feat(talk): real-world topics — an off-loop pool the host can mention [spec 13]

[wine-fall]

Implements specs/spec13/13-real-world-topics.md (new). Stacked on #201 (the clock-as-bearings fix); merge that first.

What

The self-initiated talk task had no source of information — one terminal tool, no built-ins, a persona and a transcript — so it invented its topics and landed on the same cozy imagery every cold boot (#44; the listener's last 30 ledger topics were all afternoon light / warm cups / fridge hum). ROADMAP §2 set the line: real material, off the live loop, weighted by language and timezone, region never stored.

Pool — src/rwt.ts: cache/rwt.json, entries {id, title, gist, category, fetchedAt, used}, 48 h expiry, stale after 6 h; take() marks used and persists. Fetch — Brain.fetchTopics: runTask with a new builtins seam (WebSearch on tools and allowedTools), terminal submit_topics (zod), 12 turns, Haiku, a researcher framing that is never the persona. Roll — RwtRoll, RandomCadence's shape (p 0.35, minGap 1, maxGap 4 batches; MURMUR_RWT_* env knobs). Director — one offer() per ordinary generateTalks (never anchor:*, never the coda), maybeRefresh() poked at every boundary beside the compactor. Prompt — rwtLine: "Something from out in the world, if you want it … Material, not an assignment". Knob — rwtEnabled (settings file < --no-rwt), change_settings({ rwt }) + a line in the steer rule so "stop with the news" works; the stub brain fetches nothing. Taste — $MURMUR_HOME/rwt-policy.md, the music-policy shape, seeded on first use.

Language is read where the host reads its own. The first real run produced English gists on a Chinese persona: detectLanguage() follows LANG (now en_US), but the persona was seeded when it said Chinese and records that as "Always speak in Chinese (Mandarin)". So the request language is settings.language, else personaLanguage(persona) (the seed's "speak in X" clause), else the persona's own first line of prose held up as the example — a generated persona is written in the listener's language and never names it. Never the machine locale. Spec 13 §3.5 records the measurement.

This is the fourth network call; DESIGN.md's "three" is already stale per #104 and is not edited here.

Evidence (the deterministic seams, not the model's word)

Real fetchTopics through the SDK (scratch smoke, language "Chinese", tz Asia/Singapore): 6 entries in 105 s, gists in Chinese, weighted to the region (Taiwan defence budget, a Singapore animated film, Asia Cup cricket). A make dev run (--no-music --max-segments 10, MURMUR_RWT_P=1) in .dev/dev.log:

18:31:45 rwt.policy seeded /Users/zach/.murmur/rwt-policy.md
18:33:25 rwt.refresh n=6 ms=100229
18:33:25 rwt.pool fresh=6 used=0
18:34:01 talk.refill need=1 queued=1
18:34:01 rwt.offer ea853598          ← "The Gentlemen Season 2 drops on Netflix today" (entertainment)
18:34:01 rwt.pool fresh=5 used=1
18:35:35 radio: …我有个朋友前阵子还跟我念叨，说下了班本来盘算着看部剧，那种热热闹闹的、一伙人折腾着干大事的片子，结果往沙发上一坐，看了两眼就困了。你说这不挺好么，那剧它又不跑…

That beat proved the seam — and disproved the first draft of the line. The material reached the batch and none of it reached the air: a Netflix release scrubbed to "a friend wanted to watch a show", the #44 attractor wearing a fig leaf. The prompt had forbidden "a headline read out" and let the host "leave it". A radio names things. The second commit redraws the line on register, never content: name the thing — the title, who, where, when — say what happened in a sentence or two and what you make of it, then carry on; what stays forbidden is the newsreader's rundown, the "here is the news" frame, the list. The default policy asks the fetch to keep the hard nouns for the same reason. Spec 13 §2.5 records the measurement; #202's first box matches. The by-ear questions (does it sound like a host, is the proportion right) are #202.

Re-run with the redrawn line (2026-09-04, same flags), .dev/dev.log:

11:11:36 rwt.refresh n=5 ms=153065
11:11:42 rwt.offer e7c1de27      ← 新加坡霾害升至不健康级别 (news)
11:13:05 radio: 我今早还看到一条，说新加坡那边的天，今儿个中部一带脏得厉害，那个污染的数儿一下过了一百，到了对身子不好的地步了。这也不是一天两天的事儿，是印尼那头的林子，断断续续都烧了五十多天了……自打二○一五年往后，最凶的一回。你说这烟火啊，它不吵不闹……
11:13:05 rwt.offer 3d933063      ← 尼泊尔洪灾隧道救援三百多人 (news)
11:13:57 radio: 我今早看着一条消息，心里头沉了一下。尼泊尔那边，是冰河崩下来闹的水患……昨天救援队从一座发电厂的隧道里，把三百五十号人给拽了出来……死的已经过了一千一百，还有小五千人，到现在连影子都没着落……

Place, numbers, dates, the cause — named, in the host's own voice, one item per beat, no rundown. Titles and gists now arrive in the persona's language (the pool this run left behind is Chinese and stays).

The third commit is a one-line test fix outside spec 13: test/memory-fold.test.ts (from #196) reopened a store on the wall clock after recording on a fixture clock of 2026-09-01, so the 48 h recent window went red on 2026-09-04 for every PR. It now reopens on its own clock.

Tests (test-first, per step)

test/rwt.test.ts (pool expiry / take / persistence / refreshDue; roll determinism; single-flight refresh, failure posture, avoid list; the fetch task's builtins, bound, framing, and submit_topics termination), test/prompts.test.ts (rendering present/absent, never on anchor/coda, the fetch prompt's contract lines, the steer rule), test/brain.test.ts (agenticOptions builtins; stub returns nothing), test/config.test.ts / test/app.test.ts / test/steer-tools.test.ts (flag, file, env knobs, store seeding, change_settings({rwt}), buildRwt wiring, language resolution), test/persona.test.ts (personaLanguage), test/director-rwt.test.ts (one offer per ordinary batch, none on anchor/coda, knob off, boundary poke).

Peer review

Peer review (codex gpt-5.6-sol): 3 findings, 3 applied, 0 dismissed — language is never taken from the machine locale (a localized persona is held up as its own example, and a manner clause is not a language); an unwritable cache logs instead of throwing on the talk path; expired entries no longer block a recurring title. Each has a regression test.

Follow-ups

• By-ear pass: real-world topics, and the clock as bearings [spec 13 §5, spec 04 §3.4] #202 (by-ear, one issue for this ship and fix(prompts): the clock is bearings, not a line [spec 04] #201): friend-register, proportion, gist language, typed off, and the clock/profile fix.
• Talk beats repeat the same cozy imagery on cold start (streetlights / kettle / lamplight) #44 stays open until By-ear pass: real-world topics, and the clock as bearings [spec 13 §5, spec 04 §3.4] #202's first box holds; spec 13 is its durable fix.

AI coding brief

Original request: second of two sequential PRs from a finished design — write spec 13 (English, coding-agent form, spec 07's header structure), then build it test-first under murmur-build-spec without waiting for review. Every design point was pinned in the dispatching session (pool/fetch/roll/knob/policy/debug seams).

Manual interventions: none in this session. One decision taken mid-task: no pane row for the knob (spec 12's pane stays at eight; flag/file/conversation reach it) — logged in the spec's non-goals.

Retro: the real-seam smoke found the language bug the unit suite could not (persona-recorded language vs. machine locale); running the dev-log check before the PR, as the brief demanded, is what caught it. Next time, name "where the host reads its language" in the brief as a file:line, not a prose pointer.

🤖 Generated with Claude Code

https://claude.ai/code/session_018KVNGUoCkDfWFdLwkGWsAi