Module agent

Modules§

agent
classifier
context_analyzer
cost
dispatcher
eval
history
history_pruner
history_trim: Whole-turn history trimming. One rule: keep the most recent whole turns that fit the token budget, drop the rest, never cut a turn in half.
loop_
loop_detector: Loop detection guardrail for the agent tool-call loop.
memory_inject: Unified memory-context injection: ONE policy and ONE renderer for the memory preamble, applied at the turn engine.
memory_strategy
personality: Personality system — loads workspace identity files (SOUL.md, IDENTITY.md, USER.md) and injects them into the system prompt pipeline.
personality_templates: Default starter templates for the per-workspace personality files.
pricing_catalog: Global pricing catalog: a last-resort fallback consulted by the cost engine when a model has no per-provider pricing entry in config.
prompt
system_prompt: System prompt construction for the agent loop and channel subsystem. These functions were originally in channels/mod.rs but live here to break a circular dependency between the channels and agent modules.
thinking: Thinking/Reasoning Level Control
tool_execution: Tool execution helpers extracted from loop_.
tool_receipts: HMAC-SHA256 tool execution receipts for hallucination detection.

Agent
AgentAttribution: Borrow-only Attributable holding an agent alias. Used by entry points (loop_::run, process_message, cron dispatch) that don’t construct a full Agent but still need to open an attribution_span! carrying the agent’s role + alias.
AgentBuilder
TurnMeta: Lightweight metadata for turn-level event emission. Built at emission call sites from the turn’s borrows — not a cached duplicate. The values are borrows from the turn mint site; the mint site stays the single source of truth for the (channel, agent_alias, turn_id) triple.

is_credential_key: Whether a JSON key denotes a credential value on a human-facing surface.
is_semantic_empty_terminal_completion: Whether a turn error has the canonical semantic-empty terminal reason.
process_message: Process a single message through the full agent (with tools, peripherals, memory). Used by channels (Telegram, Discord, etc.) to enable hardware and tool use.
run
scrub_credentials_value: Structured-aware credential scrub for a JSON value bound for a human-facing surface. Object entries whose key names a credential have their entire value replaced with a marker, preserving the key. Every other string leaf still runs through the text scrub_credentials so inline token=... patterns inside unrelated fields are caught too. Serialize-then-scrub would corrupt key names that merely contain a sensitive word (e.g. access_token), so this walks the value instead. Same rendering-boundary contract as scrub_credentials.
semantic_empty_terminal_completion_message: Render the canonical user-facing failure at a delivery boundary.
terminal_completion_error_message: Map typed terminal-delivery failures to their Fluent user-facing message.