The context layer, milestone 1. A local store of derived metadata and pointers over Apple Mail — never copies. See docs/04_context_layer.md for the full design.

Pointers, not copies

context.db (created mode 600) holds headers, a people graph, and Message-ID pointers. There is no body column in any table — a test enforces this. Content is fetched live through honeycrisp-mail by consumers who hold that permission; the serve layer itself needs no Mail access at all.

Usage

honeycrisp-context capture   # one incremental sweep (headers only, ~seconds)
honeycrisp-context serve     # MCP server: context_changes, context_person,
                           # context_capture_now

context_changes — what's new since a point in time: volume, senders, newly seen people, active subjects. Deterministic; every claim cites Message-IDs.
context_person — interaction profile: counts, first/last seen, recent message pointers.
context_capture_now — on-demand sweep (write-safe: it writes only the local store).

Capture keeps per-mailbox cursors, backfills 90 days on first run, and audits every sweep (counts only) to the suite's shared audit DB. Schedule it with launchd (or any timer) for continuous capture; a missed run just widens the next sweep.

The briefing pipeline (M2–M4)

honeycrisp-context brief <dir> runs the full cycle: ingest corrections from prior briefing files → extract commitments from recent sent mail (Layer 1) → answer standing questions → write today's briefing.

Standing questions — standing-questions.yaml in the Briefings folder (phone-editable). Deterministic candidates always; a configured model writes cited answers over fenced header data.
Layer 1 is opt-in and declared — context.json: {"model": {"type": "ollama", "model": "..."}} (local) or {"type": "command", "run": ["...your llm cli..."], "network": true}. network defaults to true — egress is assumed unless declared otherwise — and every Layer-1 run writes an audit row carrying the declaration. No config → deterministic only; nothing is guessed. Model-invented pointers are rejected; a failing model degrades to candidates, never a crash.
The flywheel — briefing items carry stable ids; tick a checkbox from any device (or add lines under ## Corrections) and the next run ingests them as judgments: muted senders/subjects/items disappear from future briefings while the AI-layer rows stay immutable.

Retention

Derived rows age out on capture: messages, events, and extracted commitments older than retentionDays (default 365) are pruned; set {"retentionDays": 0} in context.json to keep everything forever. People aggregates and human judgments are never pruned at any age — deleting a human's recorded judgment is not the machine's call. Every prune writes an audit row with counts.