[codex] Use LLM classification for job leads

.env.example

@@ -179,6 +179,11 @@ AGENT_STRONG_API_KEY=
 AGENT_REASONING_MODEL=
 AGENT_REASONING_BASE_URL=
 AGENT_REASONING_API_KEY=
+# HN job-lead scraping uses an LLM classifier when OpenAI-compatible credentials
+# are configured, then falls back to keyword matching if the model is unavailable.
+JOB_LEAD_CLASSIFIER_ENABLED=true
+JOB_LEAD_CLASSIFIER_MODEL=
+JOB_LEAD_CLASSIFIER_TIMEOUT_SECONDS=8.0
 # Optional deterministic agent tool integrations.
 # GitHub Issues are the canonical code-task backend.
 GITHUB_API_TOKEN=

apps/admin_dashboard/src/main.tsx

@@ -247,6 +247,16 @@ type JobLead = {
   apply_url?: string
   tags?: string[]
   confidence?: number
+  contractor_classification?: {
+    is_contractor_friendly?: boolean
+    posting_type?: string
+    tags?: string[]
+    confidence?: number
+    confidence_label?: string
+    rationale?: string
+    method?: string
+  }
+  review_summary?: string
   reviewed_by_discord_user_id?: string
   reviewed_at?: string
   discord_guild_id?: string
@@ -5580,6 +5590,16 @@ function defaultJobPostTagNames(lead: JobLead, channel?: JobPostChannel) {
   return bestTag ? [bestTag.name] : []
 }
 
+function jobLeadClassificationLabel(lead: JobLead) {
+  const method = lead.contractor_classification?.method
+  const confidenceLabel = lead.contractor_classification?.confidence_label
+  if (method && confidenceLabel) {
+    const methodLabel = method === "llm" ? "LLM" : "Keyword fallback"
+    return `${methodLabel}: ${confidenceLabel} fit`
+  }
+  return lead.review_summary || "Classification unavailable"
+}
+
 function JobLeadListItem({
   lead,
   loading,
@@ -5643,9 +5663,7 @@ function JobLeadListItem({
           <strong className="text-base">{lead.title || "Untitled lead"}</strong>
           <Badge variant={jobLeadStatusTone(lead.status)}>{titleCase(lead.status)}</Badge>
           {lead.remote ? <Badge variant="queued">Remote</Badge> : null}
-          {lead.confidence !== undefined ? (
-            <Badge variant="neutral">{Math.round(Number(lead.confidence || 0) * 100)}%</Badge>
-          ) : null}
+          <Badge variant="neutral">{jobLeadClassificationLabel(lead)}</Badge>
         </div>
         <div className="mt-2 flex flex-wrap gap-1.5">
           {lead.organization ? <Badge variant="neutral">{lead.organization}</Badge> : null}
@@ -5662,6 +5680,9 @@ function JobLeadListItem({
         <p className="mt-3 max-h-20 overflow-hidden text-sm text-muted-foreground">
           {lead.body_normalized || "No lead text captured."}
         </p>
+        {lead.review_summary ? (
+          <p className="mt-2 text-sm font-semibold text-foreground">{lead.review_summary}</p>
+        ) : null}
         <div className="mt-3 flex flex-wrap gap-x-4 gap-y-1 text-sm text-muted-foreground">
           <span>Posted {formatDate(lead.source_posted_at) || "unknown"}</span>
           <span>Captured {formatDate(lead.created_at) || "unknown"}</span>

apps/api/src/five08/backend/api.py

@@ -57,6 +57,7 @@
 from five08.clients.espo import EspoAPIError, EspoClient
 from five08.logging import configure_observability
 from five08.job_leads import (
+    job_lead_display_payload,
     JobLeadStatus,
     list_job_leads,
     review_job_lead,
@@ -4373,7 +4374,9 @@ async def dashboard_job_leads_handler(
         status=normalized_status,
         limit=limit,
     )
-    return JSONResponse(jsonable_encoder(leads))
+    return JSONResponse(
+        jsonable_encoder([job_lead_display_payload(lead) for lead in leads])
+    )
 
 
 async def dashboard_review_job_lead_handler(
@@ -4435,7 +4438,7 @@ async def dashboard_review_job_lead_handler(
             "status": payload.status,
         },
     )
-    return JSONResponse(jsonable_encoder(lead))
+    return JSONResponse(jsonable_encoder(job_lead_display_payload(lead)))
 
 
 async def dashboard_post_job_lead_handler(

apps/api/src/five08/backend/static/dashboard/.vite/manifest.json

@@ -1,6 +1,6 @@
 {
   "index.html": {
-    "file": "assets/index-CV7QH8Ei.js",
+    "file": "assets/index-Ce7YIjfO.js",
     "name": "index",
     "src": "index.html",
     "isEntry": true,

apps/api/src/five08/backend/static/dashboard/index.html

@@ -4,7 +4,7 @@
     <meta charset="UTF-8" />
     <meta name="viewport" content="width=device-width, initial-scale=1.0" />
     <title>508 Operations Dashboard</title>
-    <script type="module" crossorigin src="/dashboard/assets/index-CV7QH8Ei.js"></script>
+    <script type="module" crossorigin src="/dashboard/assets/index-Ce7YIjfO.js"></script>
     <link rel="stylesheet" crossorigin href="/dashboard/assets/index-2UypYUrJ.css">
   </head>
   <body>

apps/discord_bot/src/five08/discord_bot/cogs/jobs.py

@@ -65,6 +65,7 @@
     unregister_job_post_channel,
 )
 from five08.job_leads import (
+    format_job_lead_review_summary,
     get_job_lead,
     JobLead,
     JobLeadStatus,
@@ -3454,12 +3455,43 @@ def _truncate_job_lead_text(value: str, limit: int) -> str:
 
     @classmethod
     def _format_job_lead_review_line(cls, index: int, lead: JobLead) -> str:
-        tags = ", ".join(lead.tags[:5]) if lead.tags else "untagged"
         title = cls._truncate_job_lead_text(lead.title, 120)
-        return (
-            f"{index}. `{lead.id[:8]}` **{title}** "
-            f"({lead.confidence:.0%}; {tags})\n{lead.source_url}"
-        )
+        summary = cls._truncate_job_lead_text(format_job_lead_review_summary(lead), 180)
+        return f"{index}. `{lead.id[:8]}` **{title}**\n{summary}\n{lead.source_url}"
+
+    @classmethod
+    def _format_job_lead_review_message(cls, leads: list[JobLead]) -> str:
+        """Format a bounded Discord response for pending lead review."""
+        if not leads:
+            return "Pending job leads:\n\nNo pending job leads found."
+        limit = settings.discord_sendmsg_character_limit
+        header = "Pending job leads:"
+        lines: list[str] = []
+        shown = 0
+        for index, lead in enumerate(leads, start=1):
+            line = cls._format_job_lead_review_line(index, lead)
+            candidate_lines = [*lines, line]
+            omitted = len(leads) - len(candidate_lines)
+            footer = f"\n\nShowing {len(candidate_lines)} of {len(leads)} leads."
+            if omitted > 0:
+                footer += " Use a lower limit to inspect omitted leads."
+            candidate = f"{header}\n\n" + "\n\n".join(candidate_lines) + footer
+            if len(candidate) > limit:
+                break
+            lines = candidate_lines
+            shown = len(candidate_lines)
+
+        if not lines:
+            first = cls._truncate_job_lead_text(
+                cls._format_job_lead_review_line(1, leads[0]),
+                max(0, limit - len(header) - 32),
+            )
+            return f"{header}\n\n{first}"
+
+        footer = f"\n\nShowing {shown} of {len(leads)} leads."
+        if shown < len(leads):
+            footer += " Use a lower limit to inspect omitted leads."
+        return f"{header}\n\n" + "\n\n".join(lines) + footer
 
     @classmethod
     def _format_job_lead_thread_content(cls, lead: JobLead) -> str:
@@ -4014,12 +4046,8 @@ async def list_sourced_job_leads(
             )
             return
 
-        lines = [
-            self._format_job_lead_review_line(index, lead)
-            for index, lead in enumerate(leads, start=1)
-        ]
         await interaction.followup.send(
-            "Pending job leads:\n\n" + "\n\n".join(lines),
+            self._format_job_lead_review_message(leads),
             ephemeral=True,
         )
 

apps/worker/src/five08/worker/config.py

@@ -72,6 +72,9 @@ class WorkerSettings(SharedSettings):
     agent_reasoning_api_key: str | None = None
     agent_reasoning_base_url: str | None = None
     agent_reasoning_model: str | None = None
+    job_lead_classifier_enabled: bool = True
+    job_lead_classifier_model: str | None = None
+    job_lead_classifier_timeout_seconds: float = Field(default=8.0, gt=0)
     resume_ai_api_key: str | None = None
     resume_ai_base_url: str | None = None
     resume_ai_model: str = "gpt-4.1-mini"

docs/configuration.md

@@ -219,6 +219,9 @@ intake-completed field unset, and matches resume filenames with
 - `RESUME_AI_API_KEY`
 - `RESUME_AI_BASE_URL`
 - `RESUME_AI_MODEL`
+- `JOB_LEAD_CLASSIFIER_ENABLED`
+- `JOB_LEAD_CLASSIFIER_MODEL`
+- `JOB_LEAD_CLASSIFIER_TIMEOUT_SECONDS`
 
 Resume/profile LLM calls retry matching direct providers after Bifrost request
 failures when direct provider credentials are configured.