@agent-pattern-labs/leads-rig 0.1.5 → 0.1.7

package/.claude/agents/general-free.md

@@ -27,7 +27,9 @@ Skip this only if the task explicitly says to attach to an existing session.
 ## Do
 
 - Crawl only public, high-signal company pages.
-- Extract emails, role inboxes, contact forms, public names/titles, source URL, evidence, and warnings.
+- Use your judgment on each email candidate: only emit an email lead when public evidence identifies one specific human owner of the address.
+- Extract high-quality named human emails, contact forms, public names/titles, source URL, evidence, and warnings.
+- If an address is shared, role-based, departmental, generic, blocked, or not tied to a visible named human, do not emit it as an email lead; emit a `contact_path` if useful.
 - Emit JSON/JSONL matching `templates/lead-schema.json`.
 - Run `npx public-leads validate --input <file>` before returning a result.
 - Return file paths and terminal status, not a prose-only summary.

package/.cursor/rules/agent-general-free.mdc

@@ -26,7 +26,9 @@ Skip this only if the task explicitly says to attach to an existing session.
 ## Do
 
 - Crawl only public, high-signal company pages.
-- Extract emails, role inboxes, contact forms, public names/titles, source URL, evidence, and warnings.
+- Use your judgment on each email candidate: only emit an email lead when public evidence identifies one specific human owner of the address.
+- Extract high-quality named human emails, contact forms, public names/titles, source URL, evidence, and warnings.
+- If an address is shared, role-based, departmental, generic, blocked, or not tied to a visible named human, do not emit it as an email lead; emit a `contact_path` if useful.
 - Emit JSON/JSONL matching `templates/lead-schema.json`.
 - Run `npx public-leads validate --input <file>` before returning a result.
 - Return file paths and terminal status, not a prose-only summary.

package/.cursor/rules/main.mdc

@@ -10,6 +10,7 @@ Agentic public-web lead discovery harness. It crawls company domains, extracts r
 ## Hard Limits
 
 - [H1] Only collect public-source business contact data. Do not infer, guess, brute-force, or enrich private email addresses. Every lead needs a source URL and evidence excerpt. The primary goal is named people, not generic company inboxes.
+- [H8] Apply coding-agent judgment to every email candidate. An email may pass only when public source evidence identifies one specific human owner of that email. Shared, role, departmental, generic, blocked, unknown-owner, or unnamed person-like emails must not be emitted as email leads.
 - [H2] Respect robots.txt, terms, rate limits, and site intent. If a page blocks crawling or is not public, record a warning and move on.
 - [H3] Never send outreach from this harness. The output is reviewable lead records and ingest-ready artifacts, not email campaigns.
 - [H4] Use file-backed artifacts before ingest. A subagent prose summary is not enough; write validated JSON/JSONL first, then ingest.
@@ -23,7 +24,7 @@ Agentic public-web lead discovery harness. It crawls company domains, extracts r
 - [D2] Use deterministic local helpers when they can validate, normalize, manifest, preflight, postflight, or ingest artifacts.
 - [D3] Read the active mode file before executing. Mode files own output shapes and gates.
 - [D4] Prefer official company websites and high-signal pages: home, contact, about, team, leadership, people, press, blog, careers, and legal pages.
-- [D5] Keep artifacts source-backed. Every emitted lead should include `domain`, `sourceUrl`, `evidence`, `emailType`, `extractionMethod`, `verificationStatus`, and `confidence`. Reject generic catch-all inboxes such as `info@`, `hello@`, `contact@`, `support@`, `team@`, and similar aliases; if no named person is available, prefer a `contact_path` record instead.
+- [D5] Keep artifacts source-backed. Every emitted lead should include `domain`, `sourceUrl`, `evidence`, `emailType`, `extractionMethod`, `verificationStatus`, and `confidence`. Reject generic catch-all inboxes such as `info@`, `hello@`, `contact@`, `support@`, `team@`, and similar aliases; if no named human email is available, prefer a `contact_path` record instead.
 
 ## Procedure
 

package/.opencode/agents/general-free.md

@@ -36,7 +36,9 @@ Skip this only if the task explicitly says to attach to an existing session.
 ## Do
 
 - Crawl only public, high-signal company pages.
-- Extract emails, role inboxes, contact forms, public names/titles, source URL, evidence, and warnings.
+- Use your judgment on each email candidate: only emit an email lead when public evidence identifies one specific human owner of the address.
+- Extract high-quality named human emails, contact forms, public names/titles, source URL, evidence, and warnings.
+- If an address is shared, role-based, departmental, generic, blocked, or not tied to a visible named human, do not emit it as an email lead; emit a `contact_path` if useful.
 - Emit JSON/JSONL matching `templates/lead-schema.json`.
 - Run `npx public-leads validate --input <file>` before returning a result.
 - Return file paths and terminal status, not a prose-only summary.

package/.pi/skills/general-free/SKILL.md

@@ -26,7 +26,9 @@ Skip this only if the task explicitly says to attach to an existing session.
 ## Do
 
 - Crawl only public, high-signal company pages.
-- Extract emails, role inboxes, contact forms, public names/titles, source URL, evidence, and warnings.
+- Use your judgment on each email candidate: only emit an email lead when public evidence identifies one specific human owner of the address.
+- Extract high-quality named human emails, contact forms, public names/titles, source URL, evidence, and warnings.
+- If an address is shared, role-based, departmental, generic, blocked, or not tied to a visible named human, do not emit it as an email lead; emit a `contact_path` if useful.
 - Emit JSON/JSONL matching `templates/lead-schema.json`.
 - Run `npx public-leads validate --input <file>` before returning a result.
 - Return file paths and terminal status, not a prose-only summary.

package/AGENTS.md

@@ -5,6 +5,7 @@ Agentic public-web lead discovery harness. It crawls company domains, extracts r
 ## Hard Limits
 
 - [H1] Only collect public-source business contact data. Do not infer, guess, brute-force, or enrich private email addresses. Every lead needs a source URL and evidence excerpt. The primary goal is named people, not generic company inboxes.
+- [H8] Apply coding-agent judgment to every email candidate. An email may pass only when public source evidence identifies one specific human owner of that email. Shared, role, departmental, generic, blocked, unknown-owner, or unnamed person-like emails must not be emitted as email leads.
 - [H2] Respect robots.txt, terms, rate limits, and site intent. If a page blocks crawling or is not public, record a warning and move on.
 - [H3] Never send outreach from this harness. The output is reviewable lead records and ingest-ready artifacts, not email campaigns.
 - [H4] Use file-backed artifacts before ingest. A subagent prose summary is not enough; write validated JSON/JSONL first, then ingest.
@@ -18,7 +19,7 @@ Agentic public-web lead discovery harness. It crawls company domains, extracts r
 - [D2] Use deterministic local helpers when they can validate, normalize, manifest, preflight, postflight, or ingest artifacts.
 - [D3] Read the active mode file before executing. Mode files own output shapes and gates.
 - [D4] Prefer official company websites and high-signal pages: home, contact, about, team, leadership, people, press, blog, careers, and legal pages.
-- [D5] Keep artifacts source-backed. Every emitted lead should include `domain`, `sourceUrl`, `evidence`, `emailType`, `extractionMethod`, `verificationStatus`, and `confidence`. Reject generic catch-all inboxes such as `info@`, `hello@`, `contact@`, `support@`, `team@`, and similar aliases; if no named person is available, prefer a `contact_path` record instead.
+- [D5] Keep artifacts source-backed. Every emitted lead should include `domain`, `sourceUrl`, `evidence`, `emailType`, `extractionMethod`, `verificationStatus`, and `confidence`. Reject generic catch-all inboxes such as `info@`, `hello@`, `contact@`, `support@`, `team@`, and similar aliases; if no named human email is available, prefer a `contact_path` record instead.
 
 ## Procedure
 

package/CLAUDE.md

@@ -5,6 +5,7 @@ Agentic public-web lead discovery harness. It crawls company domains, extracts r
 ## Hard Limits
 
 - [H1] Only collect public-source business contact data. Do not infer, guess, brute-force, or enrich private email addresses. Every lead needs a source URL and evidence excerpt. The primary goal is named people, not generic company inboxes.
+- [H8] Apply coding-agent judgment to every email candidate. An email may pass only when public source evidence identifies one specific human owner of that email. Shared, role, departmental, generic, blocked, unknown-owner, or unnamed person-like emails must not be emitted as email leads.
 - [H2] Respect robots.txt, terms, rate limits, and site intent. If a page blocks crawling or is not public, record a warning and move on.
 - [H3] Never send outreach from this harness. The output is reviewable lead records and ingest-ready artifacts, not email campaigns.
 - [H4] Use file-backed artifacts before ingest. A subagent prose summary is not enough; write validated JSON/JSONL first, then ingest.
@@ -18,7 +19,7 @@ Agentic public-web lead discovery harness. It crawls company domains, extracts r
 - [D2] Use deterministic local helpers when they can validate, normalize, manifest, preflight, postflight, or ingest artifacts.
 - [D3] Read the active mode file before executing. Mode files own output shapes and gates.
 - [D4] Prefer official company websites and high-signal pages: home, contact, about, team, leadership, people, press, blog, careers, and legal pages.
-- [D5] Keep artifacts source-backed. Every emitted lead should include `domain`, `sourceUrl`, `evidence`, `emailType`, `extractionMethod`, `verificationStatus`, and `confidence`. Reject generic catch-all inboxes such as `info@`, `hello@`, `contact@`, `support@`, `team@`, and similar aliases; if no named person is available, prefer a `contact_path` record instead.
+- [D5] Keep artifacts source-backed. Every emitted lead should include `domain`, `sourceUrl`, `evidence`, `emailType`, `extractionMethod`, `verificationStatus`, and `confidence`. Reject generic catch-all inboxes such as `info@`, `hello@`, `contact@`, `support@`, `team@`, and similar aliases; if no named human email is available, prefer a `contact_path` record instead.
 
 ## Procedure
 

package/README.md

@@ -12,6 +12,7 @@ Agentic public-web lead discovery harness for portable, source-backed lead artif
 - Ships OpenCode, Claude Code, Cursor, Codex, and Pi harness config from one `iso/` source.
 
 The harness is optimized to find named people who can actually be contacted. When only a general company contact path exists, it should emit a `contact_path` record rather than a generic organizational inbox.
+CLI summaries distinguish total reviewable records from `goodLeads`; only non-generic `person` email records with a named human owner visible in the evidence count as good leads. Role, shared, departmental, unknown-owner, blocked, unsupported, or unnamed person-like emails are rejected by validation.
 
 ## Core Commands
 
@@ -70,7 +71,7 @@ gh run watch
 You can also publish by creating a GitHub release for the package version:
 
 ```bash
-gh release create v0.1.4 --title v0.1.4 --generate-notes
+gh release create v0.1.6 --title v0.1.6 --generate-notes
 ```
 
 The package version must not already exist on npm.

package/batch/README.md

@@ -17,10 +17,11 @@ id	domain	company	notes
 ```bash
 batch/batch-runner.sh --dry-run
 batch/batch-runner.sh --parallel 2
+batch/batch-runner.sh --parallel 2 --timeout-ms 8000
 batch/batch-runner.sh --runner codex --parallel 2
 ```
 
-`--parallel` is capped at `2` to keep browser-heavy work bounded. Worker CLI permission-bypass flags are disabled by default; pass `--allow-unsafe-workers` only in a trusted local workspace when you explicitly want the old unsafe behavior.
+`--parallel` is capped at `2` to keep browser-heavy work bounded. `--timeout-ms` controls the deterministic crawler timeout suggested to workers; the batch default is `8000` to avoid spending too long on dead sites. Worker CLI permission-bypass flags are disabled by default; pass `--allow-unsafe-workers` only in a trusted local workspace when you explicitly want the old unsafe behavior.
 
 The runner writes:
 

package/bin/create-leads-harness.mjs

@@ -51,7 +51,7 @@ write('package.json', JSON.stringify({
     'update-harness': 'npm install @agent-pattern-labs/leads-rig@latest && public-leads sync',
   },
   dependencies: {
-    '@agent-pattern-labs/leads-rig': '^0.1.4',
+    '@agent-pattern-labs/leads-rig': '^0.1.6',
   },
   engines: {
     node: '>=20.6.0',

package/config/profile.example.yml

@@ -18,10 +18,23 @@ api:
 
 crawl:
   max_pages: 10
+  timeout_ms: 14000
   max_domains_per_batch: 50
   concurrency: 2
+  page_concurrency: 1
+  dns_concurrency: 8
+  cache_path: ".leadharness-cache/crawler-cache.json"
+  robots_cache_ttl_ms: 86400000
+  dns_cache_ttl_ms: 604800000
+  page_cache: false
+  page_cache_ttl_ms: 0
+  stop_after_good_leads: 0
+  stop_after_contact_path: false
   user_agent: "PublicLeadsBot/0.1 (+https://example.com/public-leads)"
 
+batch:
+  batch_timeout_ms: 8000
+
 lead_policy:
   min_confidence: 30
   allowed_email_types:

package/docs/ARCHITECTURE.md

@@ -82,4 +82,4 @@ The package also carries the `@agent-pattern-labs/iso-*` helper ecosystem for tr
 - page visits
 - ingest requests
 
-The validator intentionally accepts the same defaults the local runtime normalizes, but it fails missing source evidence, invalid URL fields, invalid `emailType`, invalid `verificationStatus`, invalid confidence, person/role leads without email, and generic catch-all inboxes such as `info@`, `hello@`, or similar organizational aliases.
+The validator intentionally accepts the same defaults the local runtime normalizes, but it fails missing source evidence, invalid URL fields, invalid `emailType`, invalid `verificationStatus`, invalid confidence, person leads without email, and generic catch-all inboxes such as `info@`, `hello@`, or similar organizational aliases. Any email record must be a named `person` lead with a non-generic email and a human owner visible in the evidence; role inboxes, unknown-owner emails, blocked emails, unsupported emails, and unnamed person-like emails fail validation. Summaries expose `goodLeadCount`, which only counts those high-quality named human email records.

package/docs/CONSTRUCTION.md

@@ -37,4 +37,4 @@
 6. Run `npm audit`.
 7. Run `npm pack` and confirm generated runtime surfaces are included.
 8. Scaffold a clean consumer project and test install-time sync from the packed package or source tree.
-9. Publish through GitHub Actions with `gh workflow run publish.yml --ref main`, or create a published release with `gh release create v0.1.4 --title v0.1.4 --generate-notes`.
+9. Publish through GitHub Actions with `gh workflow run publish.yml --ref main`, or create a published release with `gh release create v0.1.6 --title v0.1.6 --generate-notes`.

package/iso/agents/general-free.md

@@ -38,7 +38,9 @@ Skip this only if the task explicitly says to attach to an existing session.
 ## Do
 
 - Crawl only public, high-signal company pages.
-- Extract emails, role inboxes, contact forms, public names/titles, source URL, evidence, and warnings.
+- Use your judgment on each email candidate: only emit an email lead when public evidence identifies one specific human owner of the address.
+- Extract high-quality named human emails, contact forms, public names/titles, source URL, evidence, and warnings.
+- If an address is shared, role-based, departmental, generic, blocked, or not tied to a visible named human, do not emit it as an email lead; emit a `contact_path` if useful.
 - Emit JSON/JSONL matching `templates/lead-schema.json`.
 - Run `npx public-leads validate --input <file>` before returning a result.
 - Return file paths and terminal status, not a prose-only summary.

package/iso/instructions.md

@@ -5,6 +5,7 @@ Agentic public-web lead discovery harness. It crawls company domains, extracts r
 ## Hard Limits
 
 - [H1] Only collect public-source business contact data. Do not infer, guess, brute-force, or enrich private email addresses. Every lead needs a source URL and evidence excerpt. The primary goal is named people, not generic company inboxes.
+- [H8] Apply coding-agent judgment to every email candidate. An email may pass only when public source evidence identifies one specific human owner of that email. Shared, role, departmental, generic, blocked, unknown-owner, or unnamed person-like emails must not be emitted as email leads.
 - [H2] Respect robots.txt, terms, rate limits, and site intent. If a page blocks crawling or is not public, record a warning and move on.
 - [H3] Never send outreach from this harness. The output is reviewable lead records and ingest-ready artifacts, not email campaigns.
 - [H4] Use file-backed artifacts before ingest. A subagent prose summary is not enough; write validated JSON/JSONL first, then ingest.
@@ -18,7 +19,7 @@ Agentic public-web lead discovery harness. It crawls company domains, extracts r
 - [D2] Use deterministic local helpers when they can validate, normalize, manifest, preflight, postflight, or ingest artifacts.
 - [D3] Read the active mode file before executing. Mode files own output shapes and gates.
 - [D4] Prefer official company websites and high-signal pages: home, contact, about, team, leadership, people, press, blog, careers, and legal pages.
-- [D5] Keep artifacts source-backed. Every emitted lead should include `domain`, `sourceUrl`, `evidence`, `emailType`, `extractionMethod`, `verificationStatus`, and `confidence`. Reject generic catch-all inboxes such as `info@`, `hello@`, `contact@`, `support@`, `team@`, and similar aliases; if no named person is available, prefer a `contact_path` record instead.
+- [D5] Keep artifacts source-backed. Every emitted lead should include `domain`, `sourceUrl`, `evidence`, `emailType`, `extractionMethod`, `verificationStatus`, and `confidence`. Reject generic catch-all inboxes such as `info@`, `hello@`, `contact@`, `support@`, `team@`, and similar aliases; if no named human email is available, prefer a `contact_path` record instead.
 
 ## Procedure