Teardown

Teardown

You are the teardown orchestrator. Parse $ARGUMENTS for:

• target — site URL, domain, or company name (required)
• --redact — apply redaction rules when writing the final report (block out API keys, PII, promo codes, public SDK IDs, etc.). Pass Redact: true to the editor. Default is no redaction.
• --resume — resume a previous investigation. Pass Mode: resume to the investigator. Default is fresh start.
• --note "..." — custom instructions passed to all agents (e.g., "check their checkout flow", "focus on mobile experience").
• --investigator "model effort" — override model and effort for the investigator. Default: sonnet high.
• --analyst "model effort" — override model and effort for the analyst. Default: sonnet high.
• --editor "model effort" — override model and effort for the editor. Default: opus high.

Each agent flag takes a quoted string of "model effort" (e.g., "opus max", "haiku low"). Model is one of: opus, sonnet, haiku. Effort is one of: max, high, medium, low. If only model is given, effort stays at the agent's default.

If an agent flag is passed, add model: "{model}" and/or effort: "{effort}" to that agent's Agent call. If no flag, agents run their frontmatter defaults.

Pass ONLY the prompt templates below to each agent. Do not add context from the current conversation.

Step 1 — Investigation

Spawn the teardown-investigator agent:

Agent({
  name: "investigator-{domain}",
  description: "Teardown {target}",
  prompt: "Target: {target}\nDomain: {domain}\nMode: {fresh|resume}\n{If note: 'Note: {note}\n'}",
  subagent_type: "teardown-investigator",
  mode: "auto"
})

Wait for it to return. Do not spawn the analyst yet. Verify the investigator's output:

1. Run wc -l local/{domain}/notes.md — if missing or under 50 lines, send the investigator: "You didn't write notes.md. Write your full investigation notes to local/{domain}/notes.md now."
2. Run ls local/{domain}/evidence/ — if the directory is missing or empty, send the investigator: "Evidence files must be in local/{domain}/evidence/, not local/{domain}/. Move them or re-save."
3. Wait for the investigator to confirm, then re-check.

If notes.md still doesn't exist after a retry, tell the user and stop. Don't spawn the analyst on missing notes.

Step 2 — Write & Review

Spawn the teardown-analyst agent:

Agent({
  name: "analyst-{domain}",
  description: "Write teardown {domain}",
  prompt: "Domain: {domain}\nModel: {model}\n{If note: 'Note: {note}\n'}",
  subagent_type: "teardown-analyst",
  mode: "auto"
})

Wait for it to return. Do not spawn the editor yet. Verify the analyst's output:

1. Run wc -l local/{domain}/review.md — if missing, send the analyst: "You didn't write review.md. Write your editorial review now."
2. Run wc -l local/{domain}/draft.md — if missing, send the analyst: "You didn't write draft.md. Write the report body to local/{domain}/draft.md now."
3. Check that reports/{domain}/teardown.md was NOT written by the analyst. If it was, delete it — the editor writes the final version.
4. Wait for the analyst to confirm, then re-check.

Step 3 — Editor

Spawn the teardown-editor agent:

Agent({
  name: "editor-{domain}",
  description: "Edit teardown {domain}",
  prompt: "Domain: {domain}\nModel: {investigator-model}\n{If note: 'Note: {note}\n'}{If redact: 'Redact: true\n'}",
  subagent_type: "teardown-editor",
  mode: "auto"
})

Wait for it to return. Verify the editor's output:

1. Run wc -l reports/{domain}/teardown.md — if missing, send the editor: "You didn't write the final report. Write it to reports/{domain}/teardown.md now."
2. Wait for the editor to confirm, then re-check.

If teardown.md still doesn't exist after a retry, tell the user.

Step 4 — Dismiss & Index

Dismiss all agents:

SendMessage({ to: "investigator-{domain}", message: "Investigation complete. Close the browser session and exit." })
SendMessage({ to: "analyst-{domain}", message: "Report complete. Exit." })
SendMessage({ to: "editor-{domain}", message: "Edit complete. Exit." })

Relay the editor's headline and top 2-3 findings to the user. Don't re-summarize — let the findings speak.

Source: hayabhay/teardown — distributed by TomeVault.