| name | extended-react |
| description | Six-phase ReAct loop (pre-check, thinking, self-critique, action, tool, post). Activates per-workflow via the thinking_level knob in config/workflows.yaml. Replaces standard ReAct for workflows where independent verification of the reasoning improves outcomes. |
| when_to_use | ["workflow has thinking_level >= med in config/workflows.yaml",{"workflow handles a feature in feature_list.json with `category":"agent | infrastructure | hook`"},"prior iterations on this feature failed; opt up the thinking_level for the retry"] |
| brd_ref | §3.6 |
Extended ReAct loop
Standard ReAct: reason → act → observe. Extended ReAct (BRD §3.6) inserts two phases:
1. pre-check + compaction drain message queue, compact if >70% budget
2. thinking (optional) chain-of-thought trace at configured depth
3. self-critique (optional) independent pass on the thinking output, Opus-grade
4. action LLM call with full tool schema
5. tool execution registry dispatch through five safety layers
6. post-processing decide iterate vs return
When the phases activate
thinking_level | Phase 2 thinking | Phase 3 self-critique |
|---|
off | skipped | skipped |
min | shallow trace | skipped |
low | shallow trace | skipped |
med | full CoT | skipped |
high | full CoT | full Opus-grade critique |
Configured per-workflow in config/workflows.yaml. The default is low.
Why this matters
- Phase 2 (thinking) externalizes the reasoning so it's auditable in telemetry.
- Phase 3 (self-critique) catches a category of error where the action pass commits early without consulting the BRD or the existing code — Opus-grade critique on the thinking output identifies the gap before tool calls fire.
- Phase 1 (pre-check) drains any queued message from the user (Pi pattern, BRD §11 alignment) and compacts if budget is tight — preventing context exhaustion mid-action.
Anti-patterns
- Setting
thinking_level: high on every workflow. Self-critique is expensive (Opus). Reserve for spec-auditor, planner, critic — workflows where independence is the load-bearing primitive.
- Skipping phase 1 to "save tokens". A 70%+ context window with stale tool outputs hurts more than it helps.
- Treating phase 3 verdict as final on the feature. Self-critique is for the reasoning. Feature acceptance is the Critic's job (BRD §5.1) and the E2E gate's job (BRD §3.8).
Source
Read-and-reimplement of the OPENDEV §2.2.6 pattern (Python reference at opendev-to/opendev-py/opendev/runtime/react_loop.py, MIT). Our Durable Functions topology required a different control flow. See BRD §10.2 §3.6.
