name: reflexion description: "Invoked by develop when iteration feedback requires a retry, not directly by users. Prevents repeating the same mistakes across attempts. Also relevant when: 'why did this fail again', 'same error twice', 'what should I do differently', 'keep making the same mistake'."
Reflexion
Reasoning Schema
Invariant Principles
- Every Failure Teaches: ITERATE verdicts contain actionable information.
- Patterns Over Instances: Single failures are learning; repeated failures are patterns.
- Root Cause Focus: Symptoms are feedback; causes are lessons.
- Knowledge Accumulates: Reflections persist across iterations and features.
- Guidance Prevents Repetition: Next attempt must address previous failure.
Inputs / Outputs
| Input | Required | Description |
|---|---|---|
feature_name |
Yes | Feature that received ITERATE verdict |
feedback |
Yes | List of feedback items from roundtable |
stage |
Yes | Stage where iteration occurred |
iteration_number |
Yes | Current iteration count |
| Output | Type | Description |
|---|---|---|
reflection_record |
Database | Stored in forged.db reflections table |
root_cause_analysis |
Inline | What went wrong and why |
retry_guidance |
Inline | Specific guidance for next attempt |
Phase Sequence
Steps 1-3: Full Analysis Pipeline
Dispatch subagent with command: reflexion-analyze
The subagent executes the complete analysis pipeline:
- Parse Feedback - Extract structured fields from each feedback item
- Categorize Root Cause - Map failures to root cause categories (Incomplete Analysis, Misunderstanding, Technical Gap, Scope Creep, Quality Shortcut, Integration Blind Spot)
- Root Cause Questions - Answer expected vs actual, why deviation occurred, what would have prevented it
- Store Reflections - Write to forged.db with PENDING status
- Generate Retry Guidance - Produce specific correction guidance for the re-invoked skill
Pattern Detection
| Pattern | Threshold | Alert |
|---|---|---|
| Same failure, same feature | 2 iterations | "Root cause not addressed" |
| Same failure, different features | 3 features | "Systemic pattern" |
| Same validator, different failures | 3 failures | "Validator focus area needs attention" |
Integration with Develop Workflow
Trigger: forge_iteration_return with ITERATE verdict
Flow: Roundtable ITERATE -> forge_iteration_return -> reflexion skill -> analyze + store + check patterns + generate guidance -> return to develop orchestrator -> re-select and re-invoke skill
Escalation
After 3 iterations on same stage with same root cause: mark ESCALATED, report attempts made, recommend human intervention.
Example
- Parse: source=Hermit, severity=blocking, stage=IMPLEMENT
- Categorize: Quality Shortcut (missing validation)
- Root cause: Rushed implementation, skipped security checklist
- Store reflection with status=PENDING
- Pattern check: Hermit flagged validation 2x before -> alert
- Generate guidance: "Add input validation to all endpoints before resubmit"
Self-Check
- All feedback items analyzed for root cause
- Root causes categorized (not just described)
- Reflections stored with PENDING status
- Pattern check performed
- Retry guidance includes specific corrections
- Escalation evaluated if iteration >= 3
If ANY unchecked: complete before returning.