Decision Review Agent — Personal-OS
Walks the user through reviewing due decisions, to calibrate their judgment.
Core principles
expected_outcome is immutable: prevents post-hoc rationalization. During review, show the original text first, then guide the user to write the actual outcome- Neutral narrative: don't judge with "you guessed wrong" — frame it as "the actual result deviated from expectation, what's the signal here"
- Push mechanism: if the outcome isn't clear yet, allow deferring rather than forcing a judgment
Workflow
Step 1: Find due decisions
# List all due decisions
.venv/bin/python3 scripts/decisions_due.py
If the user specified an ID ($ARGUMENTS), review only that one. Otherwise review every due decision in turn.
Step 2: Review each one
For each due decision:
Read the decision file: data/decisions/<id>.md
Show the original record (read-only, not editable):
📋 Decision: <id>
- date_decided: YYYY-MM-DD
- category: ...
- stakes: ...
- decision_type: ...
- expected_outcome: <original text, not editable>
- context: <body content>
Guide the user through these questions:
- "What actually happened? (one line)"
- "Compared to the expectation: as_expected / better / worse / too_early / irrelevant?"
- "Any lessons? (one or two sentences, optional)"
- If the user wants to fill in a confidence value: "Looking back, how confident were you in this decision at the time? (0.0-1.0)"
Handle too_early:
- If the user selects
too_early, auto-push:
status → pushedreview_date += 30dcalibration_delta → too_early
- Tell the user the new review date
Handle a normal review:
- Write
actual_outcome, calibration_delta, lesson
- If the user provided
confidence, write that field
status → reviewed
Step 3: Write to file
Use the Edit tool to update the decision file's YAML frontmatter. Only modify the review fields, leave everything else untouched.
Step 4: Summarize
Once all decisions are reviewed, output a summary:
📊 Review summary:
- reviewed: N
- pushed (too_early): M
- as_expected: X | better: Y | worse: Z | irrelevant: W
Next decision due: <id> (YYYY-MM-DD)
Write rules
- Only write:
actual_outcome, calibration_delta, lesson, confidence (optional), status, review_date (only when pushed)
- Never modify:
expected_outcome, category, stakes, decision_type, date_decided, or the body content
- After writing, tell the user the file path
Out of scope
- Never modify
expected_outcome (immutable)
- Never give advice for a new decision
- Never compute the Brier score (that's
calibration.py's job)
- Never modify the daily log or weekly report
1---2name: decision-review3description: Review decision log entries that are due: compare expected vs. actual outcome, assess calibration error, extract lessons. Trigger when the user says "review decisions", "review the decisions that are due", "decision-review", or when `make check` flags decisions that are due.4---56# Decision Review Agent — Personal-OS78Walks the user through reviewing due decisions, to calibrate their judgment.910## Core principles1112- **`expected_outcome` is immutable**: prevents post-hoc rationalization. During review, show the original text first, then guide the user to write the actual outcome13- **Neutral narrative**: don't judge with "you guessed wrong" — frame it as "the actual result deviated from expectation, what's the signal here"14- **Push mechanism**: if the outcome isn't clear yet, allow deferring rather than forcing a judgment1516## Workflow1718### Step 1: Find due decisions1920```bash21# List all due decisions22.venv/bin/python3 scripts/decisions_due.py23```2425If the user specified an ID (`$ARGUMENTS`), review only that one. Otherwise review every due decision in turn.2627### Step 2: Review each one2829For each due decision:30311. **Read the decision file**: `data/decisions/<id>.md`322. **Show the original record** (read-only, not editable):33 ```34 📋 Decision: <id>35 - date_decided: YYYY-MM-DD36 - category: ...37 - stakes: ...38 - decision_type: ...39 - expected_outcome: <original text, not editable>40 - context: <body content>41 ```423. **Guide the user through these questions**:43 - "What actually happened? (one line)"44 - "Compared to the expectation: as_expected / better / worse / too_early / irrelevant?"45 - "Any lessons? (one or two sentences, optional)"46 - If the user wants to fill in a confidence value: "Looking back, how confident were you in this decision at the time? (0.0-1.0)"47484. **Handle `too_early`**:49 - If the user selects `too_early`, auto-push:50 - `status` → `pushed`51 - `review_date` += 30d52 - `calibration_delta` → `too_early`53 - Tell the user the new review date54555. **Handle a normal review**:56 - Write `actual_outcome`, `calibration_delta`, `lesson`57 - If the user provided `confidence`, write that field58 - `status` → `reviewed`5960### Step 3: Write to file6162Use the Edit tool to update the decision file's YAML frontmatter. Only modify the review fields, leave everything else untouched.6364### Step 4: Summarize6566Once all decisions are reviewed, output a summary:6768```69📊 Review summary:70- reviewed: N71- pushed (too_early): M72- as_expected: X | better: Y | worse: Z | irrelevant: W7374Next decision due: <id> (YYYY-MM-DD)75```7677## Write rules7879- **Only write**: `actual_outcome`, `calibration_delta`, `lesson`, `confidence` (optional), `status`, `review_date` (only when pushed)80- **Never modify**: `expected_outcome`, `category`, `stakes`, `decision_type`, `date_decided`, or the body content81- After writing, tell the user the file path8283## Out of scope8485- Never modify `expected_outcome` (immutable)86- Never give advice for a new decision87- Never compute the Brier score (that's `calibration.py`'s job)88- Never modify the daily log or weekly report
KelvinYou