premortem

/premortem -- Pre-flight Risk Evaluation

Standalone skill that scans queued (or specified) analysis designs, runs a deterministic risk decision tree (HARD_BLOCK / HIGH / MEDIUM / LOW / SKIP), and issues an approval token consumed by /batch-analysis --approved-by TOKEN.

When to Use

• Before launching /batch-analysis to validate the queue
• When you want to check risk levels of designs without executing them
• As an automated gate in review / auto mode dispatched by the launcher

When NOT to Use

• Creating or editing designs (-> /analysis-design)
• Executing batch analysis (-> /batch-analysis)
• Reviewing completed results (-> /analysis-reflection)

Workflow

1. Parse arguments -- Claude Code invokes skills/premortem/cli.py via Python subprocess. The CLI expects design data as JSON on stdin (Claude Code reads MCP tools and pipes the result).

2. Collect design data (Claude Code responsibility, before invoking cli.py):

   • Call list_analysis_designs() and filter next_action.type == "batch_execute" (or use --design <id> / --all to select differently)
   • For each design: get_analysis_design(id), get_table_schema(source_id), search_catalog(source_id) to build source_checks_map
   • Pipe the JSON payload to cli.py stdin

3. Risk evaluation (cli.py, pure decision engine):

   • For each design: history_query.query() + risk_evaluator.evaluate()
   • Render 1-line-per-design table to stdout
   • Apply mode logic (manual/review/auto) to decide approved vs skipped

4. Interactive gate (manual mode or review+HIGH):

   • Display [s]kip / [e]dit / [a]bort / [c]ontinue per HIGH design
   • HARD_BLOCK: [c]ontinue is NOT offered

5. Token issuance:

   • token_manager.issue() writes .insight/premortem/{TIMESTAMP}.yaml
   • stdout final line: Launch with: /batch-analysis --approved-by {token_id}

Risk Levels

Every design is classified into one of five levels. The level determines what the operator (or automation) should do next.

Level	Trigger	Operator Action
HARD_BLOCK	Unregistered source, allowlist violation, or BigQuery location mismatch.	Batch MUST NOT run this design. The [c]ontinue option is withheld; use [s]kip, [e]dit (fix the design), or [a]bort (stop the whole batch).
HIGH	History median × buffer exceeds time_high_min or estimated_rows exceeds static_rows_high or success rate over history_min_samples samples drops below success_rate_high_threshold.	manual mode: operator prompt (s/e/a/c). review mode: batch STOPS (exit 2). auto mode: batch RUNS with a WARNING: HIGH risk executed without human approval line appended to summary.md.
MEDIUM	Extrapolated time between time_medium_min and time_high_min.	Proceeds automatically in every mode. Logged for visibility.
LOW	Nothing above triggered, history is healthy.	Proceeds automatically. The happy path.
SKIP	Terminal status (supported / rejected / inconclusive) or next_action cleared.	Recorded as skipped with skip_reason=terminal_status; never enters the batch queue.

Thresholds live in .insight/config.yaml under premortem.* (see .insight/config.example.yaml). Change them there, not in code.

Exit Codes

Code	Meaning
0	Token issued successfully, batch may proceed
2	review mode + HIGH detected, batch should stop
1	Unexpected error (config invalid, I/O failure, etc.)

Writing Contract (AC-1.5)

During /premortem execution, the following paths are NEVER written to:

• notebook.py / marimo session JSON
• .insight/designs/*.yaml / .insight/designs/*_journal.yaml
• .insight/runs/*/*/manifest.yaml / .insight/runs/*/run.yaml
• .insight/catalog/**

Only .insight/premortem/ receives writes (approval token).

Language Rules

• Respond to users in Japanese
• Code, IDs, tool names, and YAML fields stay in English