Construct an EV_PACKET envelope that mimics what build_round_envelope would have produced after an LLM turn.

Append an event to the log and wake the broadcaster. Delivery is cursor-based off the log (see _broadcast_loop), so every client — including o

Build the exact graph from §6.2 of the design doc.

Authorize (if needed) and execute an Arcade tool, returning the output.

Canned result for an investigative tool call on ``arg`` (host/pool/disk).

Push pipeline state to /logs so the frontend updates in real time.

Convenience: build a ToolCallEvent with JSON-encoded args.

Convert a state to a byte-stable dict for equality comparison.

Emit directly to the queue (bypasses filters).

Save agent output files and append entries to references.md.

First EV_PACKET in the WAL whose routing matched ``tool_name`` — return its parsed tool arguments (enriched by the adapter), or {}.

Authorize (if needed) and execute an Arcade tool, returning the output.

Push gpt-researcher cost delta to /logs.

Run a single agent conversation and return extracted JSON.

Common question answer function that updates context variables and routes based on the answer confidence. agent_level will be one of "basic", "in

Slim clean for list/search: strip body from each email, keep metadata only.

Reset to a clean slate — used by 'Start Simulation' so every run starts from scratch. Cancels in-flight flows, clears tickets/tracking, resets the

Parse HTML and extract plain text.

Heuristic: does the fact contain a timestamp inside the murder window?

Shows basic usage of the Drive v3 API. Prints the names and ids of the first 10 files the user has access to.

Return a ModelConfig for the given role ('lead' or 'researcher'). Provider is selected by LLM_PROVIDER env var (default: gemini). Model names

Extracts and decodes the email body, preferring 'text/plain' but falling back to 'text/html' if needed. Also extracts attachment filenames.

Public info the Detective can see (no private truths).

Marks an email as read by removing the 'UNREAD' label.

Interactive Q&A over an existing due diligence report folder.

Called by gpt-researcher internals. Filters per-call cost messages.

While the sim is live, emit one random healthy log line ~every 1.5s.

Single delivery point. On each wake, send every client the slice of the event log past its cursor — so a fresh connect (cursor 0 → full backlog) a