Post-publish print pipeline — photos, PDF, Drive upload. Edition-only for now (S188 rebuild); non-edition types await T11 wiring.
Post-publish print pipeline — photos, PDF, Drive upload. Edition-only for now (S188 rebuild); non-edition types await T11 wiring.
npx skills add https://github.com/pnils08/GodWorld --skill edition-print复制此命令并粘贴到 Claude Code 中以安装该技能
| name | edition-print |
| description | Post-publish print pipeline — photos, PDF, Drive upload. Edition-only for now (S188 rebuild); non-edition types await T11 wiring. |
| version | 1.4 |
| updated | "2026-05-30T00:00:00.000Z" |
| tags | ["media","active"] |
| effort | low |
| disable-model-invocation | true |
| argument-hint | [--cycle XX] |
/edition-print --type <type> --cycle <XX> --source <path>
Flags:
--type {edition|interview|supplemental|dispatch} — defaults to edition (back-compat).--cycle <XX> — engine cycle number.--source <path> — path to the published .txt. Defaults to editions/cycle_pulse_<type>_<cycle>_<slug>.txt for non-edition; editions/cycle_pulse_edition_<cycle>.txt for edition.Triggerable from: /write-edition post-publish chain, /interview, /dispatch, /write-supplemental. The four authoring skills converge here for visual assets the same way they converge on /post-publish for canon ingest.
The artifact is published. The text is canon. This skill turns it into something you can hold — photos, a PDF, files on Drive.
DJ Hartley is the art director for this step. His identity (.claude/agents/dj-hartley/IDENTITY.md) defines how photos are conceived — what's worth shooting, the mood, the composition. His team (Arman Gutiérrez for portraits, Brianna Lee for community) handles assignments. DJ reads the .txt and produces art direction; the image generator executes it.
DJ four-file canon-fidelity structure (S175) is LOCKED — the print pipeline wires AROUND DJ; do not modify his agent files.
Runs in its own terminal. Does not touch the artifact text.
| Type | Photos | Why |
|---|---|---|
| edition | 5–8 | Multi-article, full visual layout — DJ's standard workflow |
| supplemental | 1–3 | Topic deep-dive, fewer scenes |
| interview | 1–3 | One subject — typically a portrait + 1–2 supporting scenes |
| dispatch | 1–3 | One scene — single dominant image with optional supporting frames |
Budget is enforced by generate-edition-photos.js --type (engine-sheet T9).
For edition (the only fully-supported type after S188 rebuild):
editions/cycle_pulse_edition_<XX>.txt exists (compiled edition)output/sift_proposals_c<XX>.json exists (editorial weight signal — feeds DJ bundler)output/world_summary_c<XX>.md exists (atmospheric/neighborhood mood for DJ)/write-edition Step 6 for editions)output/production_log_c<XX>.md (consolidated per-cycle log; S195 convention)For non-edition types (interview/supplemental/dispatch): DJ-direction pipeline not wired yet (T11 pending). Photo gen for these types is temporarily unsupported by this skill until T11 extends the bundler to read alternate-start artifacts. The PDF + Drive steps still work.
Runs after the authoring skill (/write-edition, /interview, /dispatch, or /write-supplemental). Parallel with /post-publish. Both run independently — print produces visual assets, post-publish handles ingest and feedback.
Full chains:
/write-edition → /edition-print --type edition + /post-publish --type edition/<authoring-skill> → /edition-print --type <type> + /post-publish --type <type>Read the consolidated media production log: output/production_log_c<XX>.md. Append a ## Print Pipeline (<type>) section:
## Print Pipeline (<type>)
**Started:** {timestamp}
**Type:** <type>
**Source:** <path>
**Status:** IN PROGRESS
### Photos
[filled in at Step 1-2]
### PDF
[filled in at Step 3]
### Upload
[filled in at Step 4]
Three substeps. DJ Hartley directs (in-session subagent), FLUX executes (via Together AI).
node scripts/djDirect.js <cycle>
djDirect.js reads the compiled edition + sift_proposals_c<XX>.json (editorial weight) + world_summary_c<XX>.md (atmospheric mood), ranks proposals by sift score, pulls top-N article bodies (default 6), and writes a focused bundle to output/photos/e<XX>/dj_input_bundle.md (~50 KB). Engine audit JSON is intentionally excluded — not photo-relevant.
After djDirect.js prints its "next step" instructions:
output/photos/e<XX>/dj_input_bundle.mdsubagent_type="dj-hartley", passing the bundle as contextoutput/photos/e<XX>/dj_direction.jsonDJ's four agent files at .claude/agents/dj-hartley/{IDENTITY,LENS,RULES,SKILL}.md are LOCKED (S175 canon-fidelity Wave A). The bundle's §INSTRUCTION FOR DJ block carries pipeline-level guidance (canon-allowed brands, jersey-number rules, storefront rules) that overlays DJ's locked persona without modifying it.
node scripts/generate-edition-photos.js <cycle>
# OR chain QA + regen-on-fail in one pass:
node scripts/generate-edition-photos.js <cycle> --with-qa
Reads dj_direction.json, calls Together AI / FLUX.1.1-pro with each image_prompt verbatim (no synthesis), saves images + per-image sidecar JSONs ({slug}.meta.json) + manifest. With --with-qa, chains directly into Step 2.
Output: output/photos/e<XX>/ (image PNGs + {slug}.meta.json sidecars + manifest.json).
Two flows depending on whether you used --with-qa in Step 1c.
--with-qa already ran during Step 1cQA verdicts and regen attempts already executed. Read the report:
output/photos/e<XX>/qa_report.json — aggregate verdictsoutput/photos/e<XX>/{slug}.qa.json — per-image verdict + reasoningoutput/photos/e<XX>/regen_log.md — regen-on-fail tracenode scripts/photoQA.js output/photos/e<XX>
# OR: re-QA + regen-on-fail without re-paying for generation
node scripts/generate-edition-photos.js <cycle> --qa-only
photoQA.js reads each {slug}.meta.json, sends image + DJ spec to Claude Haiku Vision, returns PASS / FLAG / FAIL per the canon-aware rubric (negative-frame compliance, positive-frame compliance, tier-fidelity, tone). Canon-allowed brands (A's branding, Fox Theater, Heinold's, OMCA, etc. per docs/canon/INSTITUTIONS.md §Arts, Culture & Landmarks) are recognized.
--qa-only re-runs QA + regen-on-fail against existing photos (useful after a rubric change without re-paying for generation).
Regen-on-fail loop (T8): Each FAIL verdict triggers one retry with the same prompt (FLUX nondeterminism). If the retry passes or flags → photo updated, sidecar marked regenAttempt: 1. If the retry still FAILs → marked editorialFlag: true, pipeline continues without halting. All attempts logged to regen_log.md.
node scripts/generate-edition-pdf.js <source> --type <type>
Generates tabloid-format PDF using the masthead block from the .txt header (already standardized via T1). Output to output/pdfs/bay_tribune_<type>_c<XX>.pdf (edition keeps legacy output/pdfs/bay_tribune_e<XX>.pdf).
The PDF is local and not yet on Drive. This is the gate: open it and verify before it reaches the shared visibility surface. Moved ahead of Drive upload S246 (G-P-NEW3) — the C95 first render shipped a broken PDF to Drive at the old Step 4 before the eval pass at the old Step 6 caught the bylines-as-headlines bug, leaving two C95 PDFs on Drive (one broken). The eval review must gate the publish, not the close. Generator-side metrics are not eval-side review — the S212 generation-vs-evaluation asymmetry applies inside this skill: the generator has no holistic quality compass.
Open the rendered PDF. Visually verify three things:
Headlines match the article table. Every section's headline reads from the canonical ARTICLE TABLE (| Slot | Section | Reporter | Headline |) at the top of the edition .txt — NOT a lede-sentence extraction. EDITOR'S DESK gets its title from the article-table slot ED row. (G-PR7 — C94 PDF rendered all 6 article headlines as run-on lede sentences; the parser used lede extraction instead of article-table parsing. The compile-side canonicalShape gate in /write-edition Step 3 now catches the upstream cause, but eyeball the render anyway.)
Every approved photo is embedded at its correct section. Photo count in HTML <img> tags MUST equal Photos found: N (M dropped) from manifest line minus dropped count. Section-name normalization (FRONT_PAGE vs FRONT PAGE) is a known parser failure mode — visually confirm FP1, FP2, civic, sports, culture, business, etc. each carry their generated photo. (G-PR6 — C94 PDF dropped FP1 silently to section-name mismatch.)
Image execution reads on-canon. Open each photo. Does it read as prosperity-era Oakland (dynasty A's, building city, working neighborhoods with dignity) or struggling-city Oakland (poverty-doc aesthetic, blight-as-texture, despair-as-mood)? photoQA.js v2.0+ catches tone-vs-canon mismatch at the CANON-TONE axis (5th axis, S229), but tone is a softer signal than NEGATIVE_FRAME — eyeball review remains load-bearing. If any photo reads as struggling-city tone, FLAG and drop. (G-PR8 — C94 Beverly Hayes spec passed declarative items but rendered as poverty signifier.)
Gate rule: if any of the three fails, fix it (re-render the PDF / drop the photo / route the headline bug back to /write-edition) and re-run this step. Do NOT proceed to Step 4 upload until all three pass. Generator emits "PDF exists, 580KB, 6 photos, exit code 0" — that tells you the script ran, not that the artifact is usable.
Mags-discipline (MEMORY.md inline rule, S229): never close a print/publish skill without opening the artifact — and S246 strengthens it: never publish the artifact to Drive without opening it first.
Gate (G-P-NEW3): run this ONLY after Step 3.5 eval-side review passed all three checks. Drive is a shared visibility surface — an unverified PDF here is published-before-checked. If Step 3.5 flagged anything, you are not here yet.
node scripts/saveToDrive.js output/pdfs/bay_tribune_<type>_c<XX>.pdf <type>
Uploads PDF to Drive. Folder routing per --type: editions go to existing edition folder; interviews/supplementals/dispatches route to non-edition subfolder OR same folder with type-prefixed filename (engine-sheet T9 picks the implementation).
PDF only — photos remain local. (saveToDrive.js does not implement directory upload; see G-PR18 in print gap log.)
Edition path only — the dashboard uses the edition PDF for its UI:
output/pdfs/bay_tribune_e<XX>.pdf for dashboard to serveNon-edition types: skip. Dashboard surfaces editions, not alternate-start formats (yet).
The eval-side review already gated the pipeline at Step 3.5 (before Drive upload, per G-P-NEW3). Step 6 is the generator-side close — confirm the artifacts the prior steps produced all exist and landed:
output/photos/<type>_c<XX>/ — photos generated, count within budgetoutput/pdfs/bay_tribune_<type>_c<XX>.pdf — PDF generatedIf Step 3.5 was somehow skipped (no eval-side review logged in the production log), STOP — do not declare complete. The S229 rule holds: never close a print/publish skill without opening the artifact, and S246 (G-P-NEW3) moved that gate ahead of the Drive publish so a broken PDF can't reach the shared surface in the first place.
Done. Print pipeline complete.
At skill close, capture friction observed during the print pipeline as a gap log. /edition-print is a heavy skill at the media generator terminal; sidecar gap logs catch inefficiency the skill couldn't catch while running.
Output path: output/production_log_c<XX>_print_gaps.md (sidecar to the unified output/production_log_c<XX>.md; pipeline.34 convention production_log_c{XX}_<skill>_gaps.md).
Gap prefix: G-PR* (e.g., G-PR1, G-PR8, G-PR10).
Common categories for /edition-print gaps:
editorialFlag, manifest vs qa_report.json disagreement)Discipline: write the gap log even on clean runs. File a ROLLOUT row in pipeline.<n> pointing at the gap log per ADR-0005 §How to add work. Cross-cycle pattern observations are the highest-leverage findings (~30% of gaps across media skills are doc-drift class — single reconciliation pass closes them).
--type {edition|interview|supplemental|dispatch} flag added. Per-type photo budget table (1–3 for non-edition, 5–8 unchanged for edition). Triggerable from all four authoring skills. Plan [[plans/2026-04-26-non-edition-publishing-pipeline]] T4.djDirect.js (bundler) → dj-hartley subagent invocation in-session → generate-edition-photos.js (FLUX-direct, no internal synthesis). Step 2 documents --with-qa chain mode + --qa-only standalone QA + the regen-on-fail loop with editorial-flag marking. Prerequisites updated to require sift_proposals_c<XX>.json + world_summary_c<XX>.md. "Future (Phase 21.3)" note removed — it's current. Non-edition types (interview/supplemental/dispatch) flagged as T11-pending; DJ-direction pipeline doesn't reach them yet. Plan [[plans/2026-04-25-photo-pipeline-rebuild]].production_log_edition_c<XX>.md → consolidated production_log_c<XX>.md per S195) + G-PR18 (Step 4 "Photos also upload" claim → "PDF only — photos remain local" since saveToDrive.js has no directory-upload code). DJ slug-convention G-PR8 + word-band G-PR9 fixes live in dj-hartley agent files (Task 1.2).120-180 words band number anywhere in the text; the band lives in scripts/djDirect.js INSTRUCTION FOR DJ block (100-220 validator constants, 140-180 working comfort) per S197 G-PR9 closure. No skill-text edit required; the drift surfaced in C93 (G-PR6 logged "logged; proceeding") was already resolved by the S197 work without explicit gap-log closure. Marking closed in §Status updates of output/production_log_edition_c93_print_gaps.md.Audit architecture and system docs for staleness, inaccuracy, and drift from actual codebase state. Grouped into 7 tiers for thorough coverage. Complement to /md-audit (existence-staleness — when that ships).
Prepare all inputs for city-hall voice agents. Reads tracker, approvals, world summary, engine review, coverage ratings, previous log, canon, Mara directive. Writes pending decisions per voice.
Run the civic government — voice agents make decisions, tracker updates, canon output. Mags sifts, agents decide.
Immersive scene piece — one reporter, one location, one moment. Editions report, supplementals explain, dispatches immerse.
Capture-only — transcript becomes canon. Articles framed off interview transcripts move downstream to /write-edition (next cycle, via /sift) or /write-supplemental (any time, sports-desk subagent dispatch against the canon transcript).
Close the feedback loop. Canonize to Supermemory, update world-data, write ratings to sheets, grade reporters, update criteria files, update newsroom memory. Type-aware — edition, interview, supplemental, dispatch all converge here.