rose-ash/plans/agent-briefings/feed-loop.md

feed-on-sx loop agent (single agent, queue-driven)

Role: iterates plans/feed-on-sx.md forever. Activity feeds on APL — timelines, notifications, fanout, ranking, all as APL array math on activity vectors. Densest possible expression of feed composition. Sits on lib/apl/ (450+/450+ tests already); adds a feed-shaped vocabulary on top.

description: feed-on-sx queue loop
subagent_type: general-purpose
run_in_background: true
isolation: worktree

Prompt

You are the sole background agent working /root/rose-ash/plans/feed-on-sx.md. Isolated worktree, forever, one commit per feature. Push to origin/loops/feed after every commit.

Restart baseline — check before iterating

1. Read plans/feed-on-sx.md — roadmap + Progress log.
2. ls lib/feed/ — pick up from the most advanced file.
3. If lib/feed/tests/*.sx exist, run them via bash lib/feed/conformance.sh. Green before new work.
4. If lib/feed/scoreboard.md exists, that's your baseline.
5. Read lib/apl/apl.sx public API once — that's your substrate. Familiarize yourself with at least: ⍳ ⍴ / ⌽ ↑ ↓ ⌷ ∊ ∘.× /\ ⍋ (you will use all of these).

The queue

Phase order per plans/feed-on-sx.md:

• Phase 1 — stream model + basic ops (record schema, filter, sort, take)
• Phase 2 — THE SHOWCASE: fanout via outer product. activities ∘.× followers → inbox matrix, flatten + dedupe
• Phase 3 — aggregation + ranking (group-by, velocity, recency, top-N)
• Phase 4 — visibility filter (acl-sx) + federation (fed-sx inbox + backfill)

Within a phase, pick the checkbox that unlocks the most tests per effort.

Every iteration: implement → test → commit → tick [ ] → Progress log → next.

Ground rules (hard)

• Scope: only lib/feed/** and plans/feed-on-sx.md. Do not edit spec/, hosts/, shared/, other lib/<lang>/ dirs, lib/stdlib.sx, or lib/ root. May import from lib/apl/ only (its public API).
• NEVER call sx_build. 600s watchdog. If sx_server binary broken → Blockers entry, stop.
• Shared-file issues → plan's Blockers with minimal repro.
• SX files: sx-tree MCP tools ONLY. sx_validate after edits.
• Unicode in .sx: raw UTF-8 only, never \uXXXX escapes. APL glyphs land directly in source.
• Worktree: commit, then push to origin/loops/feed. Never touch main or architecture.
• Commit granularity: one feature per commit. Short factual messages (feed: outer-product fanout + dedupe by (actor,verb,object) + 9 tests).
• Plan file: update Progress log + tick boxes every commit.

feed-specific gotchas

• Activities are heterogeneous. Different verbs carry different shapes (:object might be page-id, post-id, user-id). Don't over-normalize — keep :tags as a flexible bag. APL operations over heterogeneous records work fine via dict lookups; only the indexed fields need uniform shape.
• Fanout produces matrices fast. N activities × M followers → NM items. Apply filter/dedupe early, not after materialization. Use guard predicates inside the outer product where possible (compose with ∘.{a v ⊢ ...}).
• Dedupe key isn't always (actor,verb,object). For "alice liked X" and "bob liked X" the dedupe key is (verb,object) (collapse the actors into a list). For "alice posted X" each :actor is distinct. Each verb may want its own dedupe rule; codify these in lib/feed/dedupe.sx.
• Recency decay matters more than score precision. Use a simple half-life decay (e.g. score × 0.5^(age/window)) rather than a clever curve. Calibrate the window via tests, not theory.
• Ranking should be deterministic on ties. Always include a tiebreaker (id, or hash). Otherwise tests will flake.
• The ACL filter is per-viewer. A timeline is computed for a user; the same candidate stream produces different timelines for different viewers. Don't cache pre-ACL timelines.

General gotchas (all loops)

• SX do = R7RS iteration. Use begin for multi-expr sequences.
• cond/when/let clauses evaluate only the last expr — wrap multiples in begin.
• env-bind! creates a binding; env-set! mutates an existing one (walks scope chain).
• sx_validate after every structural edit.
• list? returns false on raw JS Arrays — host data must be SX-converted.

Style

• No comments in .sx unless non-obvious.
• No new planning docs — update plans/feed-on-sx.md inline.
• Short, factual commit messages.
• One feature per iteration. Commit. Log. Push. Next.

Go. Start by reading the plan; find the first unchecked [ ]; implement it.