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>
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>
- New provider 'openai' calls any OpenAI-compatible /chat/completions endpoint
(OpenAI, OpenRouter, Together, Groq, NVIDIA, vLLM, Ollama, ...), set via
OPENAI_BASE_URL / OPENAI_API_KEY / OPENAI_MODEL. Now the default; auto-detects.
- Robustness for models that degenerate (e.g. Kimi repetition loops): anti-
repetition sampling (temperature + frequency penalty), and a schema-aware
retry — call_json(require_keys=...) retries a fresh sample on degenerate or
unparseable output (LLMRetryable). Relevance requires 'signals', synth 'briefs'.
- response_format=json_object (toggle via OPENAI_JSON_MODE); base URL accepts a
full /chat/completions URL or a /v1 base.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
