Outcome Reflection

Outcome Reflection

Overview

Closes the feedback loop between prediction and reality in agent task execution. After a task completes, this skill compares the predicted outcome (recorded at planning time) against the actual outcome (observed at completion), scores the accuracy on three dimensions, and persists the calibration record to memory for future use.

Over time, accumulated calibration data reveals systematic biases (e.g., consistent underestimation of implementation tasks) that planners can use to improve future predictions.

When to Use

Invoke immediately after any task that had a predicted outcome recorded:

• After HIGH/EPIC pipeline completion (planner predictions vs actual)
• After estimation tasks (token counts, time estimates, file counts)
• After architectural decision tasks (expected impact vs observed impact)
• After any task where the planner recorded explicit predictions in task metadata

Do not invoke for:

• Tasks where no prediction was recorded (nothing to compare)
• Trivial tasks (single-file edits) — overhead not justified
• Ongoing tasks — invoke only after TaskUpdate(completed)

Iron Law

NO CALIBRATION WITHOUT A PRIOR PREDICTION

If no prediction was recorded at planning time, outcome-reflection cannot score accuracy. The fix is to ensure planners record predictions. See plan-generator for prediction metadata format.

Calibration Dimensions

Dimension 1: Estimation Accuracy

Measures how accurately the agent estimated measurable quantities:

• Token count predictions vs actual
• File count predictions vs actual
• Step count predictions vs actual
• Time/effort predictions vs actual

Score: 0.0–1.0 (1.0 = exact, 0.0 = off by >10x)

score = max(0, 1 - abs(predicted - actual) / max(predicted, actual))

Dimension 2: Prediction Quality

Measures how accurately the agent predicted qualitative outcomes:

• Did the implementation meet stated requirements?
• Did the plan identify the actual blockers?
• Were edge cases predicted that actually occurred?

Score: 0.0–1.0 (1.0 = predicted outcome exactly matched, 0.0 = outcome unrecognized)

Scored by reading task completion metadata and comparing against task creation metadata.

Dimension 3: Decision Quality

Measures whether the decisions made during the task were appropriate in retrospect:

• Did the chosen approach work without significant pivots?
• Were there rework loops that a better decision would have avoided?
• Did the agent's confidence match the actual difficulty?

Score: 0.0–1.0 (1.0 = no rework, smooth execution; 0.0 = multiple pivots or failure)

Calibration Record Format

Each outcome-reflection run produces one calibration record:

{
  "taskId": "task-N",
  "taskType": "implementation|planning|estimation|architecture|security",
  "completedAt": "ISO-8601",
  "agentType": "developer|planner|architect|...",
  "predictions": {
    "estimatedTokens": 5000,
    "estimatedFiles": 3,
    "estimatedSteps": 5,
    "predictedOutcome": "Add JWT auth with refresh tokens",
    "predictedBlockers": ["Redis not available"],
    "confidence": "Medium"
  },
  "actuals": {
    "actualTokens": 7200,
    "actualFiles": 5,
    "actualSteps": 8,
    "actualOutcome": "Added JWT auth with refresh tokens; added Redis fallback",
    "actualBlockers": ["Redis not available", "JWT library version mismatch"],
    "reworkLoops": 1
  },
  "scores": {
    "estimationAccuracy": 0.72,
    "predictionQuality": 0.85,
    "decisionQuality": 0.8,
    "overall": 0.79
  },
  "flags": [],
  "notes": "Token estimate was 44% low. Consider 1.5x buffer for JWT auth tasks."
}

Workflow

Step 0: Load Memory and Task Data

# Check calibration history for this agent type
grep -r "outcome-reflection" C:/dev/projects/agent-studio/.claude/context/memory/learnings.md 2>/dev/null | tail -10

Read the completed task's metadata using TaskGet:

const taskData = TaskGet({ taskId: 'task-N' });
// Access: taskData.metadata.predictions, taskData.metadata.actuals

Step 1: Validate Input

node .claude/skills/outcome-reflection/hooks/pre-execute.cjs \
  '{"taskId":"task-N","predictions":{},"actuals":{}}'

Expected output: { "valid": true } or error listing missing fields.

Step 2: Compute Estimation Accuracy Score

For each measurable quantity where a prediction exists:

estimationScore = max(0, 1 - abs(predicted - actual) / max(predicted, actual))

Command to compute:

node .claude/skills/outcome-reflection/scripts/main.cjs \
  --taskId task-N \
  --predicted '{"tokens":5000,"files":3,"steps":5}' \
  --actual '{"tokens":7200,"files":5,"steps":8}'

Expected output: JSON with per-dimension scores and overall score.

Step 3: Score Prediction Quality (Qualitative)

Read task creation metadata (what was predicted qualitatively) and task completion metadata (what actually happened). Score on 0.0–1.0:

Step 4: Score Decision Quality

Check completion metadata for rework signals:

• reworkLoops: 0 → Decision quality: 1.0
• reworkLoops: 1 → Decision quality: 0.75
• reworkLoops: 2 → Decision quality: 0.5
• reworkLoops: 3+ → Decision quality: 0.25
• Task status: "failed" → Decision quality: 0.0

Step 5: Persist Calibration Record

Append to .claude/context/memory/learnings.md:

## [DATE] Calibration: task-N (AGENT_TYPE)
- Estimation: SCORE (note what was under/over-estimated)
- Prediction: SCORE (note what was missed)
- Decision: SCORE (note rework loops)
- Overall: SCORE
- Action: [flag for reflection | no action needed]

Also emit structured record via MemoryRecord for semantic search:

MemoryRecord({
  type: 'pattern',
  text: `Calibration for ${taskType} tasks: estimation=${score}, prediction=${score}, decision=${score}`,
  area: 'calibration',
});

Step 6: Flag High-Miss Tasks

If overall < 0.6, append a reflection request:

echo '{"id":"'$(node -e "console.log(require('crypto').randomUUID())")'","trigger":"calibration-miss","priority":"low","context":"Task task-N had calibration score SCORE. Identify root cause of estimation/prediction miss."}' >> .claude/context/runtime/reflection-spawn-request.json

Step 7: Emit Observability Event

const { sendEvent } = require('.claude/tools/observability/send-event.cjs');
sendEvent({
  tool_name: 'outcome-reflection',
  agent_id: process.env.AGENT_ID || 'unknown',
  session_id: process.env.SESSION_ID || 'unknown',
  outcome: 'success',
  metadata: { taskId, overallScore, flagged: overallScore < 0.6 },
});

Calibration Trend Analysis

After 5+ calibrations for the same task type and agent, the planner should query the trend:

node .claude/skills/outcome-reflection/scripts/main.cjs \
  --analyze --agentType developer --taskType implementation --last 10

Expected output: Mean, median, trend direction (improving/degrading/stable), and top 3 miss patterns.

Use trend data to adjust future estimates:

• Systematic underestimation → apply correction factor
• Systematic overestimation → reduce buffers
• High variance → flag for human review before HIGH tasks

Enforcement Hooks

Input validated against schemas/input.schema.json before execution. Output contract defined in schemas/output.schema.json.

Pre-execution hook: hooks/pre-execute.cjs — validates taskId, predictions object, actuals object. Post-execution hook: hooks/post-execute.cjs — emits observability event via send-event.cjs.

Anti-Patterns

• Never invoke without a prior prediction — meaningless calibration score
• Never aggregate scores before trending — premature aggregation hides per-dimension biases
• Never skip low-score flagging — misses compound if root cause is not investigated
• Never use calibration to blame agents — use it to improve planning accuracy
• Never run on trivial tasks — calibration overhead is not justified for single-file edits

Related Skills

• plan-generator — records predictions at planning time (prerequisite for calibration)
• instinct-learning — records atomic learned patterns (complementary to calibration)
• reflection-agent — investigates high-miss tasks flagged by outcome-reflection
• verification-before-completion — gate that runs before completion (runs before this skill)

Memory Protocol

Before starting: Read .claude/context/memory/learnings.md to find prior calibrations for the same agent type and task type. This provides baseline context.

After completing: Append calibration record to .claude/context/memory/learnings.md. Use MemoryRecord tool for structured pattern recording. Do NOT write directly to patterns.json.

Assume interruption: If calibration context is lost, re-read the completed task's metadata from TaskGet — all predictions and actuals should be in task metadata.