lakehouse/mcp-server/observer.ts

/**
 * Lakehouse Observer — autonomous iteration loop.
 *
 * Runs continuously alongside the agent gateway. Watches every operation,
 * logs outcomes, detects failures, and feeds learnings back so agents
 * improve over time without retraining.
 *
 * Three loops:
 * 1. OPERATION OBSERVER — wraps gateway calls, timestamps + logs every
 *    success/failure to the lakehouse
 * 2. ERROR ANALYZER — periodically reads the error log, asks a local
 *    model to diagnose patterns, writes recommendations
 * 3. PLAYBOOK BUILDER — after N successful ops of the same type,
 *    consolidates them into a reusable playbook entry
 *
 * This is the "third-party witness" J asked for — it watches what
 * agents do and helps them not repeat mistakes.
 */

import { filterChunks } from "./relevance";

const GATEWAY = process.env.GATEWAY_URL || "http://localhost:3700";
const LAKEHOUSE = process.env.LAKEHOUSE_URL || "http://localhost:3100";
const CYCLE_SECS = parseInt(process.env.OBSERVER_CYCLE || "30");
// Phase 24 — observer now listens on an HTTP port for external ops
// (scenarios bypass the MCP:3700 layer by design). Default 3800.
const OBSERVER_PORT = parseInt(process.env.OBSERVER_PORT || "3800");

// ─── Observed operation log ───

interface ObservedOp {
  timestamp: string;
  endpoint: string;
  input_summary: string;
  success: boolean;
  duration_ms: number;
  output_summary: string;
  error?: string;
  // Phase 24 — optional provenance so error analyzer and playbook
  // builder can differentiate MCP-layer ops from scenario-sourced
  // events. Scenarios set source="scenario" + staffer_id + sig_hash.
  source?: "mcp" | "scenario" | "langfuse" | "overseer_correction";
  staffer_id?: string;
  sig_hash?: string;
  event_kind?: string;
  role?: string;
  city?: string;
  state?: string;
  count?: number;
  rescue_attempted?: boolean;
  rescue_succeeded?: boolean;
  // Overseer-correction-specific (2026-04-23): lets the analyzer
  // correlate corrections with the drift that prompted them and with
  // subsequent outcomes that either validated or invalidated the advice.
  task_class?: string;
  correction?: string;
  applied_at_turn?: number;
}

const recentOps: ObservedOp[] = [];

// Phase 24 — external ingest path. Scenarios POST outcome summaries
// here so the observer's analyzer + playbook builder see them. Called
// from the Bun.serve() handler below. Same ring buffer as the MCP-
// wrapped path so downstream loops don't need to know the source.
export function recordExternalOp(op: ObservedOp): void {
  recentOps.push({ ...op, source: op.source ?? "scenario" });
  if (recentOps.length > 2000) recentOps.shift();
}

// ─── Wrapped gateway caller — every call gets observed ───

export async function observed(
  endpoint: string,
  body: any,
  description: string,
): Promise<{ data: any; op: ObservedOp }> {
  const t0 = Date.now();
  let data: any;
  let error: string | undefined;
  let success = true;

  try {
    const resp = await fetch(`${GATEWAY}${endpoint}`, {
      method: "POST",
      headers: { "Content-Type": "application/json" },
      body: JSON.stringify(body),
    });
    data = await resp.json();
    if (data.error) {
      success = false;
      error = data.error;
    }
  } catch (e: any) {
    success = false;
    error = e.message;
    data = { error: e.message };
  }

  const op: ObservedOp = {
    timestamp: new Date().toISOString(),
    endpoint,
    input_summary: description,
    success,
    duration_ms: Date.now() - t0,
    output_summary: success
      ? summarize(data)
      : `ERROR: ${error}`,
    error,
  };

  recentOps.push(op);
  if (recentOps.length > 1000) recentOps.shift();

  // Persist to lakehouse
  await persistOp(op);

  return { data, op };
}

function summarize(data: any): string {
  if (data.sql_matches !== undefined) return `hybrid: ${data.sql_matches} sql → ${data.vector_reranked} results`;
  if (data.rows) return `${data.row_count || data.rows.length} rows`;
  if (data.answer) return `answer: ${data.answer.slice(0, 80)}...`;
  if (data.sources) return `${data.sources.length} sources`;
  return JSON.stringify(data).slice(0, 100);
}

// Phase 24 honesty fix — the old persistOp used /ingest/file which
// REPLACES the dataset (flagged in feedback_ingest_replace_semantics.md).
// Every op silently wiped all prior ops. Now we append a JSONL line to
// data/_observer/ops.jsonl so the historical trace is durable. The
// observer analyzer + playbook builder read from this file when it
// outgrows the 2000-entry in-memory ring.
async function persistOp(op: ObservedOp) {
  try {
    const { mkdir, appendFile } = await import("node:fs/promises");
    await mkdir("data/_observer", { recursive: true });
    await appendFile("data/_observer/ops.jsonl", JSON.stringify(op) + "\n");
  } catch {
    // Persistence is best-effort; in-memory ring still works.
  }
}


// ─── LLM Team escalation (code_review mode) ───
//
// When recent failures on a single sig_hash cross a threshold the
// local qwen2.5 analysis is probably insufficient. J's 2026-04-24
// direction: "the observer would trigger to give more context" —
// route failure clusters to LLM Team's specialized code_review mode
// (via /api/run) so richer structured signal lands in the KB for
// scrum + auditor + playbook memory to consume next pass.
//
// Non-destructive: runs in parallel to the existing qwen2.5 analysis,
// never replaces it. Writes to data/_kb/observer_escalations.jsonl
// as a dedicated audit surface.

const LLM_TEAM = process.env.LH_LLM_TEAM_URL ?? "http://localhost:5000";
const LLM_TEAM_ESCALATIONS = "/home/profit/lakehouse/data/_kb/observer_escalations.jsonl";
const ESCALATION_THRESHOLD = 3; // N+ failures on same sig_hash triggers

// ─── KB enrichment helper (2026-04-26) ────────────────────────────
// Mirrors what scrum_master_pipeline already does on every per-file
// review: queries pathway_memory bug fingerprints + the lakehouse_arch
// matrix corpus, then asks qwen3.5:latest to synthesize a tight
// briefing. We reuse the same primitives so observer escalations carry
// the same compounding context the scrum loop builds — no new index
// surfaces, no new corpora.
//
// `task_class` is derived from the cluster (most ops use the same one);
// pathway/bug_fingerprints is permissive about a null file_path, so
// non-code clusters (scenario fills, v1.chat events) just see broader
// matches via task_class alone.
//
// Returns "" when there's no useful signal — caller treats empty as
// "no preamble" and skips the prepend.
async function buildKbPreamble(sigHash: string, cluster: ObservedOp[]): Promise<string> {
  const sample = cluster[0];
  const taskClass = sample?.event_kind
    ?? (sample?.source === "scenario" ? "scenario_fill" : "observer_escalation");

  // Step 1: pathway bug fingerprints. Best-effort; null filePath just
  // widens the query at the matrix-index level.
  let fingerprints: { flag: { kind: string }; pattern_key: string; example: string; occurrences: number }[] = [];
  try {
    const r = await fetch(`${LAKEHOUSE}/vectors/pathway/bug_fingerprints`, {
      method: "POST",
      headers: { "content-type": "application/json" },
      body: JSON.stringify({ task_class: taskClass, file_path: null, signal_class: null, limit: 5 }),
      signal: AbortSignal.timeout(5000),
    });
    if (r.ok) fingerprints = (await r.json() as any).fingerprints ?? [];
  } catch {}

  // Step 2: architectural matrix (lakehouse_arch_v1) — ADRs/PRD/plan
  // intent. Cluster summary is the search query.
  const clusterSummary = cluster.slice(-5).map(o =>
    `${o.endpoint ?? "?"} ${o.input_summary ?? ""} ${o.error ?? ""}`
  ).join(" | ");
  let matrixChunks: { doc_id?: string; chunk_text?: string; score?: number }[] = [];
  try {
    const r = await fetch(`${LAKEHOUSE}/vectors/search`, {
      method: "POST",
      headers: { "content-type": "application/json" },
      body: JSON.stringify({ index_name: "lakehouse_arch_v1", query: `${taskClass} ${clusterSummary}`, top_k: 4 }),
      signal: AbortSignal.timeout(5000),
    });
    if (r.ok) matrixChunks = (await r.json() as any).results ?? [];
  } catch {}

  // Step 3: gold-standard prior answers (lakehouse_answers_v1) — past
  // scrum reviews + observer escalations. This is where the BIG-model
  // results we save live; future small-model handlers retrieve them
  // here as scaffolding so the cheap rung gets near-paid quality.
  let answerChunks: { doc_id?: string; chunk_text?: string; score?: number }[] = [];
  try {
    const r = await fetch(`${LAKEHOUSE}/vectors/search`, {
      method: "POST",
      headers: { "content-type": "application/json" },
      body: JSON.stringify({ index_name: "lakehouse_answers_v1", query: `${taskClass} ${clusterSummary}`, top_k: 3 }),
      signal: AbortSignal.timeout(5000),
    });
    if (r.ok) answerChunks = (await r.json() as any).results ?? [];
  } catch {}

  if (fingerprints.length === 0 && matrixChunks.length === 0 && answerChunks.length === 0) return "";

  // Step 3: synthesis via local model (qwen3.5:latest, provider=ollama).
  // Compresses the raw bundle to a 1-2 sentence briefing the cloud
  // reviewer can actually use. If local model is down/slow, fall back
  // to the raw dump rather than blocking the escalation path.
  const rawBundle = [
    fingerprints.length > 0
      ? "PRIOR BUG PATTERNS (pathway memory):\n" + fingerprints.map((fp, i) =>
          `${i + 1}. [${fp.flag.kind}] ${fp.pattern_key} (×${fp.occurrences}) e.g. ${fp.example.slice(0, 120)}`
        ).join("\n")
      : "",
    matrixChunks.length > 0
      ? "RELATED ARCHITECTURE CONTEXT:\n" + matrixChunks.map((c, i) =>
          `${i + 1}. [${c.doc_id ?? "?"}] ${(c.chunk_text ?? "").slice(0, 200)}`
        ).join("\n")
      : "",
    answerChunks.length > 0
      ? "PRIOR GOLD-STANDARD ANSWERS (similar past reviews + escalations):\n" + answerChunks.map((c, i) =>
          `${i + 1}. [${c.doc_id ?? "?"}] ${(c.chunk_text ?? "").slice(0, 240)}`
        ).join("\n")
      : "",
  ].filter(Boolean).join("\n\n");

  const synthPrompt = `A failure cluster (sig_hash=${sigHash.slice(0, 8)}, ${cluster.length} occurrences, task_class=${taskClass}) is about to be escalated for diagnosis. Here are prior signals from our knowledge base:

${rawBundle}

Output a single paragraph (≤300 chars) briefing the cloud reviewer on which prior signals are most likely relevant to this cluster. If nothing matches, say so plainly. No preamble, no markdown.`;

  let synthesized = "";
  try {
    const r = await fetch(`${LAKEHOUSE}/v1/chat`, {
      method: "POST",
      headers: { "content-type": "application/json" },
      body: JSON.stringify({
        provider: "ollama",
        model: "qwen3.5:latest",
        messages: [{ role: "user", content: synthPrompt }],
        max_tokens: 200,
        temperature: 0.1,
        think: false,
      }),
      signal: AbortSignal.timeout(15000),
    });
    if (r.ok) {
      const j = await r.json() as any;
      synthesized = (j?.choices?.[0]?.message?.content ?? "").trim();
    }
  } catch {}

  const body = synthesized.length > 0 ? synthesized : rawBundle;
  return `═══ KB CONTEXT — prior signals on this task class (synthesized by qwen3.5:latest) ═══\n${body}\n═══\n\n`;
}

async function escalateFailureClusterToLLMTeam(sigHash: string, cluster: ObservedOp[]) {
  // Package the failure cluster as a single context blob. Originally
  // I routed this to LLM Team's `code_review` mode at /api/run, but
  // that mode isn't registered in llm_team_ui.py — it returned
  // "Unknown mode" on every call. Revised 2026-04-24: route directly
  // to the gateway's /v1/chat with provider=ollama_cloud + qwen3-coder:480b
  // (the coding specialist that's rung 2 of the scrum ladder, proven
  // to produce substantive structured reviews). Fire-and-forget so
  // downstream failures don't block observer's normal loop.
  const context = cluster.slice(-8).map((o, i) =>
    `[${i + 1}] endpoint=${o.endpoint} input=${o.input_summary} error=${o.error ?? "?"}`
  ).join("\n");
  const kbPreamble = await buildKbPreamble(sigHash, cluster);
  const prompt = `${kbPreamble}sig_hash=${sigHash} · ${cluster.length} failures on the same signature:\n\n${context}\n\nReview this failure cluster. Identify:\n1. Likely root cause (single sentence).\n2. Files most likely responsible (path hints).\n3. Concrete fix direction (under 3 sentences).\n4. Confidence: NN%\n\nBe specific, not generic.`;

  try {
    // 2026-04-26: switched from ollama_cloud/qwen3-coder:480b (weekly
    // 429 quota was blocking escalations) to paid OpenRouter
    // deepseek-v3.1-terminus — 671B reasoning specialist, $0.21 in /
    // $0.79 out per M tokens (under the $0.85/M ceiling J set), 164K
    // ctx. Per-escalation cost: ~$0.0006 (typical 500-token prompt +
    // 300-token completion).
    const resp = await fetch(`${LAKEHOUSE}/v1/chat`, {
      method: "POST",
      headers: { "Content-Type": "application/json" },
      body: JSON.stringify({
        provider: "openrouter",
        model: "deepseek/deepseek-v3.1-terminus",
        messages: [{ role: "user", content: prompt }],
        max_tokens: 800,
        temperature: 0.2,
      }),
      signal: AbortSignal.timeout(60000),
    });
    if (!resp.ok) {
      console.error(`[observer] escalation /v1/chat ${resp.status}: ${(await resp.text()).slice(0, 200)}`);
      return;
    }
    const j: any = await resp.json();
    const analysis = j?.choices?.[0]?.message?.content ?? "";

    // Audit row stays schema-compatible with the prior implementation —
    // downstream consumers see structured fields regardless of the
    // review-source change. Facts/entities stay empty (this call is
    // direct-model, not extract-mode); the raw analysis carries the
    // signal.
    const row = {
      ts: new Date().toISOString(),
      source: "observer_escalation",
      mode: "direct_chat_deepseek_v3_1_terminus",
      sig_hash: sigHash,
      cluster_size: cluster.length,
      cluster_staffer: cluster[0]?.staffer_id,
      cluster_endpoint: cluster[0]?.endpoint,
      prompt_tokens: j?.usage?.prompt_tokens ?? 0,
      completion_tokens: j?.usage?.completion_tokens ?? 0,
      kb_preamble_chars: kbPreamble.length,
      analysis: analysis.slice(0, 4000),
    };
    const { appendFile } = await import("node:fs/promises");
    await appendFile(LLM_TEAM_ESCALATIONS, JSON.stringify(row) + "\n");
    console.error(
      `[observer] escalated sig_hash=${sigHash.slice(0, 8)} · cluster=${cluster.length} · ${analysis.length} chars`
    );
  } catch (e) {
    console.error(`[observer] escalation failed: ${(e as Error).message}`);
  }
}

// Track which sig_hashes we've already escalated this session so we
// don't hammer LLM Team on every analyzeErrors tick when a cluster
// persists across cycles.
const escalatedSigHashes = new Set<string>();

// ─── Hand-review for scrum/agent candidate responses (2026-04-25) ───
//
// Observer is OUTSIDE the scrum loop's epistemic scope, so its verdict
// can be treated as truth about whether a candidate review is grounded.
// Two-tier evaluator:
//   1. Try cloud LLM (qwen3-coder:480b) — semantic judgment with
//      response + source excerpt + grounding stats as context.
//   2. On cloud failure (throttle/timeout) → deterministic heuristic
//      over grounding_pct + total_quotes. Marked source: "heuristic"
//      so consumers can tell which rung produced the verdict.
// Every verdict is persisted to data/_kb/observer_reviews.jsonl.

const OBSERVER_REVIEWS = "/home/profit/lakehouse/data/_kb/observer_reviews.jsonl";

interface HandReviewInput {
  file_path: string;
  model: string;
  response: string;
  source_content: string;
  grounding_stats: { total: number; grounded: number; groundedPct: number | null };
  attempt: number;
}

interface HandReviewVerdict {
  verdict: "accept" | "reject" | "cycle";
  confidence: number;
  notes: string;
  source: "cloud" | "heuristic";
}

async function handReview(input: HandReviewInput): Promise<HandReviewVerdict> {
  const t0 = Date.now();
  let verdict: HandReviewVerdict;

  try {
    verdict = await cloudHandReview(input);
  } catch (e) {
    console.error(`[observer/review] cloud failed (${(e as Error).message}); using heuristic`);
    verdict = heuristicHandReview(input);
  }

  // Persist regardless of source so we can later compare cloud vs
  // heuristic verdicts on the same input and tune the heuristic.
  const row = {
    ts: new Date().toISOString(),
    file_path: input.file_path,
    model: input.model,
    attempt: input.attempt,
    response_chars: input.response.length,
    grounding_stats: input.grounding_stats,
    verdict: verdict.verdict,
    confidence: verdict.confidence,
    notes: verdict.notes,
    source: verdict.source,
    duration_ms: Date.now() - t0,
  };
  try {
    const { appendFile } = await import("node:fs/promises");
    await appendFile(OBSERVER_REVIEWS, JSON.stringify(row) + "\n");
  } catch { /* best-effort persistence */ }

  return verdict;
}

async function cloudHandReview(input: HandReviewInput): Promise<HandReviewVerdict> {
  const grounded = input.grounding_stats.grounded;
  const total = input.grounding_stats.total;
  const pct = input.grounding_stats.groundedPct;
  // Truncate to keep the prompt under typical context windows.
  // 2000 + 4000 = ~6000 chars ≈ 1500 tokens, plus response context.
  const responseExcerpt = input.response.slice(0, 2000);
  const sourceExcerpt = input.source_content.slice(0, 4000);

  const prompt = `You are a code-review quality observer. Decide whether the following automated review is grounded in the actual source — not invented, not hallucinated.

FILE: ${input.file_path}
MODEL: ${input.model}
ATTEMPT: ${input.attempt}
ANCHOR GROUNDING: ${grounded}/${total} backtick-quoted snippets matched the source verbatim${pct !== null ? ` (${pct}%)` : ""}

REVIEW (first 2000 chars):
\`\`\`
${responseExcerpt}
\`\`\`

SOURCE EXCERPT (first 4000 chars):
\`\`\`
${sourceExcerpt}
\`\`\`

Respond ONLY with a JSON object:
{
  "verdict": "accept" | "reject" | "cycle",
  "confidence": 0-100,
  "notes": "<1-2 sentences on what makes this grounded or hallucinated>"
}

- accept: review references real symbols/lines in source; findings could be acted on.
- reject: review invents APIs, fabricates calls, contradicts source. Do NOT record.
- cycle: review is mediocre — partially grounded but wrong shape, try a stronger model.`;

  // Hand-review uses paid OpenRouter so it sidesteps the Ollama Cloud
  // throttle that drove every prior iter into the heuristic fallback.
  // Grok 4.1 fast: $0.20 in / $0.50 out per M tokens, 2M ctx. A typical
  // hand-review (~6K input + 300 output) costs ~$0.0014. Selected via
  // J directive 2026-04-25 ("best model under $0.72/M").
  const resp = await fetch(`${LAKEHOUSE}/v1/chat`, {
    method: "POST",
    headers: { "Content-Type": "application/json" },
    body: JSON.stringify({
      provider: "openrouter",
      model: "x-ai/grok-4.1-fast",
      messages: [{ role: "user", content: prompt }],
      max_tokens: 300,
      temperature: 0.0,
    }),
    signal: AbortSignal.timeout(45000),
  });
  if (!resp.ok) {
    throw new Error(`/v1/chat ${resp.status}: ${(await resp.text()).slice(0, 200)}`);
  }
  const j: any = await resp.json();
  const content = (j?.choices?.[0]?.message?.content ?? "").trim();
  // Pull JSON object from the response — model may wrap it in prose.
  const m = content.match(/\{[\s\S]*\}/);
  if (!m) throw new Error(`no JSON object in response: ${content.slice(0, 100)}`);
  const parsed = JSON.parse(m[0]);
  const v = String(parsed.verdict ?? "accept").toLowerCase();
  return {
    verdict: (v === "reject" || v === "cycle") ? v as "reject" | "cycle" : "accept",
    confidence: Number(parsed.confidence ?? 50),
    notes: String(parsed.notes ?? "").slice(0, 500),
    source: "cloud",
  };
}

function heuristicHandReview(input: HandReviewInput): HandReviewVerdict {
  // Deterministic fallback when cloud is throttled. Conservative:
  // only flip to reject when the evidence is overwhelming, otherwise
  // accept (fall-open principle — observer is policy, not blocker).
  const total = input.grounding_stats.total;
  const pct = input.grounding_stats.groundedPct;
  const respLen = input.response.length;

  // Too short to be a real review
  if (respLen < 1500) {
    return { verdict: "reject", confidence: 80, notes: `response too short (${respLen} chars)`, source: "heuristic" };
  }
  // Below 5 quotes — not enough signal to judge grounding; accept
  if (total < 5 || pct === null) {
    return { verdict: "accept", confidence: 50, notes: `insufficient quote signal (${total} quotes); accepting`, source: "heuristic" };
  }
  // Very heavy hallucination
  if (pct < 20) {
    return { verdict: "reject", confidence: 85, notes: `low grounding (${pct}% of ${total} quotes)`, source: "heuristic" };
  }
  // Mediocre — cycle to a stronger model
  if (pct < 50) {
    return { verdict: "cycle", confidence: 65, notes: `mediocre grounding (${pct}% of ${total} quotes); try stronger`, source: "heuristic" };
  }
  // Good enough
  return { verdict: "accept", confidence: 75, notes: `grounding ${pct}% of ${total} quotes`, source: "heuristic" };
}

async function maybeEscalate(failures: ObservedOp[]) {
  // Group failures by sig_hash
  const bySig = new Map<string, ObservedOp[]>();
  for (const f of failures) {
    const k = f.sig_hash ?? "__no_sig__";
    (bySig.get(k) ?? bySig.set(k, []).get(k)!).push(f);
  }
  for (const [sigHash, cluster] of bySig) {
    if (sigHash === "__no_sig__") continue;
    if (cluster.length < ESCALATION_THRESHOLD) continue;
    if (escalatedSigHashes.has(sigHash)) continue;
    escalatedSigHashes.add(sigHash);
    // Fire-and-forget — don't block the existing analyzer loop.
    escalateFailureClusterToLLMTeam(sigHash, cluster).catch(() => {});
  }
}

// ─── Error analyzer loop ───

async function analyzeErrors() {
  // Read recent failures
  const failures = recentOps.filter(op => !op.success);
  if (failures.length === 0) return;

  // NEW 2026-04-24: escalate recurring sig_hash clusters to LLM Team
  // code_review mode. Runs in parallel to the local qwen2.5 analysis
  // below — non-blocking, richer downstream signal for scrum/auditor.
  maybeEscalate(failures).catch(() => {});

  const errorSummary = failures.slice(-10).map(f =>
    `[${f.endpoint}] ${f.input_summary}: ${f.error}`
  ).join("\n");

  // Ask local model to diagnose. Phase 44 migration (2026-04-27):
  // /v1/chat instead of legacy /ai/generate so /v1/usage tracks the
  // call + Langfuse traces it. Same upstream model (qwen2.5 local).
  try {
    const resp = await fetch(`${LAKEHOUSE}/v1/chat`, {
      method: "POST",
      headers: { "Content-Type": "application/json" },
      body: JSON.stringify({
        model: "qwen2.5",
        provider: "ollama",
        messages: [{
          role: "user",
          content: `You are a system reliability observer. Analyze these recent failures and suggest fixes:

${errorSummary}

For each error:
1. What likely caused it?
2. How should the agent adjust its approach?
3. Should this be added to the playbook as a "don't do this"?

Be specific and actionable. Under 200 words.`,
        }],
        max_tokens: 400,
        temperature: 0.2,
      }),
    });
    const analysis = await resp.json() as any;
    const analysisText = analysis?.choices?.[0]?.message?.content ?? "";
    if (analysisText) {
      console.error(`[observer] Error analysis:\n${analysisText}`);
      // Log the analysis as a playbook entry
      await fetch(`${GATEWAY}/log`, {
        method: "POST",
        headers: { "Content-Type": "application/json" },
        body: JSON.stringify({
          operation: `error_analysis: ${failures.length} failures`,
          approach: "LLM-analyzed error patterns",
          result: analysisText.slice(0, 500),
          context: errorSummary.slice(0, 500),
        }),
      });
    }
  } catch (e) {
    console.error(`[observer] Analysis failed: ${e}`);
  }
}

// ─── Playbook consolidation ───

async function consolidatePlaybooks() {
  const successes = recentOps.filter(op => op.success);
  if (successes.length < 5) return;

  // Group by endpoint
  const groups: Record<string, ObservedOp[]> = {};
  for (const op of successes) {
    const key = op.endpoint;
    if (!groups[key]) groups[key] = [];
    groups[key].push(op);
  }

  for (const [endpoint, ops] of Object.entries(groups)) {
    if (ops.length < 3) continue;

    const avgDuration = ops.reduce((s, o) => s + o.duration_ms, 0) / ops.length;
    const pattern = ops.slice(-3).map(o => o.input_summary).join("; ");

    await fetch(`${GATEWAY}/log`, {
      method: "POST",
      headers: { "Content-Type": "application/json" },
      body: JSON.stringify({
        operation: `consolidated: ${ops.length} successful ${endpoint} calls`,
        approach: `common pattern: ${pattern.slice(0, 200)}`,
        result: `avg_duration=${avgDuration.toFixed(0)}ms, ${ops.length} successes`,
        context: `endpoint=${endpoint}`,
      }),
    }).catch(() => {});
  }
}

// ─── HTTP listener for external ops (Phase 24) ───

// Scenarios POST per-event outcomes here so the observer's analyzer +
// playbook builder see them alongside MCP-wrapped ops. Read-only stats
// also exposed at /stats for external health checks.
function startHttpListener() {
  Bun.serve({
    port: OBSERVER_PORT,
    hostname: "0.0.0.0",
    fetch(req) {
      const url = new URL(req.url);
      if (req.method === "GET" && url.pathname === "/health") {
        return new Response(JSON.stringify({ status: "ok", ops_in_ring: recentOps.length }));
      }
      if (req.method === "GET" && url.pathname === "/stats") {
        const bySource = new Map<string, number>();
        for (const o of recentOps) {
          const k = o.source ?? "mcp";
          bySource.set(k, (bySource.get(k) ?? 0) + 1);
        }
        return new Response(JSON.stringify({
          total: recentOps.length,
          successes: recentOps.filter(o => o.success).length,
          failures: recentOps.filter(o => !o.success).length,
          by_source: Object.fromEntries(bySource),
          recent_scenario_ops: recentOps
            .filter(o => o.source === "scenario")
            .slice(-10)
            .map(o => ({ ts: o.timestamp, ok: o.success, staffer: o.staffer_id, kind: o.event_kind, role: o.role })),
        }));
      }
      // ─── Hand-review endpoint (2026-04-25) ───
      // scrum/agent posts a candidate response + source content + grounding
      // stats. Observer evaluates via cloud LLM (qwen3-coder:480b) with
      // semantic context and returns {verdict, confidence, notes}. On
      // cloud throttle, falls back to a deterministic heuristic over the
      // grounding stats so the loop keeps moving with honest signal.
      //
      // This is the policy layer scrum was missing — pre-2026-04-25 the
      // scrum_master applied a hardcoded grounding-rate threshold inline,
      // which baked judgment into the wrong layer. Now scrum reports data
      // (response + source + stats) and observer decides accept/reject/cycle.
      if (req.method === "POST" && url.pathname === "/review") {
        return req.json().then((body: any) => handReview(body))
          .then((verdict) => new Response(JSON.stringify(verdict), {
            headers: { "content-type": "application/json" },
          }))
          .catch((e: Error) =>
            new Response(JSON.stringify({ verdict: "accept", notes: `observer error: ${e.message}`, source: "heuristic" }), {
              status: 200,  // fall-open shape — scrum keeps moving on observer failure
              headers: { "content-type": "application/json" },
            }));
      }
      if (req.method === "POST" && url.pathname === "/event") {
        return req.json().then((body: any) => {
          const op: ObservedOp = {
            timestamp: body.timestamp ?? new Date().toISOString(),
            endpoint: body.endpoint ?? "scenario:fill",
            input_summary: body.input_summary ?? `${body.event_kind ?? "?"} ${body.role ?? "?"}×${body.count ?? "?"} in ${body.city ?? "?"}, ${body.state ?? "?"}`,
            success: !!body.success,
            duration_ms: Number(body.duration_ms ?? 0),
            output_summary: body.output_summary ?? (body.success ? "filled" : (body.error ?? "failed")),
            error: body.error,
            // Respect the client's provenance if set (langfuse bridge
            // sends source:"langfuse", etc.). Default to "scenario"
            // to keep legacy Phase 24 callers working.
            source: body.source ?? "scenario",
            staffer_id: body.staffer_id,
            sig_hash: body.sig_hash,
            event_kind: body.event_kind,
            role: body.role,
            city: body.city,
            state: body.state,
            count: body.count,
            rescue_attempted: !!body.rescue_attempted,
            rescue_succeeded: !!body.rescue_succeeded,
          };
          recordExternalOp(op);
          persistOp(op).catch(() => {});
          return new Response(JSON.stringify({ accepted: true, ring_size: recentOps.length }));
        }).catch((e: Error) =>
          new Response(JSON.stringify({ error: e.message }), { status: 400 }));
      }
      // ─── Relevance filter (2026-04-25) ───
      // Drops "adjacency pollution" from matrix-retrieved chunks before
      // they reach the reviewer LLM. Caller (scrum/agent) posts the
      // focus file + candidate chunks; observer scores via heuristic
      // (path/symbol/token signals) and returns kept + dropped lists.
      // Pure function — no I/O, safe to call hot.
      if (req.method === "POST" && url.pathname === "/relevance") {
        return req.json().then((body: any) => {
          const focus = body.focus_file ?? body.focus ?? {};
          const chunks = body.chunks ?? [];
          const threshold = typeof body.threshold === "number" ? body.threshold : 0.3;
          const result = filterChunks(focus, chunks, threshold);
          return new Response(JSON.stringify(result), {
            headers: { "content-type": "application/json" },
          });
        }).catch((e: Error) =>
          new Response(JSON.stringify({ error: e.message }), { status: 400 }));
      }
      return new Response("not found", { status: 404 });
    },
  });
  console.error(`[observer] HTTP listener bound to 0.0.0.0:${OBSERVER_PORT}`);
}

// ─── Overseer corrections tailer (2026-04-23) ───

// The gateway's /v1/respond loop writes T3 overseer corrections to
// data/_kb/overseer_corrections.jsonl. Tail it once per cycle and
// inject each new row into the same recentOps ring that analyzeErrors
// + consolidatePlaybooks read — so a correction that just fired shows
// up alongside the outcomes it was meant to repair, and the analyzer
// can flag patterns like "three corrections on staffing.fill with the
// same advice — underlying problem isn't a drift, it's a data gap".
const CORRECTIONS_PATH = process.env.OVERSEER_CORRECTIONS_PATH
  ?? "/home/profit/lakehouse/data/_kb/overseer_corrections.jsonl";
let correctionsCursor = 0;  // byte offset

async function tailOverseerCorrections(): Promise<number> {
  const f = Bun.file(CORRECTIONS_PATH);
  if (!(await f.exists())) return 0;
  const size = f.size;
  if (size <= correctionsCursor) return 0;

  // Read only the suffix since the last cursor; keeps tail work
  // bounded even as the file grows.
  const text = await f.slice(correctionsCursor, size).text();
  correctionsCursor = size;

  let forwarded = 0;
  for (const line of text.split("\n")) {
    if (!line.trim()) continue;
    let row: any;
    try { row = JSON.parse(line); } catch { continue; }
    const op: ObservedOp = {
      timestamp: row.created_at ?? new Date().toISOString(),
      endpoint: `overseer:${row.model ?? "gpt-oss:120b"}`,
      input_summary: `${row.task_class ?? "?"}: ${row.reason ?? "escalation"}`,
      // Correction itself is neither success nor failure — it's a
      // mitigation attempt. We mark success=true so analyzeErrors
      // doesn't count it as a failure, but the preview lets the
      // analyzer see what was tried.
      success: true,
      duration_ms: Number(row.usage?.latency_ms ?? 0),
      output_summary: String(row.correction ?? "").slice(0, 200),
      source: "overseer_correction",
      sig_hash: row.sig_hash,
      task_class: row.task_class,
      correction: String(row.correction ?? ""),
      applied_at_turn: Number(row.applied_at_turn ?? 0),
    };
    recordExternalOp(op);
    forwarded++;
  }
  return forwarded;
}

// ─── Main loop ───

async function main() {
  console.error(`[observer] started — cycle=${CYCLE_SECS}s, gateway=${GATEWAY}, port=${OBSERVER_PORT}`);

  // Run a health check first
  const health = await fetch(`${GATEWAY}/health`).then(r => r.ok ? r.text() : null).catch(() => null);
  if (!health) {
    console.error("[observer] gateway unreachable — exiting");
    process.exit(1);
  }
  console.error(`[observer] gateway healthy: ${JSON.stringify(health)}`);

  // Phase 24 — bind HTTP listener so scenarios can POST outcomes.
  startHttpListener();

  // Main loop
  let cycle = 0;
  while (true) {
    await Bun.sleep(CYCLE_SECS * 1000);
    cycle++;

    // Every cycle: tail the overseer corrections KB stream, then
    // analyze errors. Order matters — if an overseer correction just
    // landed for a sig_hash that previously failed, the analyzer
    // should see both.
    const newCorrections = await tailOverseerCorrections();
    if (newCorrections > 0) {
      console.error(`[observer] pulled ${newCorrections} new overseer correction(s) into ring`);
    }
    await analyzeErrors();

    // Every 5 cycles: consolidate playbooks
    if (cycle % 5 === 0) {
      await consolidatePlaybooks();
    }

    const scenarioOps = recentOps.filter(o => o.source === "scenario").length;
    const langfuseOps = recentOps.filter(o => o.source === "langfuse").length;
    const correctionOps = recentOps.filter(o => o.source === "overseer_correction").length;
    const stats = {
      cycle,
      total_ops: recentOps.length,
      successes: recentOps.filter(o => o.success).length,
      failures: recentOps.filter(o => !o.success).length,
      scenario_ops: scenarioOps,
      langfuse_ops: langfuseOps,
      overseer_corrections: correctionOps,
    };
    console.error(`[observer] cycle ${cycle}: ${JSON.stringify(stats)}`);
  }
}

// Export the observed wrapper for other agents to use
export { main as startObserver };

// Run if executed directly
if (import.meta.main) {
  main().catch(console.error);
}