Personal digest (planned)
A second way to ask whodar a question. Today whodar answers "who do I talk to about X." The digest answers "what did I miss that matters to me." It scans the channels the running user can see, scores each message for personal relevance, and returns a ranked, actionable roll-up on demand, on a schedule, or live.
This is a new mode beside the existing pipeline, not a new resolver. Resolvers answer a topic query and return people and channels. The digest reads raw messages for one person over a time window and returns messages and suggested actions. It consumes the Slack client directly, before the connector aggregates messages into the people graph.
Why it belongs here
Every hard piece already exists. The Slack Web API client reads users, channels, and history. The messaging client sends direct messages and opens a Socket Mode connection. The LLM layer supports a local model through Ollama and a cloud model through Anthropic or OpenAI. The policy layer governs egress with strict, redacted, and open modes and an org lock. The bot frontend runs over Socket Mode. The digest reuses all of it and adds only a per-message relevance path.
Data flow
The digest sits beside source to connector to index. It reads the Slack client's history for each visible channel since a stored watermark, scores each message, buckets the survivors, and renders or delivers them.
slack client -> messages -> relevance -> buckets -> render or DM
Relevance
Signals evaluate cheapest first. Direct signals are deterministic: a mention of the running user, a reply in a thread the user is in, a direct message, and registered keywords such as project names or topics. Heuristics add weight from reaction volume, messages from key people drawn from the identity graph, embedded links, and action words like need, must, deadline, or broken. An optional LLM pass runs only on messages that clear the rules threshold, classifies whether the message is actionable for the user, and suggests a next step. The backend is the existing LLM layer, local or cloud by the user's choice.
The score maps to a bucket. Alert means act now. FYI means read it. Digest is low-priority context for the roll-up. Each item appears once, ranked, with the signals that produced it, a permalink, and a suggested action.
Governance
The cloud LLM pass must call the policy egress check before sending any message text off the machine. Strict mode denies it and the digest falls back to rules-only or the local model. Private channel scanning respects the existing private-channel policy toggle. The token is read from the OS keychain and never logged.
Faces
The command line mode is first: a digest subcommand that prints a ranked report, top items first. Scheduled delivery runs the same engine on a cron and sends the roll-up as a direct message through the existing messaging client. Live alerts reuse the Socket Mode bot to react to messages as they arrive and ping the user when one clears the alert threshold.
Extras whodar needs
These are the additive changes. None alter the existing query pipeline.
The Slack message type carries type, subtype, user, bot ID, text, and timestamp today. The digest also needs the thread timestamp, reaction totals, and a permalink. Extend the message type and the history call to include reactions, add a replies call to follow threads the user is in, and add a permalink lookup.
Self identity resolves through the existing auth test call, which returns the running user's ID for mention and thread matching. Direct-message delivery needs a conversations-open call to open the user's own IM channel before posting.
The token model differs from the bot. The digest needs a user token so it sees every channel the user sees, with scopes for channel history, group history, IM history, user read, and search read. Document this alongside the bot token.
A new internal digest package holds the relevance scorer, the bucketer, and the renderer. The scorer is a single-method interface with a rules implementation and an optional LLM implementation, chained so a nil LLM scorer means rules only.
A watermark store under the data directory records the last seen timestamp per channel so a run reads only new messages. It starts file-backed and can move to SQLite behind the same interface for history and a scoring audit trail.
A new digest command wires fetch, score, bucket, and render, with flags for the lookback window, channel scope, minimum score, LLM backend, delivery target, and output shape. The cloud backend is gated by the policy egress check.
Build order
Extend the Slack message type and history for threads, reactions, and permalinks. Add self IM open for delivery. Add the watermark store. Add the digest package with the rules scorer, bucketer, and renderer. Add the digest command for the command line face. Add the optional LLM scorer. Add scheduled DM delivery, then live alerts over the existing bot. The first five deliver a working rules-only command line digest.