Hooks Reference

🔵 Hooks intercept the agent lifecycle. These are YAML files in config/agents/<agent>/hooks/, connected via hook_files: (or by name in hooks:). The kind: field selects the hook. Below is a summary table of all hooks with links to sections.

Hook Types

kind	Purpose	Section
`faq_cache`	cache (question → answer) in DuckDB; serve repeats, skipping the LLM	Context and Memory
`inject`	declaratively inject system/user messages	Context and Memory
`escalate`	escalation to a human when the agent cannot help	Routing and Reliability
`memory`	long-term per-user memory (facts + summary)	Context and Memory
`rag`	retrieval from the archive: injects relevant chunks into context (BM25, no embeddings)	Context and Memory
`tool_cache`	cache of read-only tool results by `(scope, tool, args)`; short-circuit repeats	Context and Memory
`link_check`	output guardrail: checks the final answer's references for groundedness (against the archive) and structural safety; opt-in `live` HTTP	Security and Quality
`condense`	condensation of follow-ups into a standalone query using history; full audit	Context and Memory
`compact_results`	compression of long tool outputs (head/tail + preserving references)	Context and Memory
`context_summary`	LLM summarization of evicted history in the context window (replaces crude concatenation)	Context and Memory
`format_guard`	deterministic answer formatting: bare URL → links, dedup, one link per line	Security and Quality
`schema_guard`	validation of the answer or tool arguments against JSON Schema; `notice` / `rework` / `deny`	Security and Quality
`language_guard`	answer language: capture the user's language, optional directive, `notice`/`rework` on mismatch	Security and Quality
`guard`	stateless DLP membrane: entity detection (email/phone/cards/INN/SNILS/keys…) and action — `redact`/`deny`/`suppress`/`notice`/`rework` on input, output, and tool boundaries	Security and Quality
`pseudonymize`	reversible pseudonymization: surrogates `[EMAIL_1]` to the model, local DuckDB vault, detokenization of the answer/tool arguments	Security and Quality
`loop_guard`	breaking tool-call loops: `deny` repeats beyond the threshold	Routing and Reliability
`groundedness`	output guardrail: checks the final answer's claims for confirmation by the RAG archive; `notice`/`rework`	Security and Quality
`circuit_breaker`	config for the native provider circuit breaker (per-agent): skip a failed provider, falling back to failover	Routing and Reliability
`model_router`	classifier-based model routing (ONNX/LLM/hybrid): choose `model_config` per message	Routing and Reliability
`intent_router`	classifier-based triage: `reply`/`inject`/`notice`/`escalate` by intent	Routing and Reliability
`lexicon`	self-learning query normalizer: slots + gap + golden-gated auto-training (YAML packs + DuckDB overlay)	Context and Memory
`llm_cache`	native inference cache: an exact prompt repeat is served from DuckDB without the provider (turn loop, `llm(...)`, `ask` judges)	Routing and Reliability
`script_hook`	scripted hook in Rhai/JavaScript (see `07-scripts`)	Extension
`agent_events`	native agent action log: writes an event at selected lifecycle points; the `GET /v1/events` route exposes it externally with long-poll and cursor drain	Extension
`dll_hook`	hook from a native Rust DLL (C ABI, see `20-native-plugins`)	Extension

Sections

Fundamentals — lifecycle points, composition, HookContext, local testing.
Context and Memory — faq_cache, inject, memory, rag, tool_cache, condense, compact_results, context_summary, lexicon.
Security and Quality — guard, pseudonymize, link_check, groundedness, language_guard, format_guard, schema_guard.
Routing and Reliability — escalate, loop_guard, circuit_breaker, model_router, intent_router, llm_cache.
Extension — script_hook, dll_hook, native agent_events (agent action log).