--- name: lofn-daily description: Run the Lofn daily pipeline backed by Codex — fetch real-world facts, then generate the day's music (24 songs) and images (24→top 6) through the full Lofn pipeline with the daily rules (tri-source method, dual 3+3 constraint, emotional duality, library-only selection). Use for "daily run", "today's dailies", "run the daily pipeline", "do the daily drop", or a scheduled creative drop. Down-scalable for a quick test. Do NOT use for a single one-off competition piece (use `lofn`) or QA-only audits. ---

# Lofn Daily — Codex-backed daily run

The canonical recurring Lofn drop, ported to Codex. Faithful to `vault/DAILY_PIPELINE.md`. Where the original fired from an OpenClaw cron at 22:25 ET and delivered to Telegram, **you (Codex) run it on demand and present results in chat + save to disk.** The research is done by **this session itself with real fetches — never hallucinated, never delegated to a research subagent.**

``` PHASE 1 Research — fetch 20–25 verified real-world facts → 00_research_brief.md (you, inline) PHASE 2 Generate (in parallel): MUSIC → lofn pipeline → 6 pairs × 4 = 24 songs → best 6 IMAGE → lofn pipeline → 24 prompts → top 12 → top 6 PHASE 3 QA (lofn-qa) → save under output/daily/YYYY-MM-DD/ → present the drop ```

---

## 0. Run scope (confirm before a big run) The full daily run is **two complete pipelines** (24 songs + 24 images) — large. Pick a scope: - **Full daily** — music (24→6) **and** image (24→12→6). The real thing. - **Single modality** — just music **or** just image (still full cardinality). - **Test slice** — 1 modality, **2 pairs × 2 variations**, library personality/panel, skip render-ranking. Proves the wiring fast and cheap. *(Good default when the user says "test the daily run".)*

State the chosen scope in the brief and in the run INDEX so QA knows the intended cardinality. Down-scaling is explicit, not silent.

---

## PHASE 1 — Research brief (you, inline, real fetches) Use **WebFetch / WebSearch** (the Codex equivalents of the legacy `web_fetch`). Follow `skills/lofn-core/steps/00_music_research.md` (25-fact music research) and, when the image lane is in scope, the NightCafe-themed research in `skills/lofn-core/steps/00_research.md`. Fetch from the daily source table (`vault/DAILY_PIPELINE.md`) — at minimum:

**OpenClaw ledger standard:** this is a dispatch summary, not the saved brief format. `00_research_brief.md` MUST expand the full `vault/DAILY_PIPELINE.md` ledger into one row each for `F01` through `F25`. Do not collapse ranges such as `F1-F3` or `F21-F25`; use `OK`, `NO DATA`, `UNAVAILABLE`, or `SCOPE-SKIPPED` per row.

| Code | Source | Extract | |------|--------|---------| | F1–F3 | NightCafe daily challenge (`nightcafe.studio/pages/daily-challenge`) | challenge #, theme, what wins (image lane) | | F4–F5 | USGS quakes (`earthquake.usgs.gov/.../significant_day.geojson`) | magnitude, place, depth | | F6–F7 | NASA APOD (`api.nasa.gov/planetary/apod?api_key=DEMO_KEY`) | title, first sentence, color/light, **image structure** | | F8 | Poetry Foundation poem of the day | poet, most physical line | | F9–F10 | Bandcamp Daily (`daily.bandcamp.com`) | album, genre tags, **exact sonic-texture quote** | | F11 | Protein Data Bank molecule of the month | molecule, structural descriptor | | F13 | Color API (`thecolorapi.com/id?hex=MMDD`) | hex, name, emotional association | | F17 | Oblique Strategies (`stoney.sb.org/eno/oblique.html`) | exact phrase verbatim | | F18 | Space weather (`services.swpc.noaa.gov/products/summary/solar-wind-speed.json`) | solar wind speed, Kp | | F19 | Hacker News (`news.ycombinator.com`) | top 3 titles — what builders discuss | | F20 | BBC World RSS (`feeds.bbci.co.uk/news/world/rss.xml`) | top 3 headlines — world's emotional temperature | | F21–F25 | Public Domain Review, Almanac moon, NOAA buoy 46059 | esoteric visual detail, moon folklore, wave data |

Mark any JS-gated/unavailable source `UNAVAILABLE` and continue. Add 3–5 obscure theme-specific facts (not the obvious Wikipedia entry). Also write the 5 **EXISTENCE** prompts (interior-life questions songs can answer) from `00_music_research.md`.

**Save** `output/daily/YYYY-MM-DD/00_research_brief.md` with the Tri-Source Summary + the 3+3 seeded split (see below). Today's date is available in context — use it for the directory.

> ⛔ **One controller per directory — take the lock BEFORE the research brief.** This paragraph used to be advice, and on 2026-07-24 a second controller wrote into a live run's directory and destroyed eleven hours of it. Advice does not interlock. The first action of the run, before this step writes anything: > > ```bash > python3 scripts/run_lock.py acquire output/daily/YYYY-MM-DD --run-slug <run-slug> --engine codex > ``` > > **Exit 3 means STOP** — another run holds that directory. Do not inspect it and decide for yourself; give this run its own directory (`output/daily/YYYY-MM-DD-<run-slug>/`), or resume the other run by its exact slug, or ask The Scientist. See `EXECUTION.md` §5.1. Then `heartbeat` at every wave and `release` after the INDEX. If the lock says the same run is already under way, **resume** it from the RUN_STATE manifest (§ "Run-state manifest & resume") rather than re-running passed pairs.

### Advisory learnings note (dispatch brief only — NEVER the ICB) Before dispatching either modality, **tag-walk `vault/COMPETITION_LEARNINGS.md`** (and `vault/LESSONS_INDEX.md` if it exists) for the **3–5 entries that intersect THIS run's theme/venue/modality** — e.g. a container/object theme pulls the Container Test, a NightCafe image lane pulls the warm-palette/anti-austerity lessons, a portrait theme pulls the portrait shifts. Surface them **as ADVISORY NOTES in the dispatch brief / Phase-0 reasoning ONLY.** Hard rules: - **NEVER injected into the ICB / `CREATIVE_CONTEXT.md`.** The ICB stays read-only and lesson-free. These notes live in the brief the coordinator reasons over, not in the verbatim block every subagent receives. - Each note carries its **confidence %** (from the entry) and is run through the mandatory **"would this have hurt our best past entry?"** gate before it is allowed to influence anything. A lesson that would have hurt a past win is dropped, not applied. - **Venue/modality-scoped:** an image-venue lesson must NOT leak into the music lane; a NightCafe-voting lesson must NOT leak into non-competition runs. - **Advisory, never a hard constraint.** A note can inform the pair brief's framing; it can never auto-reject a candidate or become a gate. Promotion to a hard constraint is a **human** decision. - **Triggered-INDIGNATION is EXEMPT from suppression.** A lesson such as "INDIGNATION underperforms on NightCafe" may inform image-venue *selection* advice, but it must NEVER suppress an INDIGNATION piece the panel deliberately chose, and never touches the music lane or the ≥1-INDIGNATION duality rule.

State in the brief: "Advisory learnings consulted: <N entries, tags>; INDIGNATION exempt; advisory-only." If zero entries intersect, say so. Write-back of one curated entry per shipped/selected piece happens in `lofn-qa` / Phase 3 — not here.

---

## PHASE 2 — Generate (daily rules layered on the `lofn` pipeline) For each in-scope modality, run the **`lofn`** pipeline (Phase 0 Golden Seed → Phase 1 3-panel orchestrator → modality steps → QA) using the research brief as `{input}`, **plus these daily-only rules:**

### Tri-Source Methodology (declare BEFORE writing any artifact) Every daily piece integrates three sources; state them explicitly in the metaprompt and each pair brief: - **Source 1 — CONTENT / emotional stakes:** today's world facts (quakes, APOD, F19 HN, F20 BBC, moon, solar weather). Songs/images are *resonance*, not reportage. **⛔ One-fact rule (music):** the tri-source method feeds the THEME and FORM — it is not a lyric quota. **At most ONE numeric fact is sung per song**, at the emotional hinge, responded to rather than recited (`lofn-music` Golden Move rule 2; `gates.yaml → max_sung_numeric_facts`). A verse reciting the day's sunspot number, solar-wind speed, moon percentage, AND quake depth is a weather report in meter — a repair. The other facts inform the pair briefs and stay there. - **Source 2 — SONIC/AESTHETIC VOCABULARY:** the exact Bandcamp review language (F9–F10) imported into prompts — grounds the sound/look in something specific and real, not generic genre labels. - **Source 3 — MATERIAL STRUCTURE:** the NASA APOD image structure (or a PDR artifact) translated into a **mandatory form rule** — e.g. "comet with long tail" → long trailing fade-out outro; "3×1 tile panel with meanders" → 3-section form with transitional bridges; "bilateral wing venation" → mirrored call-and-response.

### Dual 3+3 Constraint (set at pair-assignment time, Phase 1 step 5) - **Axis A — ACCESSIBLE vs AMBITIOUS:** pairs 1–3 ACCESSIBLE, pairs 4–6 AMBITIOUS. Final top 6 = best 3 from each arm; rank **within each arm only** (never 5+1 or 6+0 by global score). - **Axis B — NEWS vs EXISTENCE:** **max 3** pairs anchored to today's research/news; **min 3** pairs explore existence/interior-life/universal experience. All 6 on one theme = a lecture, not a record.

### Emotional Duality & diversity - **≥1 AWE song and ≥1 INDIGNATION song** in the set. - 6 different verse architectures / camera grammars across the 6 pairs (the standing distinctiveness rule). Vary stanza lengths intentionally. - **Variation angles are per-pair, never a shared template set.** A global "V4 = glitch chapel for everyone" scheme is how the 2026-06-26 daily produced two pairs singing the same song with nouns swapped; each pair derives its own 4 angles from its own concept (`EXECUTION.md` §3 item 3). - **Daily music pair isolation is mandatory.** Steps 05–11 must run as isolated pair runs; a central all-pairs helper may not author lyrics, section maps, stanza scaffolds, hook grammar, rhyme logic, or production arcs for multiple pairs. Shared validators and file writers are OK. If the pairs sound like the same song with nouns swapped, rerun each affected pair from Step 05. - **AWE stays terror-adjacent.** The daily's comfort gravity is real (kitchens, cups, reassurance) — every AWE song still answers the two pre-draft questions (*where is the body standing / what could hurt it here*) and carries a clean fear (`lofn-music` Golden Move rules 1 & 4).

### Library-only selection For daily runs, **always select personality + panel from the existing libraries** (`personalities_index.md` / `panels_index.md`) — **no generation.** Freshly generated personalities over-fit the day's theme and lose the battle-tested DNA. (Generation is reserved for competition/Scientist-special runs.)

### Modality specifics - **MUSIC** (`lofn-music`): 24 songs (6×4); each ≤1000-char two-field Suno prompt, female vocals default, EMO headers, 70–120 lines; `06_audio_handoff.md` carries 2 Golden Songs. Run music in parallel with image. - **IMAGE** (`lofn-image`): 24 prompts → rank → **top 12 → top 6**; figurative legible primary subject (thumbnail test); noun-first present-tense ≥80 words; warm palette leads on NightCafe-style venues (INDIGNATION underperforms there — see `vault/COMPETITION_WORKFLOW.md`); aspect 3:4 for upload challenges else 9:16. Apply the **Container Test** (`COMPETITION_WORKFLOW.md`) and **Action-Verb rule** (action theme → cinematic wide, not portrait).

Run the two modalities concurrently (independent `lofn` runs writing to `music/` and `images/` subdirs). Each fans its 6 pairs out as parallel subagents per `.agents/skills/lofn/EXECUTION.md`.

> ### ⚙️ Concurrency: cap-and-stagger (do NOT run all 12 chains at once) > The full daily is **two pipelines × 6 pairs = 12 concurrent chains** — enough to blow the tool/context budget if fired together. **Cap and stagger, don't serialize:** > - **Cap the in-flight pair-subagents** (default ~6 at a time, not all 12). Launch one modality's 6-pair wave, t