Capability-set targeting + Claude Code headless provider + resilience fixes

Capability targeting (new feature):
- loop --capability <SET|ALL> scopes which capability library the loop reasons
  against. data/capability_sets/<name>.json (same schema as capabilities.json)
  holds one-off targeted products; drop a file in and pass its name, no code
  changes. First set: agentminder (Broadcom AgentMinder, agentic-AI identity).
- capabilities.set_active_set() swaps the active library; keyword pre-filter is
  bypassed for small targeted sets so every signal is evaluated. New
  'capability-sets' command lists them; run banner shows the active set.

LLM provider — the '429' was a disguised policy block:
- Anthropic rejects Claude Code OAuth tokens on /v1/messages for non-Claude-Code
  traffic with a fake rate_limit_error (no anthropic-ratelimit-* headers). New
  'claude-cli' provider runs inference through the sanctioned headless path
  (claude -p --append-system-prompt --model), now the default provider.
- OAuth keychain token auto-refresh via headless claude when expired.

Resilience:
- STATE_DIR is overridable (macOS endpoint-security locked output/state via
  com.apple.macl); load_memory tolerates PermissionError; exa usage counter
  tolerates malformed files; run.py handles Ctrl-C cleanly (no traceback).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

sales-agent/USAGE.md

@@ -25,6 +25,7 @@ python3 run.py loop --no-domain-filter       # Exa: broad-web news (default)
 | `--tier T` | Only `strategic` / `enterprise` / `growth` accounts |
 | `--rep NAME` | Only accounts assigned to a rep (case-insensitive name or `rep_id` substring) |
 | `--limit N` | Process at most N accounts. Order: **strategic tier first, then least-recently-scanned** (never-scanned first). The run cursor (`output/state/run_cursor.json`) persists, so successive `--limit` runs walk the whole territory — e.g. a nightly `loop --limit 30` covers 342 accounts in ~12 days. |
+| `--capability SET` | Capability set to reason against: `ALL` (full 40-capability portfolio, default) or a named set from `data/capability_sets/` (e.g. `agentminder`). In a targeted set the keyword pre-filter is bypassed so EVERY signal is evaluated against the chosen product. List sets: `python3 run.py capability-sets`. **Add a new one-off product**: drop `data/capability_sets/<name>.json` (same schema as `capabilities.json`) and pass `--capability <name>` — no code changes. |
 | `--domain-filter` | Scope Exa news to the account's own domain (newsroom/PR only) |
 | `--no-domain-filter` | Broad-web news about the account (default) |
 

sales-agent/data/capability_sets/agentminder.json

@@ -0,0 +1,50 @@
+{
+  "version": "1.0",
+  "lastUpdated": "2026-07-16",
+  "_source": "https://www.broadcom.com/products/identity/agentminder",
+  "_note": "One-off targeted capability set. Select at runtime: `loop --capability agentminder`. Add more sets as data/capability_sets/<name>.json with this same schema.",
+  "capabilities": [
+    {
+      "id": "cap_iam_agentminder",
+      "grouping": "Cybersecurity — Identity & Access Management",
+      "category": "Agentic AI Security",
+      "name": "AgentMinder — Secure Agentic AI (Identity & Governance for AI Agents)",
+      "description": "Agentic security and governance platform that turns autonomous AI agents into production-grade 'digital employees': governs AI agents as first-class identities with least-privilege access, credential protection, centralized authentication/authorization, and lifecycle governance, on a cloud-native container-based architecture.",
+      "businessProblems": [
+        "AI agents/copilots being deployed with no identity, credential, or access governance",
+        "Non-human identities (service accounts, bots, AI agents) proliferating outside IAM controls",
+        "No least-privilege or audit trail for what autonomous agents can access and do",
+        "Compliance/regulatory exposure from ungoverned agentic AI acting on production systems",
+        "Security teams blocking AI initiatives because agent access cannot be controlled"
+      ],
+      "buyingSignals": [
+        "Announcements of agentic AI, AI agents, or copilot deployments in production",
+        "AI transformation or GenAI initiatives touching core business systems",
+        "Hiring for AI security, AI governance, ML platform, or non-human identity roles",
+        "Executive posts about AI agents as 'digital workers' or 'digital employees'",
+        "AI governance, responsible-AI, or AI-risk programs and committees being stood up",
+        "Security incidents, breaches, or audit findings involving automation or bots",
+        "Regulatory or compliance pressure on AI usage (finance, healthcare, government)",
+        "Partnerships with AI platform vendors (OpenAI, Anthropic, Microsoft Copilot, agent frameworks)"
+      ],
+      "antiSignals": [
+        "Already owns AgentMinder per CRM",
+        "Standardized on a competing agentic-AI security platform",
+        "Explicit company policy prohibiting AI agent deployment"
+      ],
+      "targetBuyer": {
+        "titles": ["CISO", "Chief AI Officer", "VP Security", "Director IAM", "VP AI/ML Platform", "CTO"],
+        "departments": ["Security", "IT", "AI/Data", "Risk & Compliance"]
+      },
+      "keywords": [
+        "agentminder", "agentic ai", "ai agent", "ai agents", "autonomous agent",
+        "digital employee", "digital worker", "copilot", "non-human identity", "nhi",
+        "machine identity", "ai governance", "ai security", "least privilege",
+        "agent identity", "genai", "generative ai", "ai risk", "responsible ai",
+        "service account", "credential", "zero trust"
+      ]
+    }
+  ],
+  "bundleSignals": [],
+  "globalAntiSignals": []
+}

sales-agent/run.py

@@ -22,6 +22,8 @@ from src.memory import store, feedback
 
 
 def cmd_loop(args):
+    from src import capabilities
+    capabilities.set_active_set(args.capability)
     if args.no_domain_filter:
         config.EXA_DOMAIN_FILTER = False  # per-run override
     if args.domain_filter:
@@ -95,6 +97,19 @@ def cmd_recalibrate(args):
         print(f"  {k}: weight={v['weight']} (n={v['samples']})")
 
 
+def cmd_capability_sets(args):
+    from src import capabilities
+    import json as _json
+    for name in capabilities.available_sets():
+        if name == "ALL":
+            n = len(_json.loads((config.DATA_DIR / "capabilities.json").read_text())["capabilities"])
+            print(f"  ALL          {n} capabilities (full portfolio, default)")
+        else:
+            doc = _json.loads((config.DATA_DIR / "capability_sets" / f"{name}.json").read_text())
+            caps = ", ".join(c["name"].split(" — ")[0] for c in doc["capabilities"])
+            print(f"  {name:12} {len(doc['capabilities'])} capability(ies): {caps}")
+
+
 def cmd_accounts(args):
     for a in crm.load_accounts():
         prods = a.get("currentProducts") or []
@@ -137,6 +152,9 @@ def main(argv=None):
     p_loop.add_argument("--limit", type=int,
                         help="process at most N accounts (strategic + least-recently-run first); "
                              "successive runs resume where the last left off")
+    p_loop.add_argument("--capability", default="ALL", metavar="SET",
+                        help="capability set to reason against: ALL (full portfolio, default) "
+                             "or a name from data/capability_sets/ (e.g. agentminder)")
     p_loop.add_argument("--no-domain-filter", action="store_true",
                         help="broad-web news about the account (default behavior)")
     p_loop.add_argument("--domain-filter", action="store_true",
@@ -168,13 +186,20 @@ def main(argv=None):
     p_re = sub.add_parser("recalibrate", help="Recompute feedback weights (weekly job)")
     p_re.set_defaults(func=cmd_recalibrate)
 
+    sub.add_parser("capability-sets", help="List available capability sets"
+                   ).set_defaults(func=cmd_capability_sets)
     sub.add_parser("accounts", help="List target accounts").set_defaults(func=cmd_accounts)
     sub.add_parser("briefs", help="List all generated briefs").set_defaults(func=cmd_briefs)
     sub.add_parser("exa-usage", help="Show Exa.ai request budget").set_defaults(func=cmd_exa_usage)
 
     args = parser.parse_args(argv)
     config.ensure_dirs()
+    try:
         args.func(args)
+    except KeyboardInterrupt:
+        print("\nInterrupted — no partial briefs were written; unprocessed "
+              "signals will be retried on the next run.")
+        return 130
 
 
 if __name__ == "__main__":

sales-agent/src/capabilities.py

@@ -3,6 +3,12 @@
 The design uses a vector ANN pre-filter to cut LLM calls ~60%. For the MVP we
 approximate that with keyword overlap scoring (no embedding service required).
 The interface is the same: given a signal, return candidate capabilities.
+
+CAPABILITY SETS: the loop can target a subset of the portfolio at runtime.
+- "ALL" (default): data/capabilities.json — the full portfolio.
+- "<name>":        data/capability_sets/<name>.json — a one-off targeted set
+  (same schema). Add new products by dropping another file in that folder and
+  passing `loop --capability <name>`.
 """
 from __future__ import annotations
 
@@ -12,10 +18,43 @@ from functools import lru_cache
 
 from . import config
 
+# Selected via set_active_set() before the loop runs; "ALL" = full portfolio.
+_active_set = "ALL"
+
+
+def available_sets() -> list[str]:
+    sets_dir = config.DATA_DIR / "capability_sets"
+    names = sorted(p.stem for p in sets_dir.glob("*.json")) if sets_dir.exists() else []
+    return ["ALL"] + names
+
+
+def set_active_set(name: str) -> None:
+    """Select which capability library the loop reasons against."""
+    global _active_set
+    name = (name or "ALL").strip()
+    if name.upper() == "ALL":
+        _active_set = "ALL"
+    else:
+        path = config.DATA_DIR / "capability_sets" / f"{name}.json"
+        if not path.exists():
+            raise SystemExit(
+                f"Unknown capability set '{name}'. Available: {', '.join(available_sets())}"
+            )
+        _active_set = name
+    load_library.cache_clear()
+
+
+def active_set() -> str:
+    return _active_set
+
 
 @lru_cache(maxsize=1)
 def load_library() -> dict:
+    if _active_set == "ALL":
         return json.loads((config.DATA_DIR / "capabilities.json").read_text())
+    return json.loads(
+        (config.DATA_DIR / "capability_sets" / f"{_active_set}.json").read_text()
+    )
 
 
 def capability_by_id(cap_id: str) -> dict | None:
@@ -33,7 +72,13 @@ def prefilter(signal_content: str, top_k: int = 3) -> list[dict]:
     """Return capabilities whose keywords overlap the signal, best first.
 
     Stand-in for the vector ANN pre-filter in the design (section 5.2).
+    In a targeted set (<=5 capabilities) the pre-filter is skipped: the whole
+    point of a targeted run is to evaluate EVERY signal against the chosen
+    product, and the LLM prompt stays small anyway.
     """
+    caps = load_library()["capabilities"]
+    if _active_set != "ALL" and len(caps) <= 5:
+        return caps
     sig_tokens = _tokens(signal_content)
     scored: list[tuple[float, dict]] = []
     for cap in load_library()["capabilities"]:

sales-agent/src/config.py

@@ -6,9 +6,6 @@ from pathlib import Path
 
 ROOT = Path(__file__).resolve().parent.parent
 DATA_DIR = ROOT / "data"
-OUTPUT_DIR = ROOT / "output"
-STATE_DIR = OUTPUT_DIR / "state"
-BRIEFS_DIR = OUTPUT_DIR / "briefs"
 
 
 def _load_dotenv() -> None:
@@ -27,6 +24,15 @@ def _load_dotenv() -> None:
 
 _load_dotenv()
 
+# Paths (after dotenv so OUTPUT_DIR/STATE_DIR overrides in .env apply).
+# STATE_DIR is overridable because macOS TCC/endpoint-security software can
+# permanently lock a directory tree (com.apple.macl xattr) — pointing at a
+# fresh dir is the recovery path (state is regenerable). Relative values are
+# resolved against the sales-agent root.
+OUTPUT_DIR = ROOT / os.environ.get("OUTPUT_DIR", "output")
+STATE_DIR = (ROOT / os.environ.get("STATE_DIR", "")) if os.environ.get("STATE_DIR") else OUTPUT_DIR / "state"
+BRIEFS_DIR = OUTPUT_DIR / "briefs"
+
 # --- LLM ---
 # Provider abstraction: openai | gemini | anthropic. If LLM_PROVIDER is unset we
 # auto-detect from whichever key is present (OpenAI-compatible first). There is NO
@@ -66,6 +72,16 @@ LLM_MODEL = os.environ.get("LLM_MODEL", "claude-sonnet-5")
 # keychain ("Claude Code-credentials"), read at runtime so refreshes are picked up.
 ANTHROPIC_OAUTH_TOKEN = os.environ.get("ANTHROPIC_OAUTH_TOKEN", "").strip()
 ANTHROPIC_USE_KEYCHAIN = os.environ.get("ANTHROPIC_USE_KEYCHAIN", "1").strip() == "1"
+# When the keychain token is expired, run a headless `claude` command to make
+# Claude Code refresh it, then re-read — the loop self-heals with no human step.
+ANTHROPIC_AUTOREFRESH = os.environ.get("ANTHROPIC_AUTOREFRESH", "1").strip() == "1"
+
+# "claude-cli" provider: run inference through Claude Code's official headless
+# mode (`claude -p`). This is the sanctioned way to use a Claude subscription
+# programmatically — direct API calls with the Claude Code OAuth token are
+# policy-blocked (they return a fake 429 "Error"). Requires `claude` on PATH.
+CLAUDE_CLI_MODEL = os.environ.get("CLAUDE_CLI_MODEL", "sonnet").strip()
+CLAUDE_CLI_TIMEOUT = int(os.environ.get("CLAUDE_CLI_TIMEOUT", "300"))
 # Effort level for reasoning models (Sonnet 5 etc.): low | medium | high | max.
 ANTHROPIC_EFFORT = os.environ.get("ANTHROPIC_EFFORT", "low").strip()
 
@@ -74,14 +90,14 @@ GEMINI_MODEL = os.environ.get("GEMINI_MODEL", "gemini-2.5-flash")
 
 _provider = os.environ.get("LLM_PROVIDER", "").strip().lower()
 if not _provider:
-    if ANTHROPIC_OAUTH_TOKEN or ANTHROPIC_API_KEY:
+    if ANTHROPIC_API_KEY:
         _provider = "anthropic"
     elif OPENAI_API_KEY:
         _provider = "openai"
     elif GEMINI_API_KEY:
         _provider = "gemini"
     else:
-        _provider = "none"
+        _provider = "claude-cli"  # subscription via Claude Code headless mode
 # Selecting a provider whose key is missing => no usable provider.
 # ("anthropic" is exempt here: the token may come from the keychain at runtime.)
 if _provider == "openai" and not OPENAI_API_KEY:
@@ -100,6 +116,8 @@ LLM_CONFIGURED = LLM_PROVIDER != "none"
 def llm_label() -> str:
     if LLM_PROVIDER == "none":
         return "NONE (no LLM provider configured)"
+    if LLM_PROVIDER == "claude-cli":
+        return f"Claude Code headless ({CLAUDE_CLI_MODEL}, subscription)"
     if LLM_PROVIDER == "openai":
         host = OPENAI_BASE_URL.split("//")[-1].split("/")[0]
         return f"OpenAI-compatible ({OPENAI_MODEL} @ {host})"

sales-agent/src/connectors/exa.py

@@ -39,11 +39,14 @@ def _current_month() -> str:
 def _load_usage() -> dict:
     config.ensure_dirs()
     p = _usage_path()
+    try:
         if p.exists():
             data = json.loads(p.read_text())
-        if data.get("month") == _current_month():
+            if data.get("month") == _current_month() and isinstance(data.get("count"), int):
                 return data
-    # New file or a new month -> reset the counter.
+    except (OSError, ValueError):
+        pass  # unreadable/corrupt file -> reset below
+    # New file, new month, or malformed -> reset the counter.
     return {"month": _current_month(), "count": 0}
 
 

sales-agent/src/llm.py

@@ -164,9 +164,36 @@ def _call_gemini(system: str, user: str, max_tokens: int) -> dict:
 _keychain_token_cache: dict = {}
 
 
+def _read_keychain_cred() -> dict:
+    import subprocess
+    raw = subprocess.run(
+        ["security", "find-generic-password", "-s", "Claude Code-credentials", "-w"],
+        capture_output=True, text=True, timeout=10,
+    ).stdout.strip()
+    return json.loads(raw).get("claudeAiOauth", {})
+
+
+def _refresh_via_claude_cli() -> bool:
+    """Make Claude Code refresh its own OAuth token by running a tiny headless
+    command. Returns True if the CLI ran (the keychain should then be fresh)."""
+    import subprocess
+    try:
+        print("   [llm] keychain OAuth token expired — auto-refreshing via `claude -p` ...")
+        subprocess.run(
+            ["claude", "-p", "reply: ok", "--model", "haiku"],
+            capture_output=True, text=True, timeout=120,
+        )
+        return True
+    except Exception as e:
+        print(f"   [llm] auto-refresh failed: {str(e)[:80]}")
+        return False
+
+
 def _anthropic_oauth_token() -> str:
     """Resolve an OAuth access token: env var first, then the Claude Code
-    credential stored in the macOS keychain (kept fresh by Claude Code)."""
+    credential in the macOS keychain. If the keychain token is expired and
+    ANTHROPIC_AUTOREFRESH=1 (default), a headless `claude` call refreshes it
+    automatically — the loop self-heals without human intervention."""
     if config.ANTHROPIC_OAUTH_TOKEN:
         return config.ANTHROPIC_OAUTH_TOKEN
     if not config.ANTHROPIC_USE_KEYCHAIN:
@@ -176,20 +203,19 @@ def _anthropic_oauth_token() -> str:
     if _keychain_token_cache.get("expires", 0) > now:
         return _keychain_token_cache["token"]
     try:
-        import subprocess
+        cred = _read_keychain_cred()
-        raw = subprocess.run(
+        expired = cred.get("accessToken") and cred.get("expiresAt", 0) / 1000 < now
-            ["security", "find-generic-password", "-s", "Claude Code-credentials", "-w"],
+        if expired and config.ANTHROPIC_AUTOREFRESH and _refresh_via_claude_cli():
-            capture_output=True, text=True, timeout=10,
+            cred = _read_keychain_cred()  # re-read after refresh
-        ).stdout.strip()
+            expired = cred.get("accessToken") and cred.get("expiresAt", 0) / 1000 < time.time()
-        cred = json.loads(raw).get("claudeAiOauth", {})
         token = cred.get("accessToken", "")
-        exp_ms = cred.get("expiresAt", 0)
+        if token and expired:
-        if token and exp_ms and exp_ms / 1000 < now:
             raise LLMNotConfigured(
                 "The Claude Code OAuth token in the keychain is EXPIRED "
-                f"(expired {time.strftime('%Y-%m-%d %H:%M', time.localtime(exp_ms/1000))}). "
+                f"(expired {time.strftime('%Y-%m-%d %H:%M', time.localtime(cred.get('expiresAt', 0)/1000))}) "
-                "Run any command in the Claude Code CLI (e.g. `claude -p hi`) to refresh it, "
+                "and auto-refresh did not produce a fresh one. Run `claude -p hi` manually, "
-                "or set ANTHROPIC_OAUTH_TOKEN / ANTHROPIC_API_KEY in .env."
+                "or set ANTHROPIC_OAUTH_TOKEN (long-lived token from `claude setup-token`) "
+                "or ANTHROPIC_API_KEY in .env."
             )
         if token:
             _keychain_token_cache.update({"token": token, "expires": now + 60})
@@ -230,7 +256,38 @@ def _call_anthropic(system: str, user: str, max_tokens: int) -> dict:
     return _extract_json(text)
 
 
+def _call_claude_cli(system: str, user: str, max_tokens: int) -> dict:
+    """Inference through Claude Code's official headless mode (`claude -p`).
+
+    This is the sanctioned way to run subscription inference programmatically:
+    direct /v1/messages calls with a Claude Code OAuth token are policy-blocked
+    (the API returns a disguised 429 'Error' for non-Claude-Code traffic).
+    """
+    import subprocess
+    cmd = [
+        "claude", "-p", user,
+        "--append-system-prompt", system + "\nReturn ONLY a JSON object. No prose, no markdown fences.",
+        "--model", config.CLAUDE_CLI_MODEL,
+    ]
+    try:
+        proc = subprocess.run(cmd, capture_output=True, text=True,
+                              timeout=config.CLAUDE_CLI_TIMEOUT)
+    except FileNotFoundError:
+        raise LLMNotConfigured("`claude` CLI not found on PATH (needed for LLM_PROVIDER=claude-cli).")
+    except subprocess.TimeoutExpired:
+        raise LLMRetryable(f"claude CLI timed out after {config.CLAUDE_CLI_TIMEOUT}s")
+    if proc.returncode != 0:
+        err = (proc.stderr or proc.stdout or "").strip()[:300]
+        # Usage-limit messages from the CLI are worth waiting out like a 429.
+        if "limit" in err.lower():
+            raise LLMRetryable(f"claude CLI usage limit: {err}")
+        raise LLMError(f"claude CLI failed (rc={proc.returncode}): {err}")
+    return _extract_json(proc.stdout.strip())
+
+
 def _dispatch(system: str, user: str, max_tokens: int) -> dict:
+    if config.LLM_PROVIDER == "claude-cli":
+        return _call_claude_cli(system, user, max_tokens)
     if config.LLM_PROVIDER == "openai":
         return _call_openai(system, user, max_tokens)
     if config.LLM_PROVIDER == "gemini":

sales-agent/src/memory/store.py

@@ -19,8 +19,13 @@ def _account_path(account_id: str) -> Path:
 def load_memory(account_id: str) -> dict:
     config.ensure_dirs()
     p = _account_path(account_id)
+    try:
         if p.exists():
             return json.loads(p.read_text())
+    except (PermissionError, OSError) as e:
+        # macOS TCC/endpoint security can lock individual files (com.apple.macl);
+        # start fresh rather than killing the whole run — memory is regenerable.
+        print(f"   [memory] cannot read {p.name} ({e.__class__.__name__}) — starting fresh")
     return {
         "accountId": account_id,
         "signalHashes": [],          # [{hash, discoveredAt, signalType}]

sales-agent/src/pipeline.py

@@ -161,8 +161,12 @@ def run_all(
     total = len(accounts)
     if limit and limit > 0:
         accounts = accounts[:limit]
+    from . import capabilities
+    cap_set = capabilities.active_set()
+    n_caps = len(capabilities.load_library()["capabilities"])
     print(f"Running prospecting loop over {len(accounts)}/{total} account(s) "
           f"(strategic + least-recently-run first) — LLM: {config.llm_label()}")
+    print(f"Capability set: {cap_set} ({n_caps} capabilit{'y' if n_caps == 1 else 'ies'})")
     if config.EXA_ENABLED:
         st = exa.usage_status()
         scope = "domain-scoped" if config.EXA_DOMAIN_FILTER else "broad web"