Personal digest: a private, DM'd feed roundup

ARCHITECTURE.md

@@ -188,6 +188,10 @@ Registry:
 | `suggest_feeds` | no (validates only) | — |
 | `add_feed` | yes | no |
 | `remove_feed` | yes | no |
+| `list_personal_feeds` | no | — |
+| `suggest_personal_feeds` | no (validates only) | — |
+| `add_personal_feed` | yes | no |
+| `remove_personal_feed` | yes | no |
 | `set_presence` | self only (own status/activity, persisted) | no |
 | `set_nickname` | self only (own guild nickname) | no |
 | `server_stats` | no | — |
@@ -257,6 +261,21 @@ triggerable via the `run_digest` tool. There is **no user input anywhere in this
   are truncated to 500 chars before the model sees them.
 - **Exactly-once posting.** Items are marked **seen** (`seen` table) only *after* a successful post,
   so a failed post retries the same items next time rather than dropping them.
+- **A personal, DM'd sibling.** `run_personal_digest_job` is the same mechanism — fetch, dedupe,
+  summarize — pointed at a second, separately-curated `personal_feeds` list, delivered to
+  `PERSONAL_DIGEST_CHANNEL_ID` if set, else DMed directly to the owner — same fallback shape and
+  the same "deploy owner's choice of destination and privacy, not Roger's" reasoning as Giga
+  Brain's periodic check-in (§12). It shares the `digest` brain's model and daily budget; it's a
+  second job, not a second brain. Curated the same way — `suggest_personal_feeds` /
+  `add_personal_feed` / `remove_personal_feed` / `list_personal_feeds` mirror the public digest's
+  four curation tools exactly. Scheduled unconditionally, same as Giga Brain's interval check —
+  the job itself decides "not configured" (no feeds) rather than the caller gating on whether a
+  seed env var happens to still be set, so feeds curated live via chat are actually delivered.
+- **One caveat: `seen` is shared.** Dedup is keyed on `(feed_url, entry_id)` globally, not per
+  list — a URL curated into *both* `feeds` and `personal_feeds` is only ever delivered by
+  whichever job runs first that day (personal digest defaults to `PERSONAL_DIGEST_HOUR=7`, before
+  the public digest's `DIGEST_HOUR=8`). Don't add the same feed to both lists if you want it in
+  both digests.
 
 ## §10 Persistence
 
@@ -272,6 +291,7 @@ behaviour adds rows, not migrations.
 | `admin_log` | Owner admin conversation memory, per channel (§6) |
 | `gigabrain_log` | Owner gigabrain conversation memory, per channel (§12) |
 | `feeds` | The curated digest feed list (§9) |
+| `personal_feeds` | The owner's personal digest feed list, curated separately from `feeds` (§9) |
 | `meta` | Small key/value bot state (persisted presence outfit, gigabrain's last-run date) — never pruned |
 
 ## §11 LLM layer & budgets

README.md

@@ -114,8 +114,10 @@ Feature-complete across the planned phases:
 - **Digest** — a scheduled daily RSS/Atom summary (also triggerable via `/roger run the digest
   now`), deduped so nothing posts twice. Roger curates its own feed list: `DIGEST_FEEDS` seeds it
   once, then Roger validates candidates against the live web and adds or drops them on request.
+  A second, privately-curated feed list can also be DM'd to the owner only
+  (`PERSONAL_DIGEST_FEEDS`), on its own schedule.
 
-Runs as a non-root, read-only-rootfs container. ~110 tests cover the guard rules, the tool loop
+Runs as a non-root, read-only-rootfs container. ~275 tests cover the guard rules, the tool loop
 (including channel creation with access presets and the confirm-gated edit, post, and reorder
 tools), the rate limiter, and the digest and feed-curation paths.
 

ROADMAP.md

@@ -0,0 +1,63 @@
+# Roadmap
+
+[`BACKLOG.md`](BACKLOG.md) is explicit that it's *"not a feature wishlist — it's the
+production-hardening layer."* This file is the other half: new capability, not hardening — ideas
+for Roger doing more without the owner personally driving every step, and for growing a small,
+personal-interest Discord server that doesn't have a lot going on day to day.
+
+Community/culture planning (what to post about, tone, events — not code) lives separately in
+[`community-discord`](https://github.com/R055LE/community-discord), since Roger and the server it
+runs in are different things. This file is the "how Roger could help" side.
+
+Same effort key as BACKLOG: **S** ≈ an afternoon, **M** ≈ a day or two, **L** ≈ multi-day.
+
+---
+
+## 1. Personal digest — **S/M** — *shipped*
+
+A feed roundup DM'd to the owner only, separate list and schedule from the public digest. Directly
+answers "I'm out of the loop" — reuses Digest's existing RSS/feeds/curation plumbing rather than
+building anything new, per [ADR-style discussion in the
+spec](docs/superpowers/specs/2026-08-18-personal-digest-design.md).
+
+*Why it's first:* smallest of the four, reuses infrastructure that already exists and is already
+tested, and is the most direct fix for the stated pain.
+
+## 2. Proactive public content ("Spark") — **L** — *idea only*
+
+The flagship. A new scheduled capability that posts unprompted to a public channel to spark
+discussion — reacting to real fetched items (AI/tech headlines, story recs, art highlights, open
+questions), not free-generated commentary that could confidently state something wrong. Same risk
+shape as Digest (§9 of `ARCHITECTURE.md`): scheduled, no user in the path, posts a message, nothing
+destructive.
+
+Needs its own design pass — content grounding, cadence, tone, and probably its own ADR given it's a
+new class of unprompted public output. Not spec'd yet.
+
+*Why it's not first:* biggest unknown, and the personal digest's grounding-vs-hallucination question
+(item 1) is a smaller version of the same problem worth solving first.
+
+## 3. Model allocation pass — **S** — *idea only*
+
+Give content-generation brains (Spark once it exists, Giga Brain now) their own tuned/creative model
+chain distinct from admin's tool-calling chain. There's real headroom to do this — Giga Brain's spend
+has been low. Mostly config + evaluation, not new code.
+
+*Why it's not standalone:* rides along with items 1 and 2 rather than being picked up on its own —
+worth doing once there's actual creative-generation output to tune against.
+
+## 4. Parked — self-serve commands for other members
+
+Not a stated pain point today (the server's small, and the ask was about the owner's own bottleneck,
+not delegating to others). Noted so it doesn't get lost if the server grows enough that it becomes
+one.
+
+---
+
+## Deliberately not doing (for now)
+
+- **Scraping non-RSS sources** for items 1/2. Feedparser-only, matching Digest's existing
+  constraint. A site with no feed is a future idea, not a blocker.
+- **A distinct model/budget for the personal digest.** It shares Digest's `MODEL_DIGEST` chain and
+  `DAILY_TOKENS_DIGEST`/`DAILY_USD_DIGEST` cap — a second job, not a second brain. Revisit if the
+  editorial angle (item 2 territory) ever gets folded into it.

compose.yaml

@@ -43,6 +43,9 @@ services:
       DIGEST_FEEDS: ${DIGEST_FEEDS:-}
       DIGEST_CHANNEL_ID: ${DIGEST_CHANNEL_ID:-}
       DIGEST_HOUR: ${DIGEST_HOUR:-8}
+      PERSONAL_DIGEST_FEEDS: ${PERSONAL_DIGEST_FEEDS:-}
+      PERSONAL_DIGEST_CHANNEL_ID: ${PERSONAL_DIGEST_CHANNEL_ID:-}
+      PERSONAL_DIGEST_HOUR: ${PERSONAL_DIGEST_HOUR:-7}
       OPS_CHANNEL_ID: ${OPS_CHANNEL_ID:-}
       METRICS_PORT: ${METRICS_PORT:-9108}
       TZ: ${TZ:-America/Detroit}