Lofn Step 11 Packager — Codex-backed fable/opus refinement-prompt builder

⚖️ AUTHORITY (2026-07-01): the .claude/skills/ twin of this skill is the CANONICAL policy source; this Codex mirror binds to it and to .agents/skills/lofn/EXECUTION.md §8 (Policy Deltas — golden-output quarantine, no-skip/NON-CANONICAL, itemized packet, per-pair variation angles, judge separation, the publish bar, gate mid-bands). On any disagreement, the .claude file wins.

Assembles the full context bundle for a Lofn music run into one paste-ready prompt per pair, so a high-capability model (Codex‑Fable 5 / Opus) can re-produce the Step 11 enhanced Suno package with complete creative authority. Each prompt embeds: the entire archive personality YAML (~50KB), the Suno v5.5 construction rules, the golden songs index, the full run context (seed → panel debate → metaprompt → pair assignments → creative context/ICB), the production mandates, the Step 10 package (and any prior inline Step 11 draft), and — as the final block — the canonical Manual Refinement Instructions (the executable task).

This is the Codex-native port of the legacy skills/step11-packager/. It is self-contained: the packager-specific references (Suno v5.5 rules, golden songs index, the default LOFN-PRIME YAML) are vendored under references/ so the skill runs even outside the repo. The repo archive is only a fallback for named personalities not vendored here. This skill swaps the execution layer: vendored/repo-relative paths, the Codex-native run layout, and cross-platform (Windows, UTF-8) I/O. The legacy skills/step11-packager/ is left untouched. See .agents/skills/lofn/EXECUTION.md.

When to use vs. lofn-music

• lofn-music runs the pipeline 00→11 inline (it already produces pair_NN_step10_final_package_enhanced.md).
• This skill is the manual-refinement packager: it does not generate songs. It wraps a finished (or partway) run's per-pair Step 10 outputs into the heavy refinement prompt so the smart model can rebuild Step 11 from scratch with all context in one paste — for a deeper polish pass, a Fable/Opus rerun, or a head-to-head against the inline step-11 draft.

Primary workflow — scripted (recommended)

scripts/build_fable_prompt.py auto-discovers everything from the run directory. On this machine the interpreter is python (Python 3.14).

# Build all pairs for a run (auto-detect repo root, pairs, personalities, context):
python .agents/skills/lofn-step11-packager/scripts/build_fable_prompt.py output/daily/2026-06-19

# Opus-named variant:
python .agents/skills/lofn-step11-packager/scripts/build_fable_prompt.py output/daily/2026-06-19 --mode opus

# Specific pairs only:
python .agents/skills/lofn-step11-packager/scripts/build_fable_prompt.py output/daily/2026-06-19 --pairs 1,3,5

# Override personality (archive filename or absolute path) / custom output / explicit repo root:
python .agents/skills/lofn-step11-packager/scripts/build_fable_prompt.py output/daily/2026-06-19 --personality lofn-prime-mini.yaml
python .agents/skills/lofn-step11-packager/scripts/build_fable_prompt.py output/daily/2026-06-19 -o /tmp/prompts
python .agents/skills/lofn-step11-packager/scripts/build_fable_prompt.py output/daily/2026-06-19 --repo-root E:/git/lofn

What it does:

1. Auto-detects the repo root (from the script location, then the run dir) — no hardcoded workspace path.
2. Discovers per-pair Step 10 files: <run>/music/pair_NN_step10_package.md (Codex-native) or <run>/audio/pair_XX_step10_production_wrap.md (legacy) — backslash-safe.
3. Resolves each pair's personality from 05_pair_assignments.md (LOFN-PRIME AWE/INDIGNATION mode by default; a named personality maps to its archive YAML; otherwise LOFN-PRIME).
4. Embeds the full personality YAML, the Suno rules, and the golden songs index — vendored copies in this skill's references/ are used first; for a named personality not vendored, it falls back to the repo archive skills/orchestration/personalities/ if present.
5. Embeds the full run context (research brief, golden seed, panel debate, metaprompt, pair assignments, creative context/ICB) and the production mandates.
6. Embeds the full Step 10 package and, if present, the prior inline Step 11 draft (pair_NN_step10_final_package_enhanced.md) as "improve on this" (suppress with --no-enhanced).
7. Appends the canonical Manual Refinement Instructions (three-block output spec, Disc_Channel + EMO format, <5000-char lyrics cap) as the FINAL block.
8. Writes <run>/music/pair_NN_step11_{fable|opus}_prompt.md per pair, in UTF-8.

Then: execute the refinement (Codex-native)

The packager only builds the prompt — it does not call any model. To produce the enhanced package, hand each pair_NN_step11_*_prompt.md to the refinement engine:

• Subagent (default): spawn one Agent per pair (model fable for the Fable-5 ceremony tier, or inherit Opus), feeding it the prompt file as its task. It returns the three-block enhanced package.
• Manual: paste a file into a Codex‑Fable/Opus chat for a hand pass.

The model's output MUST be the three canonical blocks (## SUNO STYLE PROMPT 850–1000 chars · ## SUNO EXCLUDE PROMPT 400–900 chars · ## SUNO ENHANCED LYRICS with [Theme:]+[SONG FORM:], 5-line Disc_Channel strip, em-dash EMO section headers, entire lyrics field < 5000 chars) plus every supporting block. The private bundle may contain Golden Song payloads for external review, but the generated output MUST NOT reproduce them; emit a concise ## Golden Move Calibration instead. Suno lyrics MUST NOT carry [11] or other pipeline citation markers. Save as pair_NN_step11_enhanced_suno.md, then route through lofn-qa.

Fallback workflow — manual build

If the script can't run, assemble by hand per the legacy spec in skills/step11-packager/SKILL.md §"Fallback Workflow" — same mandatory inclusions, but read context from the Codex-native filenames (core_seed.md, 03_panel_debate.md, 04_metaprompt.md, 05_pair_assignments.md, CREATIVE_CONTEXT.md) and Step 10 from <run>/music/pair_NN_step10_package.md.

Mandatory inclusions (every prompt file)

1. Context header (run, run ID, pair, personality, mode)
2. Full personality YAML — the ENTIRE file, zero bytes trimmed (LOFN-PRIME = lofn-prime-mini.yaml, ~1800 lines)
3. Full Suno v5.5 construction rules (7 principles + 7-position order + limits + format)
4. Full run context — never rely on the reviewer opening repo files
5. Golden Song References index (two examples with payloads)
6. Production mandates — full text, not summaries
7. Full Step 10 output — all sections, not just music prompt + lyrics
8. Prior Step 11 draft if it exists (improve, don't echo)
9. Manual Refinement Instructions as the FINAL block — the executable task

Anti-patterns — never do these

• ❌ Abbreviated personality block instead of the full YAML file
• ❌ Extracting only "music prompt + lyrics" from Step 10
• ❌ Truncating files with "[… truncated …]" markers
• ❌ Building prompts under 50KB (the script warns) — the YAML alone is ~50KB; a real bundle is 120–200KB

⛔ FUSION IS PERMANENTLY BANNED

OpenRouter Fusion (openrouter/fusion) drained all OpenRouter credits. Never invoke, mention, offer, or build Fusion-mode prompts. There is no --mode fusion. The refinement engine is Codex (Fable/Opus) only.

References

Vendored (self-contained, used first):

• references/suno_rules_condensed.md — Suno v5.5 rules (embedded by the script)
• references/golden_songs_index.md — golden song payloads
• references/personalities/lofn-prime-mini.yaml — the default LOFN-PRIME DNA (daily/fallback persona)

Repo fallback (for named personalities not vendored, non-destructive):

• skills/orchestration/personalities/*.yaml — full archive personality DNA. To make a named-personality run self-contained too, copy its YAML into references/personalities/.
• skills/step11-packager/SKILL.md — the legacy spec this port mirrors