--- name: immo-find description: Discover new listings — scan portals, process pipeline inbox, batch discovery user_invocable: true args: mode argument-hint: "[scan | pipeline | batch | auto | notify]" --- # immo-find — Discover Listings ## Mode Routing | Input | Mode | |-------|------| | (empty / no args) | `discovery` — Show command menu | | `scan` | `scan` — Scan all configured portals | | `pipeline` | `pipeline` — Process pending URLs from inbox | | `batch` | `batch` — Batch process multiple listings (delegates to /immo-assess) | | `auto` | `auto` — Full automated cycle: scan → triage → notify | | `notify` | `notify` — Send email summary of current pipeline state | --- ## Discovery Mode (no arguments) ``` immo-find — Discover Listings /immo-find scan → Scan configured portals for new listings /immo-find pipeline → Process pending URLs from data/pipeline.md /immo-find batch → Batch evaluate multiple pending listings /immo-find auto → Full cycle: scan → triage → notify (for /loop) /immo-find notify → Send email summary of current pipeline Automated: /loop 1h /immo-find auto Pipeline: data/pipeline.md (add URLs manually or via scan) Script: npm run scan (headless Playwright for non-CiC portals) ``` --- ## Context Loading ### Scan mode: Read `modes/scan.md`. Load `portals.yml` + `config/profile.yml` + `data/scan-history.tsv`. **Two scan methods** (determined by `scan_method` in portals.yml): #### 1. Playwright portals (`scan_method: playwright`) Run headlessly via script: `node scripts/scan.mjs` Can run in background, unattended. Handles Immowelt, Kleinanzeigen, Vonovia, etc. #### 2. CiC portals (`scan_method: cic`) Scanned interactively via Claude-in-Chrome using the user's real browser. Required for portals with aggressive bot detection (ImmoScout24). **CiC tab management:** ALWAYS create a dedicated tab for CiC scanning via `tabs_create_mcp`. Never reuse existing tabs. When done, close only the tab(s) you created via `tabs_close_mcp` — never close tabs you didn't create. **CiC scan workflow:** 1. Read `portals.yml` for all portals with `scan_method: cic` and `enabled: true` 2. **Create a new CiC tab** via `mcp__claude-in-chrome__tabs_create_mcp` — note the tabId 3. For each CiC portal: a. Navigate to `search_url` via `mcp__claude-in-chrome__navigate` (URL should include `&sorting=2` for newest first) b. If CAPTCHA appears ("Ich bin kein Roboter"): wait 5-10 seconds, then re-check — most CAPTCHAs auto-solve. Only ask user if still blocked after waiting. c. **Pagination loop** (up to 100 listings total): - Read the extraction snippet from `scripts/portals/{portal}-cic.js` - Run the snippet via `mcp__claude-in-chrome__javascript_tool` — returns `{count, total, hasNextPage, listings}` - Pipe the listings JSON to `node scripts/process-scan.mjs` - Check process-scan output: if most listings are duplicates (≥80% already seen), stop paginating - If `hasNextPage` is true and under 100 total: click "Nächste Seite" via `mcp__claude-in-chrome__find` or navigate to `search_url&pagenumber={N}` - Repeat 4. **Close the tab** you created via `mcp__claude-in-chrome__tabs_close_mcp` (only your tab) 5. Show scan summary **Available CiC extraction snippets:** - `scripts/portals/immoscout24-cic.js` — ImmoScout24 (`.listing-card` containers, returns pagination info) **Combined scan order:** 1. First: run `node scripts/scan.mjs` for Playwright portals (can be backgrounded) 2. Then: scan CiC portals interactively 3. Show combined summary ### Pipeline mode: Read `modes/_shared.md` + `modes/pipeline.md`. If 3+ pending URLs: delegate to subagent. ### Batch mode: Read `modes/_shared.md` + `modes/batch.md`. Delegates individual evaluations to /immo-assess. ### Auto mode (for `/loop` usage): Full automated cycle designed for `loop 1h /immo-find auto`. The purpose is to deliver scored, actionable recommendations — not raw links. Every step must complete before notifying. **Step 1 — Playwright scan:** ``` node scripts/scan.mjs ``` Capture stdout. Note how many new listings were added. **Step 2 — CiC scan (ImmoScout24 and other `scan_method: cic` portals):** 1. Read `portals.yml` for CiC portals 2. If any CiC portals enabled: a. Create a new CiC tab (note the tabId) b. For each CiC portal: navigate → extract via JS snippet → pipe to `process-scan.mjs` (paginate all pages, up to 100) c. Close your tab 3. Note how many new listings were added **Step 3 — Pipeline triage (metadata-only discard):** Read `data/pipeline.md`. For each pending `- [ ]` entry, check metadata against profile criteria: - Rooms < min_rooms → mark as `DISCARDED` - m² < min_m2 → mark as `DISCARDED` - Price > max_kaltmiete × 1.2 → mark as `DISCARDED` - Title contains Tauschwohnung keywords → mark as `DISCARDED` - Location is wrong city → mark as `DISCARDED` - Price suspiciously low (likely extraction error) → mark as `DISCARDED` Update `data/pipeline.md` in place. **Step 4 — Pipeline evaluation (ALL pending listings):** First, check which pending URLs already have reports (cross-reference `data/pipeline.md` URLs against `reports/*.md`). Skip those — mark as processed with the existing score. For EVERY remaining pending URL that has NO report yet, launch a dedicated Agent: ``` Agent( subagent_type="general-purpose", prompt="You are evaluating a real estate listing for immo-ops. LISTING URL: {url} Read these files for context: - config/profile.yml — search criteria (max Kaltmiete 1500, min 60m², min 3 rooms, must-haves: balkon_or_terrasse + keller) - modes/_shared.md — scoring system (8 blocks A-H, weighted average) - modes/evaluate.md — full evaluation workflow and report format STEPS: 1. Create a CiC tab via mcp__claude-in-chrome__tabs_create_mcp 2. Navigate to the listing URL 3. If CAPTCHA: report it and stop. If 'Angebot nicht gefunden': mark as EXPIRED and stop. 4. Extract ALL details via javascript_tool (use the extraction snippet pattern from the codebase) 5. Detect Tauschwohnung → if yes, mark as DISCARDED and stop 6. Score all 8 blocks (A-H) with numeric scores and one-line justification per block 7. Write report to reports/{NNN}-{location-slug}-{rooms}r-{date}.md 8. Write tracker TSV to batch/tracker-additions/{NNN}-{slug}.tsv 9. Close your CiC tab (only the tab you created) 10. Report back: {URL} | {score}/5 | {one-line summary} Next listing number: {NNN} ", description="immo-assess {expose_id}" ) ``` Launch agents **one at a time** (CiC tabs can't run in parallel — each agent needs exclusive browser access). Wait for each agent to complete before launching the next. After all agents complete: `node scripts/merge-tracker.mjs` **Step 5 — Verification (MUST pass before notify):** Before sending any notification, verify: - [ ] All enabled portals have been scanned (no portal skipped without reason) - [ ] Pipeline has 0 pending `- [ ]` entries (all evaluated or discarded) - [ ] `node scripts/verify-pipeline.mjs` passes If verification fails, DO NOT notify. Instead report the failure and stop. **Step 6 — Notify:** Only after verification passes: 1. **Push notification** via `PushNotification` — short summary (under 200 chars): `immo-ops: {N} new — {counts per target} (top: #{id} {score}/5)` 2. **Email draft** via Gmail MCP — HTML with sections per search target, tables with scored listings, pro/con, color-coded. If no new listings were found in this cycle, skip both silently. See **Notify mode** below for email format. ### Notify mode: Two channels: push notification (instant, short) + email draft (detailed, for review). **Config:** Read `searcher.notification_email` and all `searches[]` entries from `config/profile.yml`. **When to send:** - In auto mode: only if new listings were found in this scan cycle - Standalone `/immo-find notify`: always send current state **1. Push notification** (always, when there are results): ``` PushNotification({ message: "immo-ops: {N} new — {counts per target, e.g. '3 Miete, 2 Haus, 1 Grundstück'} (top: #{id} {score}/5)", status: "proactive" }) ``` Under 200 chars. Lead with total count, break down by target, mention top pick. **2. Email draft** (detailed): **What to include:** Scored listings from `data/pipeline.md` with score ≥ 3.0 (max 50). Exclude DISCARDED, DUPE, and sub-3.0 entries — these are not actionable and waste the reader's attention. Sort by score descending within each section. Mention the count of excluded sub-3.0 listings in the footer. **Enrichment from reports:** For each listing with a report in `reports/`, read the `## Summary` section and extract: - The bold assessment phrase (e.g., "Strong candidate, worth pursuing") - Key pro: first positive point from the summary - Key con: first concern after "However:" or "Main concerns:" or "Key concern:" Include these as a second row under each listing in the table (smaller font, gray text). **Email structure — grouped by search target:** Header block (before the sections): - `

immo-ops scan results

` - Subtitle `

`: `{timestamp} • {N} new listings evaluated • {M} scoring 3.0+` - Scan note `

`: which portals were scanned (Playwright vs browser), which search groups are disabled, anything skipped (CAPTCHA etc.) The email body is organized into sections, one per search target from `config/profile.yml`. Each section has a header and its own table. **Section header:** `

` with emoji, search name, and a gray count badge: ```html

🏢 Berlin flat rental (Miete / Wohnung · {N} listings)

``` - `🏠` house purchase (Kauf/Haus), `🏢` flat rental (Miete/Wohnung), `🌳` plot purchase (Kauf/Grundstück) - Active section with zero scored listings: show header with "(no listings yet)", skip the table - **Disabled search groups: still show the header**, grayed out — border `#bbb`, `color:#888`, badge text `(disabled)` — so the reader sees the full search scope at a glance **Matching listings to sections:** Use the search group tag from the pipeline entry (added during scan). For older pipeline entries without a tag, infer from listing type (miete/kauf) and property type (wohnung/haus/grundstück) based on the report or URL. **Table format per section:** HTML with `htmlBody` parameter. Columns: #, Score, Listing (linked to portal URL), Price, Size, Rooms. The # column shows the listing number from the tracker. Score is `{score}` without "/5" (e.g. `4.6`). Header row: `#f0f0f0` background, all cells `border:1px solid #ddd`. **Listing cell:** linked short descriptive title, portal name inline after the link in gray: ```html {short title — area} ({Portal}) ``` Color-code rows: - Green background (`#e8f5e9`): score 3.5+ (worth pursuing) - Yellow background (`#fff8e1`): score 3.0–3.4 (compromises) - Red background (`#ffebee`): score below 3.0 (not recommended) - White: no score yet **CRITICAL — Gmail strips row-level styles in the draft composer.** Set the background on EVERY cell, not just the row, and use both the legacy attribute and inline style. The `bgcolor` attribute survives every email sanitizer: ```html ... ... ``` All styles inline on the elements — do not rely on a `