responses (default) — OpenAI Responses API via OpenAIModel (enables enable_store=True for prefix caching)
    • openai_chat_completions — OpenAI Chat Completions API via OpenAIModel
    • anthropic_messages — Anthropic Messages API via ClaudeModel (also works with Kimi-compatible Anthropic endpoints)
  • PLAN_TASK_LANGFUSE: Set to 1, true, yes, or on to install Langfuse observability on the plan-and-task World before Runner.run() starts.
  • PLAN_TASK_LANGFUSE_ENVIRONMENT: Optional Langfuse environment label. Defaults to plan-and-task.
  • PLAN_TASK_LANGFUSE_RELEASE: Optional release label sent with plan-and-task traces.
  • PLAN_TASK_LANGFUSE_SESSION_ID: Optional session ID for grouping plan-and-task traces. The Langfuse SDK v4 adapter sends this as a trace-level session attribute via propagate_attributes(...); metadata-only session IDs do not power the Langfuse Sessions UI.
  • LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY, and LANGFUSE_HOST or LANGFUSE_BASE_URL: Langfuse connection settings used when PLAN_TASK_LANGFUSE is enabled.
  • PLAN_TASK_INTERACTIVE: Set to 0 to disable interactive stdin.
  • DEBUG: Set to 1 to make this example call configure_logging() with debug logging. All plan_task_* structured log events will then appear on stderr via structlog.

Langfuse Observability

Install the optional extra before enabling Langfuse for this example:

uv pip install -e ".[langfuse]"

When PLAN_TASK_LANGFUSE is enabled, main.py calls install_plan_task_langfuse_observability() after build_plan_task_world(...) creates the World and before Runner.run(...) starts. In interactive mode, every UserInputReceivedEvent starts a user.turn trace that covers the complete chain from that user input until the next user input or process exit: prompt normalization, retrieval/compaction, LLM generations, tool calls, subagent spans, retries, errors, context pressure, and completion scores all stay inside that turn trace. One-shot runs without interactive input keep the runner trace for backward compatibility. Completed LLM, tool, and subagent observations export their ECS-recorded end timestamps through the Langfuse SDK v4 public lifecycle; preserving historical start timestamps requires explicitly validating your SDK version and enabling LangfuseConfig(enable_private_v4_historical_otel=True) because that path uses private SDK hooks. Raw prompts, tool arguments, and outputs are captured by default for backward compatibility; use LangfuseConfig(capture_input=False, capture_output=False) if raw content should not leave the process.

Subagents are exported as subagent.<name> spans inside the active user.turn trace. Their child-world LLM calls are exported as generation observations under that subagent span, and child-world tool/retrieval/API work is exported as child spans/events under the same turn trace rather than creating another top-level Langfuse trace. When a child-world generation requests a tool, that tool observation stays attached to the requesting generation so the Langfuse hierarchy shows the exact delegation chain.

Use environment variables or a secret manager for LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY, and the Langfuse host value. Do not put concrete keys in scripts, docs, or command history. On exit, the CLI calls flush() and shutdown() on the observability handle so buffered trace events are sent before the process terminates.

Anthropic-compatible Langfuse smoke run:

PLAN_TASK_LANGFUSE=1 \
PLAN_TASK_LANGFUSE_ENVIRONMENT=dev \
PLAN_TASK_LANGFUSE_RELEASE=local-test \
PLAN_TASK_LANGFUSE_SESSION_ID="plan-task-dev-1" \
LLM_API_FORMAT=anthropic_messages \
LLM_MODEL=deepseek-v4-flash \
uv run python examples/e2e/plan_and_task/main.py

Set LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY, LANGFUSE_BASE_URL, LLM_BASE_URL, and LLM_API_KEY in your shell or secret manager before running the command.

Anthropic / Kimi Example

DEBUG=1 \
  LLM_API_FORMAT=anthropic_messages \
  LLM_BASE_URL=https://api.anthropic.com \
  LLM_API_KEY=sk-... \
  LLM_MODEL=kimi-for-coding \
  uv run python examples/e2e/plan_and_task/main.py

The ClaudeModel appends /v1/messages to LLM_BASE_URL, so the actual endpoint called is https://api.anthropic.com/v1/messages. Anthropic cache stats (cache_read_input_tokens, cache_creation_input_tokens) are normalized and surfaced as plan_task_llm_cache_stats events with cache_hit_rate.

Log Events (observable with DEBUG=1)

This example explicitly enables logging in main.py when DEBUG=1; the base ecs-agent library remains silent until configure_logging() is called.

Event Level File Description
plan_task_workflow_id_derived info runtime.py Workflow ID derived from LLM or fallback; method=llm|fallback, slug=
plan_task_workflow_id_llm_failed warning runtime.py LLM slug derivation failed; exception=
plan_task_draft_written info scratchbook_adapter.py Draft written to disk; path=
plan_task_state_loaded debug scratchbook_adapter.py Runtime state read from disk; phase=
plan_task_event_appended debug scratchbook_adapter.py Event appended to events.jsonl; event_type=
plan_task_memory_appended debug scratchbook_adapter.py Memory entry appended; task_id=
plan_task_subagents_marked_stale info scratchbook_adapter.py In-flight subagents staled on restart; stale_count=, task_ids=
plan_task_task_queue_initialized info task_exec.py Task queue built and state updated; task_count=, current_task_id=, phase=
plan_task_subagent_dispatched info task_exec.py Subagent session recorded for a task; task_id=, session_id=
plan_task_task_completed info task_exec.py Task completed; next_task_id=, workflow_done=
plan_task_circuit_breaker_triggered warning task_exec.py Task retry budget exhausted; retry_count=, max_retries=
plan_task_dependency_cycle_detected warning task_exec.py Cyclic dependency found before raise; cycle_ids=
plan_task_reviews_not_approved warning task_exec.py Task start blocked by missing reviews; missing_phases=
plan_task_finalize_blocked warning controller.py Plan finalization blocked by missing verdicts; missing_phases=
plan_task_plan_artifact_missing warning controller.py Plan artifact file not found before raise; path=
plan_task_plan_not_finalized warning plan_schema.py Plan status is not finalized before raise; status=
plan_task_command_plan_start info main.py /plan:start succeeded; workflow_id=, description_len=
plan_task_command_plan_resume info main.py /plan:resume succeeded; workflow_id=, phase=
plan_task_command_plan_finalize info main.py /plan:finalize succeeded; workflow_id=
plan_task_system_prompt_switched info main.py System prompt replaced from PLAN_MAIN_AGENT to TASK_MAIN_AGENT; entity_id=, from_prompt=, to_prompt=
plan_task_command_task_start info main.py /task:start succeeded; task_count=, current_task_id=
plan_task_task_start_auto_loaded_state info main.py /task:start <workflow_id> auto-loaded persisted state; workflow_id=, phase=
plan_task_command_task_resume info main.py /task:resume succeeded; workflow_id=
plan_task_command_task_replan info main.py /task:replan succeeded; task_id=
plan_task_command_task_abort info main.py /task:abort succeeded; task_id=
plan_task_command_plan_status debug main.py /plan:status invoked
plan_task_command_task_status debug main.py /task:status invoked; phase=
plan_task_command_error warning main.py A slash command raised ValueError; command=, exception=
plan_task_llm_usage info billing.py Per-invocation token counts; prompt_tokens=, completion_tokens=, total_tokens=, cached_input_tokens=, cache_creation_tokens=, cache_read_tokens=
plan_task_llm_cache_stats info billing.py Per-invocation cache hit-rate; cache_read_tokens=, total_prompt_tokens=, cache_hit_rate=. Emitted when using ApiFormat.OPENAI_RESPONSES with DashScope (returns input_tokens_details.cached_tokens), OpenAI (returns prompt_tokens_details.cached_tokens), or Anthropic (returns cache_read_input_tokens). Not emitted with DashScope Chat Completions API, which does not expose cached token counts.
plan_task_session_billing_summary info billing.py Cumulative token totals at end of session; invocation_count=, total_prompt_tokens=, total_completion_tokens=, total_tokens=, total_cached_input_tokens=
accounting_invocation_recorded info ecs_agent.accounting Per-invocation cost + cache hit-rate from AccountingSubscriber; total_cost=, cache_hit_rate=
plan_task_auto_transition_write_plan info controller.py QA review approved — auto-transitioned from DRAFT_QA_REVIEW to WRITE_PLAN; workflow_id=
plan_task_auto_transition_plan_finalized info controller.py Plan QA review approved — auto-transitioned from PLAN_QA_REVIEW to PLAN_FINALIZED; workflow_id=
plan_task_auto_trigger_plan_writer info main.py QA approved — injected write-plan trigger message to start plan_writer subagent automatically; workflow_id=, source= (omitted when triggered by live QA event; source=reconcile_after_resume when triggered on /plan:resume)

Implementation Details

  • Testable World Factory: build_plan_task_world(model, base_dir=None, *, compaction_threshold_tokens=..., compaction_method=...) is a public function that returns (world, agent_id, adapter_ref, runtime_state), enabling direct world setup in tests without running the CLI. adapter_ref is a list[ArtifactAdapter | None] — starts as [None] and is populated in-place by the /plan:start handler after the workflow ID is derived.
  • Framework-Native Auto Compaction: build_plan_task_world(...) accepts compaction_threshold_tokens and compaction_method, installs CompactionConfigComponent, initializes ConversationArchiveComponent, and registers CompactionSystem() at priority -30. The example reuses the shared framework compaction pipeline rather than maintaining a bespoke plan-and-task summarizer.
  • workflow_id Auto-Derivation: /plan:start <description> calls derive_workflow_id_from_llm() to ask the LLM to generate a short, meaningful English slug from the description (e.g., "writing-assistant-multi-agent"). Falls back to slug_from_description() on provider error or invalid output. The derived ID controls the scratchbook directory for all subsequent operations in that session.
  • Progressive Draft Editing: The planning interview fills draft.md one section at a time using read_file (to get LINE#HASH annotated content) plus hash-anchored edit_file(op=..., pos=..., end=..., content=...) calls. The LLM reads the file first to capture N#HASH references, then replaces exactly the placeholder line or range. Full-file rewrites via write_file are explicitly prohibited by the system prompt.
  • Atomic Writes: All artifact updates use atomic file operations to prevent corruption.
  • Circuit Breaker: TaskExec implements a retry budget to prevent infinite loops on failing tasks.
  • Review Gating: Finalization is strictly blocked until DRAFT_ADVISOR_REVIEW, DRAFT_QA_REVIEW, and PLAN_QA_REVIEW all have approved verdicts.
  • Advisor Retry Loop: When the advisor returns revise or blocked, the system prompt instructs the planner LLM to apply the feedback to draft.md via edit_file and re-call the advisor. Only an approved advisor verdict unlocks the QA step. Non-approved verdicts for a phase replace any prior non-approved verdict (upsert semantics). Once a phase reaches approved, that verdict is sticky — any subsequent upsert attempt for the same phase is silently ignored. approved verdicts always have notes=None; non-approved verdicts retain their notes for debugging.
  • Slash Command Re-trigger Guard: After /task:start the command message stays as the last role="user" entry in the conversation (tool results use role="tool" and do not replace it). _handle_task_start therefore checks whether SystemPromptConfigSpec already contains TASK_MAIN_AGENT_SYSTEM_PROMPT; if so, it returns None immediately, letting the LLM continue task execution without re-initializing the task queue.
  • Compaction State Reset on Workflow Switch: _reset_compaction_state(...) removes CurrentCompactionSummaryComponent, clears ConversationArchiveComponent.archived_summaries, and removes RenderedSystemPromptComponent before /plan:start, /plan:resume, or auto-loading state from /task:start <workflow_id>. This keeps prompt summaries aligned with the active workflow instead of a previous session.
  • Plan Template: templates/workflow_plan_template.md is an annotated reference showing the exact workflow_plan.md format: YAML frontmatter, ## Overview with ### Dependency Graph, ## Tasks section, per-task ### Task: <task_id> + ```yaml ``` blocks, and an optional ## Appendix AC cross-reference table. The format spec is also embedded verbatim into WRITE_PLAN_SYSTEM_PROMPT and build_write_plan_prompt() as _WORKFLOW_PLAN_FORMAT so the plan_writer subagent has an unambiguous reference without reading a file.
  • Dependency Resolution: Tasks are executed in topological order based on their dependencies list.
  • Review Verdict Lifecycle: Each phase holds at most one verdict in review_verdicts. Upsert semantics: non-approved verdicts replace earlier non-approved verdicts for the same phase; approved is terminal and cannot be overwritten. approved verdicts always have notes=None. The plan_version field has been removed from ReviewVerdict.
  • Status Lifecycle: Whenever the state machine transitions to a new phase, status is set to "active". Terminal handlers (handle_task_abort → "aborted", handle_plan_finalize → "ready", handle_task_replan with scope change → "needs_review") override status after the transition.
  • Token Prefix Caching: This example uses ApiFormat.OPENAI_RESPONSES with enable_store=True (default config). The DashScope Responses API endpoint (/api/v2/apps/protocols/compatible-mode/v1) returns usage.input_tokens_details.cached_tokens, which normalize_openai_usage() maps to cached_input_tokens. On warm calls where the system prompt prefix is cached, cached_input_tokens > 0 and plan_task_llm_cache_stats will be emitted with a non-zero cache_hit_rate. The DashScope Chat Completions API (/compatible-mode/v1) does not return cached token counts and does not support the Responses protocol — switching back to it would make cache observability unavailable.