Mount layout: precomputed digests for common cross-source questions

[khaliqgant]

Context

Companion to #106 and the filter-view issue above. The May 8 benchmark surfaced an opportunity that goes beyond reducing cost: there are common questions ("what did I do yesterday", "what's open right now", "what's recent") that MCP can't structurally answer in one call — they require synthesis across integrations, which forces a multi-step agent loop on either substrate.

Relayfile can answer these in one Read by precomputing them.

Proposal

Add a digests/ directory to the mount, populated by the sync daemon on a schedule (or on-write where cheap):

~/relayfile-mount/
├── digests/
│   ├── yesterday.md             # regenerated each morning at 00:00 local
│   ├── today.md                 # regenerated every N minutes
│   ├── this-week.md             # regenerated daily
│   ├── recent-activity.md       # last 24h, regenerated hourly
│   ├── my-open-work.md          # open PRs + tickets + draft pages, regenerated on writes
│   └── stale-work.md            # > 7 days idle, regenerated daily

Each digest is a small markdown file (1-5 KB) summarizing the relevant cross-source state.

Example: digests/yesterday.md

# Yesterday — 2026-05-07

## GitHub (54 items, 11 repos)
**relayfile-adapters** (most active): #38, #40, #42, #44, #46 ...
**relayfile**: #95, #98, #99
**cloud**: #284, #473, #476, #487
...

## Linear (1 issue)
- AGE-13 \"Demo issue\" updated 11:35

## Notion (31 pages edited)
- 13:12-14:25: Demos, Relayfile Integration, Relaycast, strategy pages...
- 20:03: Script

The agent reads this one file and answers. Cost: a single Read. Wall-clock: milliseconds. Quality: same as the multi-step run that produced the digest in the first place.

Why this matters

• The multi-source synthesis task is the most expensive query relayfile faces today and the one MCP loses thoroughness on. Precomputing it kills both costs simultaneously.
• This is genuinely something MCP can't do — there's no standard MCP tool for "summarize cross-integration activity". Each adapter would need its own. Relayfile's daemon already sees all the data.
• The digest itself is a correctness anchor: regenerated by the same code path each time, so the answer is consistent across queries. Versus an agent that synthesizes from scratch each run with subtle variation.

Predicted benchmark impact

For "what did I work on yesterday" specifically:

• Today: Sonnet RF $0.38 / 108s / 22 turns
• After Mount layout: add indexes, semantic aliases, and a SKILL.md to reduce agent file-shopping #106 + filter views: ~$0.15 / 30s / ~8 turns
• After this: ~$0.02 / 5s / 2 turns (1 Read of the digest, 1 Read of the synthesis prompt's response)

That's 19× cheaper than today's relayfile run, 10× cheaper than today's MCP run, and arguably more accurate because the digest is computed exhaustively at sync time, not under per-query token pressure.

Implementation notes

• Digests live as plain markdown — agents read them natively, no schema needed.
• Front-matter declares freshness:

  ---
  generated_at: 2026-05-09T00:00:00Z
  covers: 2026-05-08
  staleness: \"regenerated daily at 00:00 local\"
  ---
  

• The daemon decides regen cadence per digest. Some are cheap (yesterday is fixed once the day ends); some need rolling updates (recent-activity).
• Document the catalog in LAYOUT.md so the agent knows which digests exist.
• For digests that don't exist yet (e.g., user just connected), the directory entry can be a stub: # yesterday.md\\n\\nNot yet generated. Run \relayfile digest regenerate yesterday`.`

Out of scope

• User-defined digests. v1 is a fixed catalog; user-defined templates can come later.
• Personalized digests across users in shared mounts. Single-user for v1.

Acceptance

• digests/yesterday.md, digests/today.md, digests/recent-activity.md, digests/my-open-work.md populated and refreshed on a schedule.
• Re-running the May 8 benchmark prompt on relayfile shows < 3 turns and < $0.05 on Sonnet because the agent reads digests/yesterday.md and answers.
• LAYOUT.md (Mount layout: add indexes, semantic aliases, and a SKILL.md to reduce agent file-shopping #106) documents the catalog.