Improve signal quality (Exa news category) and remove mock fallback

Exa harvester: - EXA_CATEGORY=news (default) restricts to news articles, excluding evergreen careers/marketing landing pages that were producing irrelevant signals. - Event-focused query (acquisition/earnings/launch/leadership), dropped hiring/career terms that pulled recruiting pages. - Domain scoping now OFF by default (broad-web news about the account beats first-party careers pages); add EXA_EXCLUDE_DOMAINS and loop --domain-filter. LLM layer: - Remove the mock reasoner entirely (no provider => loop refuses to run). - On HTTP 429, WAIT (LLM_RATELIMIT_WAIT_SECONDS, default 900s; honors Retry-After) and retry up to LLM_RATELIMIT_MAX_RETRIES (24) instead of degrading to mock output. - If the LLM is ultimately unavailable for an account, skip it and leave its signals unconsumed (retried next run) — never fabricate briefs. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-24 10:41:41 -04:00
parent 2374ef137d
commit 9131e83c08
9 changed files with 169 additions and 173 deletions
--- a/sales-agent/.env.example
+++ b/sales-agent/.env.example
@@ -12,15 +12,17 @@ OPENAI_API_KEY=
 OPENAI_MODEL=gpt-4o-mini
 OPENAI_JSON_MODE=1          # set 0 if the endpoint rejects response_format=json_object

+# There is NO mock fallback. On HTTP 429 the loop WAITS this long and retries
+# (honors Retry-After if present), so it grinds through tight rate limits.
+LLM_RATELIMIT_WAIT_SECONDS=900
+LLM_RATELIMIT_MAX_RETRIES=24
+
 # Other providers:
 GEMINI_API_KEY=
 GEMINI_MODEL=gemini-2.5-flash
 ANTHROPIC_API_KEY=
 LLM_MODEL=claude-sonnet-4-6

-# Force the offline mock reasoner even if a key is set:
-# USE_MOCK_LLM=1
-
 # Scoring thresholds (design Phase 5)
 PRIORITY_THRESHOLD=0.72
 STANDARD_THRESHOLD=0.45
@@ -36,6 +38,12 @@ EXA_MONTHLY_REQUEST_CAP=1000
 EXA_PER_RUN_REQUEST_CAP=25
 EXA_RESULTS_PER_QUERY=5
 EXA_LOOKBACK_DAYS=45
-# Scope news to each account's own domain (1, default) to cut ticker/analyst
-# noise, or search the broad web (0). Override per run: loop --no-domain-filter
-EXA_DOMAIN_FILTER=1
+# Exa content category. "news" keeps out careers/marketing landing pages — the
+# biggest signal-quality lever. "" disables category filtering.
+EXA_CATEGORY=news
+# Scope to the account's own domain (1) or broad-web news about it (0, default).
+# With category=news, broad web finds actual coverage; domain scoping finds the
+# company's own newsroom. Override per run: loop --no-domain-filter / --domain-filter
+EXA_DOMAIN_FILTER=0
+# Hosts to always exclude (comma-separated), e.g. talent.example.com,careers.example.com
+EXA_EXCLUDE_DOMAINS=
--- a/sales-agent/README.md
+++ b/sales-agent/README.md
@@ -20,7 +20,7 @@ and iterate. Deliberate v1 simplifications (we'll revisit each tomorrow):
 | Postgres + Redis + pgvector | **JSON files** under `output/state/` | Hash dedup works; semantic dedup is a documented v2 hook. |
 | Vector ANN capability pre-filter | **Keyword-overlap pre-filter** | Same interface, no embedding service. |
 | Slack Bolt + SendGrid | **Console + markdown files** | Briefs render Slack-style; digest writes to `output/digest.md`. |
-| LLM relevance + synthesis | **Gemini or Anthropic when a key is set, else a deterministic mock reasoner** | Provider-agnostic (`LLM_PROVIDER`); loop runs offline without a key. |
+| LLM relevance + synthesis | **OpenAI-compatible / Gemini / Anthropic** (required) | Provider-agnostic (`LLM_PROVIDER`). **No mock fallback** — on a rate limit it waits and retries; without a key it refuses to run. |

 ## Run it

@@ -75,9 +75,16 @@ LLM_MODEL=claude-sonnet-4-6
 ```

 If `LLM_PROVIDER` is blank it auto-detects from whichever key is present (OpenAI-
-compatible preferred). With no key, a deterministic mock reasoner produces grounded
-briefs so the full pipeline still runs offline. All providers use plain `urllib` —
-no SDKs. LLM calls are batched (one relevance call + one synthesis call per account).
+compatible preferred). All providers use plain `urllib` — no SDKs. LLM calls are
+batched (one relevance call + one synthesis call per account).
+
+**No mock fallback.** Quality is the priority: the loop refuses to run without a
+configured provider, and on HTTP 429 it **waits and retries** (`LLM_RATELIMIT_WAIT_
+SECONDS`, default 900s; honors `Retry-After`) up to `LLM_RATELIMIT_MAX_RETRIES`
+(default 24) rather than emitting low-quality output. If the LLM is ultimately
+unavailable for an account, that account is skipped and its signals are left
+unconsumed (retried next run) — never fabricated. Degenerate/unparseable responses
+(e.g. a model repetition loop) are re-sampled `LLM_CONTENT_RETRIES` times.

 ## Exa.ai news connector & free-plan protection

@@ -100,23 +107,26 @@ request per loop run (so the default 3 sample accounts = 3 requests/run).
 python3 run.py exa-usage      # show remaining monthly budget
 ```

-### Domain scoping (default on)
+### Signal quality: news category & domain scoping

-By default, news search is **scoped to each account's own `domain`** (via Exa
-`includeDomains`) — newsroom, press releases, investor-relations — which cuts out
-stock-ticker and analyst noise from a bare company-name search. Toggle it:
+The single biggest lever on signal quality is **`EXA_CATEGORY=news`** (default): it
+restricts results to news articles and keeps out evergreen pages (careers, marketing,
+generic corporate). The query is also event-focused (acquisition, earnings, product
+launch, leadership change, ...) rather than pulling recruiting pages.
+
+Domain scoping is **off by default** — with `category=news`, broad-web search finds
+actual coverage *about* the account (PRNewswire, CNBC, trade press). Scope to the
+company's own newsroom only if you want first-party press releases:

 ```bash
-python3 run.py loop --no-domain-filter   # widen to broad third-party web coverage
-# or set EXA_DOMAIN_FILTER=0 in .env to default to broad
+python3 run.py loop --domain-filter      # account's own domain (newsroom/press only)
+python3 run.py loop --no-domain-filter   # broad-web news about the account (default)
+# or set EXA_DOMAIN_FILTER / EXA_CATEGORY / EXA_EXCLUDE_DOMAINS in .env
 ```

-Note: domain scoping only returns results when the account's real domain has indexed
-content. The fictional sample accounts return 0 domain-scoped articles by design —
-real target-account domains will return their first-party news.
-
 Tune all limits in `.env` (`EXA_MONTHLY_REQUEST_CAP`, `EXA_PER_RUN_REQUEST_CAP`,
-`EXA_RESULTS_PER_QUERY`, `EXA_LOOKBACK_DAYS`). Set `EXA_ENABLED=0` to force fixtures.
+`EXA_RESULTS_PER_QUERY`, `EXA_LOOKBACK_DAYS`, `EXA_CATEGORY`). Set `EXA_ENABLED=0` to
+force fixtures.

 ## How the loop maps to the design

--- a/sales-agent/run.py
+++ b/sales-agent/run.py
@@ -23,6 +23,8 @@ from src.memory import store, feedback
 def cmd_loop(args):
    if args.no_domain_filter:
        config.EXA_DOMAIN_FILTER = False  # per-run override
+    if args.domain_filter:
+        config.EXA_DOMAIN_FILTER = True
    pipeline.run_all(tier=args.tier, account_id=args.account)
    print("\n--- digest preview ---")
    print(digest_mod.write_digest())
@@ -85,7 +87,9 @@ def main(argv=None):
    p_loop.add_argument("--account", help="single account id")
    p_loop.add_argument("--tier", choices=["strategic", "enterprise", "growth"])
    p_loop.add_argument("--no-domain-filter", action="store_true",
-                        help="include broad third-party news instead of scoping Exa to the account domain")
+                        help="broad-web news about the account (default behavior)")
+    p_loop.add_argument("--domain-filter", action="store_true",
+                        help="scope Exa to the account's own domain (newsroom/press only)")
    p_loop.set_defaults(func=cmd_loop)

    p_dig = sub.add_parser("digest", help="Print the daily intelligence digest")
--- a/sales-agent/src/agents/relevance.py
+++ b/sales-agent/src/agents/relevance.py
@@ -65,33 +65,6 @@ def _to_result(signal: NormalizedSignal, data: dict) -> RelevanceResult | None:
    )


-def _mock_evaluate(signal: NormalizedSignal, candidates: list[dict], account: dict) -> dict:
-    """Deterministic stand-in for the LLM: keyword overlap + seniority heuristics."""
-    if not candidates:
-        return {"relevant": False}
-    top = candidates[0]
-    seniority = (signal.author or {}).get("seniority", "")
-    strength = "strong" if (seniority == "C-suite" or "[" in signal.content[:3]) else "moderate"
-    if signal.source == "jobs" and signal.signal_type == "hiring_surge":
-        strength = "strong"
-    return {
-        "relevant": True,
-        "capabilityIds": [top["id"]],
-        "rationale": (
-            f"Signal references themes central to {top['name']} "
-            f"(e.g. {', '.join(top['keywords'][:3])}). Source={signal.source}, "
-            f"author seniority={seniority or 'n/a'}."
-        ),
-        "signalStrength": strength,
-        "isExpansion": top["id"] in account.get("currentProducts", []),
-        "idealContact": {
-            "title": top["targetBuyer"]["titles"][0],
-            "department": top["targetBuyer"]["departments"][0],
-        },
-        "suggestedAngle": top["businessProblems"][0],
-    }
-
-
 def _signals_block(prepared: list[tuple[NormalizedSignal, list[dict]]]) -> str:
    lines = []
    for i, (s, cands) in enumerate(prepared):
@@ -107,17 +80,9 @@ def _signals_block(prepared: list[tuple[NormalizedSignal, list[dict]]]) -> str:
    return "\n".join(lines)


-def _mock_batch(prepared, account) -> list[tuple[NormalizedSignal, RelevanceResult]]:
-    out = []
-    for s, cands in prepared:
-        r = _to_result(s, _mock_evaluate(s, cands, account))
-        if r:
-            out.append((s, r))
-    return out
-
-
 def _evaluate_batch(prepared, account) -> list[tuple[NormalizedSignal, RelevanceResult]]:
-    """One LLM call for a chunk of (signal, candidates) pairs."""
+    """One LLM call for a chunk of (signal, candidates) pairs. Raises on LLM error
+    (no mock fallback) so the pipeline can skip the account rather than fabricate."""
    cap_by_id: dict[str, dict] = {}
    for _, cands in prepared:
        for c in cands:
@@ -130,11 +95,7 @@ def _evaluate_batch(prepared, account) -> list[tuple[NormalizedSignal, Relevance
        capability_summary=capabilities.library_summary_for_prompt(list(cap_by_id.values())),
        signals_block=_signals_block(prepared),
    )
-    try:
    data = llm.call_json(SYSTEM, user, max_tokens=2200, require_keys={"signals"})
-    except llm.LLMError as e:
-        print(f"   [relevance] batched LLM error, falling back to mock: {e}")
-        return _mock_batch(prepared, account)

    by_index = {}
    for item in data.get("signals", []):
@@ -161,8 +122,7 @@ def filter_for_relevance(
    if not prepared:
        return []

-    runner = _mock_batch if llm.is_mock() else _evaluate_batch
    out: list[tuple[NormalizedSignal, RelevanceResult]] = []
    for start in range(0, len(prepared), BATCH_SIZE):
-        out.extend(runner(prepared[start : start + BATCH_SIZE], account))
+        out.extend(_evaluate_batch(prepared[start : start + BATCH_SIZE], account))
    return out
--- a/sales-agent/src/agents/synthesizer.py
+++ b/sales-agent/src/agents/synthesizer.py
@@ -65,40 +65,9 @@ def _target_contacts(account: dict, ideal: dict) -> list[dict]:
    return contacts


-def _mock_brief_body(cap: dict, items, account) -> dict:
-    strong = [r for _, r in items if r.signal_strength == "strong"]
-    urgency = "high" if strong else ("medium" if len(items) > 1 else "low")
-    confidence = min(0.95, 0.45 + 0.18 * len(strong) + 0.08 * len(items))
-    lead = items[0][0]
-    tps = [
-        f"{cap['name']} directly addresses: {cap['businessProblems'][0]}.",
-        f"Signal evidence: {items[0][1].rationale}",
-    ]
-    if len(items) > 1:
-        tps.append(f"Reinforced by {len(items)-1} additional signal(s) across "
-                   f"{', '.join(sorted({s.source for s, _ in items}))}.")
-    hook_who = (lead.author or {}).get("name", "there")
-    return {
-        "headline": f"{account['accountName']}: {len(items)} signal(s) point to {cap['name']}.",
-        "executiveSummary": (
-            f"{account['accountName']} is showing intent around {cap['name'].lower()}. "
-            f"Lead signal: {lead.signal_type} via {lead.source}. "
-            f"{len(strong)} strong / {len(items)} total contributing signals."
-        ),
-        "talkingPoints": tps,
-        "outreachHook": (
-            f"{hook_who} — saw the recent signal around "
-            f"{cap['keywords'][0]}. We've helped similar {account.get('industry','')} "
-            f"teams here. Worth 20 minutes?"
-        ),
-        "urgencyLevel": urgency,
-        "confidenceScore": round(confidence, 2),
-    }
-
-
 def _batched_bodies(groups: list[tuple[str, dict, list]], account: dict) -> dict[str, dict]:
-    """One LLM call producing a brief body per capability group. Empty dict on error
-    so the caller can fall back to the mock body per group."""
+    """One LLM call producing a brief body per capability group. Raises on LLM error
+    (no mock fallback). Capabilities the model omits are simply not synthesized."""
    blocks = []
    for cap_id, cap, items in groups:
        blocks.append(
@@ -112,11 +81,7 @@ def _batched_bodies(groups: list[tuple[str, dict, list]], account: dict) -> dict
        current_products=", ".join(account.get("currentProducts", [])) or "none",
        capability_blocks="\n\n".join(blocks),
    )
-    try:
    data = llm.call_json(SYSTEM, user, max_tokens=4096, require_keys={"briefs"})
-    except llm.LLMError as e:
-        print(f"   [synthesize] batched LLM error, falling back to mock: {e}")
-        return {}
    return {
        b["capabilityId"]: b
        for b in data.get("briefs", [])
@@ -144,12 +109,14 @@ def synthesize(
    if not groups:
        return []

-    # One LLM call for all of this account's briefs (mock builds them locally).
-    bodies = {} if llm.is_mock() else _batched_bodies(groups, account)
+    # One LLM call for all of this account's briefs (raises on failure -> caller skips).
+    bodies = _batched_bodies(groups, account)

    briefs: list[OpportunityBrief] = []
    for cap_id, cap, items in groups:
-        body = bodies.get(cap_id) or _mock_brief_body(cap, items, account)
+        body = bodies.get(cap_id)
+        if not body:
+            continue  # model didn't return a brief for this capability — skip, don't fabricate

        created = now_utc()
        ideal = items[0][1].ideal_contact
--- a/sales-agent/src/config.py
+++ b/sales-agent/src/config.py
@@ -28,12 +28,13 @@ def _load_dotenv() -> None:
 _load_dotenv()

 # --- LLM ---
-# Provider abstraction: openai | gemini | anthropic | mock. If LLM_PROVIDER is
-# unset we auto-detect from whichever key is present (OpenAI-compatible first),
-# falling back to a deterministic mock reasoner so the loop still runs offline.
+# Provider abstraction: openai | gemini | anthropic. If LLM_PROVIDER is unset we
+# auto-detect from whichever key is present (OpenAI-compatible first). There is NO
+# mock fallback: the loop refuses to run without a configured provider, and on
+# rate limits it WAITS and retries rather than emitting low-quality output.
 #
 # "openai" is any OpenAI-compatible /chat/completions endpoint — OpenAI, OpenRouter,
-# Together, Groq, vLLM, Ollama, LM Studio, etc. — configured by three env vars:
+# Together, Groq, NVIDIA, vLLM, Ollama, LM Studio, etc. — configured by:
 #   OPENAI_BASE_URL  (e.g. https://api.openai.com/v1)
 #   OPENAI_API_KEY
 #   OPENAI_MODEL     (e.g. gpt-4o-mini)
@@ -50,8 +51,13 @@ OPENAI_TEMPERATURE = float(os.environ.get("OPENAI_TEMPERATURE", "0.5"))
 OPENAI_FREQUENCY_PENALTY = float(os.environ.get("OPENAI_FREQUENCY_PENALTY", "0.4"))
 OPENAI_PRESENCE_PENALTY = float(os.environ.get("OPENAI_PRESENCE_PENALTY", "0.0"))
 # Retries when a model returns degenerate/unparseable content (stochastic — a
-# retry usually succeeds). Distinct from the network/429 backoff.
+# retry usually succeeds).
 LLM_CONTENT_RETRIES = int(os.environ.get("LLM_CONTENT_RETRIES", "3"))
+# On HTTP 429 (rate limit) we WAIT this long and retry (no mock fallback). Honors
+# a Retry-After response header when present. Lets the loop grind slowly but
+# correctly through a tight rate limit instead of degrading quality.
+LLM_RATELIMIT_WAIT_SECONDS = int(os.environ.get("LLM_RATELIMIT_WAIT_SECONDS", "900"))
+LLM_RATELIMIT_MAX_RETRIES = int(os.environ.get("LLM_RATELIMIT_MAX_RETRIES", "24"))

 ANTHROPIC_API_KEY = os.environ.get("ANTHROPIC_API_KEY", "").strip()
 LLM_MODEL = os.environ.get("LLM_MODEL", "claude-sonnet-4-6")
@@ -68,24 +74,22 @@ if not _provider:
    elif ANTHROPIC_API_KEY:
        _provider = "anthropic"
    else:
-        _provider = "mock"
-# A forced USE_MOCK_LLM=1, or selecting a provider whose key is missing, => mock.
-if os.environ.get("USE_MOCK_LLM", "").strip() == "1":
-    _provider = "mock"
-elif _provider == "openai" and not OPENAI_API_KEY:
-    _provider = "mock"
+        _provider = "none"
+# Selecting a provider whose key is missing => no usable provider.
+if _provider == "openai" and not OPENAI_API_KEY:
+    _provider = "none"
 elif _provider == "gemini" and not GEMINI_API_KEY:
-    _provider = "mock"
+    _provider = "none"
 elif _provider == "anthropic" and not ANTHROPIC_API_KEY:
-    _provider = "mock"
+    _provider = "none"

 LLM_PROVIDER = _provider
-USE_MOCK_LLM = LLM_PROVIDER == "mock"
+LLM_CONFIGURED = LLM_PROVIDER != "none"


 def llm_label() -> str:
-    if LLM_PROVIDER == "mock":
-        return "MOCK reasoner (no LLM key)"
+    if LLM_PROVIDER == "none":
+        return "NONE (no LLM provider configured)"
    if LLM_PROVIDER == "openai":
        host = OPENAI_BASE_URL.split("//")[-1].split("/")[0]
        return f"OpenAI-compatible ({OPENAI_MODEL} @ {host})"
@@ -111,10 +115,18 @@ EXA_MONTHLY_REQUEST_CAP = int(os.environ.get("EXA_MONTHLY_REQUEST_CAP", "1000"))
 EXA_PER_RUN_REQUEST_CAP = int(os.environ.get("EXA_PER_RUN_REQUEST_CAP", "25"))
 EXA_RESULTS_PER_QUERY = int(os.environ.get("EXA_RESULTS_PER_QUERY", "5"))
 EXA_LOOKBACK_DAYS = int(os.environ.get("EXA_LOOKBACK_DAYS", "45"))
-# Scope news search to each account's own domain (newsroom/press/IR) to cut
-# third-party ticker/analyst noise. Toggle off in .env or per-run with
-# `loop --no-domain-filter` to include broad third-party coverage.
-EXA_DOMAIN_FILTER = os.environ.get("EXA_DOMAIN_FILTER", "1").strip() == "1"
+# Exa content category. "news" restricts results to news articles (press releases,
+# coverage) and keeps out evergreen pages like careers/marketing landing pages —
+# the single biggest signal-quality lever. Set to "" to disable category filtering.
+EXA_CATEGORY = os.environ.get("EXA_CATEGORY", "news").strip()
+# Domain scoping is OFF by default now: with category=news, broad-web search finds
+# actual news ABOUT the account from outlets, whereas scoping to the company's own
+# domain surfaces careers/marketing pages. Turn on for first-party newsroom only.
+EXA_DOMAIN_FILTER = os.environ.get("EXA_DOMAIN_FILTER", "0").strip() == "1"
+# Hosts to always exclude (comma-separated) — e.g. careers/talent subdomains.
+EXA_EXCLUDE_DOMAINS = [
+    d.strip() for d in os.environ.get("EXA_EXCLUDE_DOMAINS", "").split(",") if d.strip()
+]

 # --- Dedup / suppression windows ---
 SEMANTIC_DEDUP_DAYS = 90
--- a/sales-agent/src/connectors/exa.py
+++ b/sales-agent/src/connectors/exa.py
@@ -84,18 +84,18 @@ def _count_request() -> None:


 def _build_query(account: dict, domain_scoped: bool) -> str:
+    # Event-focused: target datable business events, NOT evergreen pages. We avoid
+    # generic words like "hiring"/"careers" that pull recruiting landing pages.
    name = account["accountName"]
+    events = (
+        "acquisition, merger, earnings results, product launch, partnership, "
+        "expansion, restructuring, layoffs, leadership change, executive appointment, "
+        "data center, cloud, modernization, regulatory, strategy"
+    )
    if domain_scoped:
-        # Domain is already restricting results, so keep the query topical/broad.
-        return (
-            f"{name} announcements: funding, acquisition, merger, leadership change, "
-            f"strategy, restructuring, product launch, hiring, earnings, partnership"
-        )
-    # No domain restriction -> anchor harder on the company name to limit drift.
-    return (
-        f'"{name}" company news: funding, acquisition, merger, leadership change, '
-        f"strategy, restructuring, product launch, hiring, earnings, partnership"
-    )
+        return f"{name} news and announcements: {events}"
+    # No domain restriction -> anchor on the exact company name to limit drift.
+    return f'"{name}" news: {events}'


 def harvest_news(account: dict) -> list[RawSignal] | None:
@@ -117,8 +117,12 @@ def harvest_news(account: dict) -> list[RawSignal] | None:
        "startPublishedDate": start,
        "contents": {"text": {"maxCharacters": 1000}},
    }
+    if config.EXA_CATEGORY:
+        payload["category"] = config.EXA_CATEGORY
    if domain_scoped:
        payload["includeDomains"] = [domain]
+    if config.EXA_EXCLUDE_DOMAINS:
+        payload["excludeDomains"] = config.EXA_EXCLUDE_DOMAINS
    req = urllib.request.Request(
        _ENDPOINT,
        data=json.dumps(payload).encode("utf-8"),
@@ -162,7 +166,8 @@ def harvest_news(account: dict) -> list[RawSignal] | None:
            )
        )
    remaining = usage_status()["monthly_remaining"]
+    cat = config.EXA_CATEGORY or "any"
    scope = f"domain={domain}" if domain_scoped else "broad web"
-    print(f"   [exa] {len(signals)} article(s) for {account['accountName']} ({scope}) "
+    print(f"   [exa] {len(signals)} {cat} article(s) for {account['accountName']} ({scope}) "
          f"(monthly budget left: {remaining}/{config.EXA_MONTHLY_REQUEST_CAP})")
    return signals
--- a/sales-agent/src/llm.py
+++ b/sales-agent/src/llm.py
@@ -1,8 +1,10 @@
 """LLM access layer (provider-agnostic).

-Dispatches to a configured provider — Gemini or Anthropic — via plain urllib
-(no SDKs). When no key is configured it falls back to a deterministic mock
-reasoner so the full loop runs offline. All paths return parsed JSON dicts.
+Dispatches to a configured provider — OpenAI-compatible, Gemini, or Anthropic —
+via plain urllib (no SDKs). There is NO mock fallback: a missing provider raises,
+and on HTTP 429 the call WAITS (honoring Retry-After, else a configured interval)
+and retries, so the loop grinds through a rate limit rather than degrading output.
+All paths return parsed JSON dicts.
 """
 from __future__ import annotations

@@ -14,14 +16,15 @@ import urllib.error

 from . import config

-# Backoff on HTTP 429 (rate limit) before giving up and letting callers fall to mock.
-_RETRY_WAITS = [4, 10, 20]
-

 class LLMError(RuntimeError):
    pass


+class LLMNotConfigured(LLMError):
+    """No LLM provider is configured — the loop must not run."""
+
+
 class LLMRetryable(LLMError):
    """A failure worth retrying with a fresh sample (degenerate/unparseable output)."""

@@ -44,9 +47,23 @@ def _extract_json(text: str) -> dict:
        raise LLMRetryable(f"Malformed JSON from LLM ({e}): ...{text[max(0, end-120):end + 1]}")


+def _fmt_secs(s: int) -> str:
+    return f"{s // 60}m{s % 60:02d}s" if s >= 60 else f"{s}s"
+
+
+def _retry_after_seconds(e: "urllib.error.HTTPError") -> int | None:
+    """Parse a Retry-After header (seconds form) if the server provides one."""
+    try:
+        val = e.headers.get("Retry-After") if e.headers else None
+        return int(val) if val and val.strip().isdigit() else None
+    except Exception:
+        return None
+
+
 def _post(url: str, payload: dict, headers: dict, provider: str) -> dict:
    data = json.dumps(payload).encode("utf-8")
-    for attempt in range(len(_RETRY_WAITS) + 1):
+    # +1 for the initial attempt before any rate-limit wait.
+    for attempt in range(config.LLM_RATELIMIT_MAX_RETRIES + 1):
        req = urllib.request.Request(
            url, data=data,
            headers={"content-type": "application/json", **headers}, method="POST",
@@ -56,17 +73,23 @@ def _post(url: str, payload: dict, headers: dict, provider: str) -> dict:
                return json.loads(resp.read().decode("utf-8"))
        except urllib.error.HTTPError as e:
            detail = e.read().decode("utf-8")[:400] if hasattr(e, "read") else ""
-            # Rate limited: wait and retry before surfacing the error.
-            if e.code == 429 and attempt < len(_RETRY_WAITS):
-                wait = _RETRY_WAITS[attempt]
-                print(f"   [llm] {provider} rate-limited (429); backing off {wait}s "
-                      f"(retry {attempt + 1}/{len(_RETRY_WAITS)})")
+            # Rate limited: WAIT (Retry-After if given, else configured interval)
+            # and retry. No mock fallback — we'd rather grind slowly than degrade.
+            if e.code == 429 and attempt < config.LLM_RATELIMIT_MAX_RETRIES:
+                wait = _retry_after_seconds(e) or config.LLM_RATELIMIT_WAIT_SECONDS
+                from .types import iso, now_utc
+                from datetime import timedelta
+                resume = iso(now_utc() + timedelta(seconds=wait))
+                print(f"   [llm] {provider} rate-limited (429). Waiting {_fmt_secs(wait)} "
+                      f"then retrying (attempt {attempt + 1}/{config.LLM_RATELIMIT_MAX_RETRIES}; "
+                      f"resume ~{resume}).")
                time.sleep(wait)
                continue
            raise LLMError(f"{provider} API error {e.code}: {detail}")
        except urllib.error.URLError as e:
            raise LLMError(f"Network error calling {provider} API: {e}")
-    raise LLMError(f"{provider} API still rate-limited after retries")
+    raise LLMError(f"{provider} API still rate-limited after "
+                   f"{config.LLM_RATELIMIT_MAX_RETRIES} waits")


 def _call_openai(system: str, user: str, max_tokens: int) -> dict:
@@ -163,10 +186,13 @@ def call_json(
    network/HTTP errors and 429s are handled in _post. If `require_keys` is given,
    a response missing all of those top-level keys is treated as degenerate and
    retried (catches valid-but-empty junk like `{}` from a repetition loop).
-    After retries are exhausted the error propagates and callers fall back to mock.
+    After content retries are exhausted the error propagates (no mock fallback).
    """
-    if config.USE_MOCK_LLM:
-        raise _MockSignal()  # callers catch this and run their mock path
+    if not config.LLM_CONFIGURED:
+        raise LLMNotConfigured(
+            "No LLM provider configured. Set OPENAI_API_KEY (and OPENAI_BASE_URL/"
+            "OPENAI_MODEL), or GEMINI_API_KEY, or ANTHROPIC_API_KEY in .env."
+        )
    attempts = max(1, config.LLM_CONTENT_RETRIES + 1)
    for i in range(attempts):
        try:
@@ -183,11 +209,3 @@ def call_json(
                      f"({i + 1}/{attempts - 1}): {str(e)[:100]}")
                continue
            raise
-
-
-class _MockSignal(Exception):
-    """Internal sentinel: tells the caller to use its deterministic mock branch."""
-
-
-def is_mock() -> bool:
-    return config.USE_MOCK_LLM
--- a/sales-agent/src/pipeline.py
+++ b/sales-agent/src/pipeline.py
@@ -1,7 +1,7 @@
 """The account cycle — the closed loop, run per account (design section 7.3)."""
 from __future__ import annotations

-from . import config
+from . import config, llm
 from .agents import harvester, normalizer, relevance, synthesizer, scorer
 from .delivery import render
 from .connectors import crm, exa
@@ -27,18 +27,25 @@ def run_account_cycle(account: dict, verbose: bool = True) -> list[dict]:
        store.save_memory(mem)
        return []

-    # 3. Relevance filter (LLM reasoning vs. capability library)
+    # 3 + 4. Relevance filter + synthesis (LLM). On LLM failure we skip the account
+    # and do NOT consume the signals, so they are retried on the next run (rather
+    # than emitting low-quality output). 429s are waited out inside the LLM layer.
+    try:
        relevant = relevance.filter_for_relevance(signals, account)
        if verbose:
            print(f"   {len(relevant)} relevant signal(s)")
-    # Record all processed signals so we don't re-evaluate them next run.
-    store.record_signals(mem, signals)
        if not relevant:
+            store.record_signals(mem, signals)  # evaluated, found irrelevant -> mark seen
            store.save_memory(mem)
            return []
-
-    # 4. Synthesize into opportunity briefs
        briefs = synthesizer.synthesize(relevant, account)
+    except llm.LLMError as e:
+        print(f"   ! LLM unavailable for {account['accountName']}: {str(e)[:160]}")
+        print("     skipping account — signals NOT consumed, will retry next run.")
+        return []
+
+    # Signals fully processed by the LLM -> mark seen so we don't re-evaluate them.
+    store.record_signals(mem, signals)

    # 5. Score + prioritize
    briefs = scorer.score_and_prioritize(briefs, account)
@@ -75,6 +82,11 @@ def run_account_cycle(account: dict, verbose: bool = True) -> list[dict]:

 def run_all(tier: str | None = None, account_id: str | None = None) -> list[dict]:
    config.ensure_dirs()
+    if not config.LLM_CONFIGURED:
+        print("ERROR: no LLM provider configured. Set OPENAI_API_KEY (+ OPENAI_BASE_URL/"
+              "OPENAI_MODEL), GEMINI_API_KEY, or ANTHROPIC_API_KEY in sales-agent/.env.")
+        print("Refusing to run without a reasoning model (no mock fallback).")
+        return []
    accounts = crm.load_accounts(tier=tier, account_id=account_id)
    if not accounts:
        print("No matching accounts.")