autholog-mend

# autholog-mend — raw writing intake Recover GLG's public raw writing as a `notes/` autholog. This is the raw → room → outgoing magnets → world-image skill lane (the old monolithic `/mend` command split here and into `tag-mend`). For one-document core-garden restoration and its playful image/prompt round, enter `/authologplay`; this skill remains the ROSSE intake and room-choice contract. ## Contract - **One raw piece, one autholog.** Never hide a new piece inside a related note. - **Reuse a real empty room.** Search thin/hold/temp/outdated notes, but do not evict a short note whose topic is still alive. - **Raw voice is sacred.** Preserve exact text in `#+begin_quote` + `[!danger]`; do not fix spelling or prompts. - **Add explanation.** Give time/place, concepts, misreading boundaries, links, and the reason this room fits. - **Connect magnets, do not rebuild them.** On the new autholog only, point `관련메타` / `관련노트` at useful `meta/` notes and keep established English tags. Do not refresh/eval meta dblocks so a write “shows up” in the magnet — GLG’s export script does that. Regexp fixes on a broken dblock definition are fine; body rewrites of meta notes are not part of this lane. `notes/` is GLG's rail; agent exposition stays in `botlog/`. - **One autholog, one world image.** Every autholog ultimately has one GLGMAN Universe image: 200 authologs means 200 images. - **Prompt headings are the visual index.** Put the complete prompt below a heading tagged `:PROMPT:` such as `** 프롬프트 전문 :PROMPT:`. This is a heading tag, never a filetag. - **Text comes first.** Keep an adequate image despite minor prompt drift and describe the delivered pixels honestly. If GLG defers image work, finish the text; backfill legacy authologs gradually, never as a bulk image job. - **Preserve room 히스토리.** A moved seed goes to its proper meta/hub and remains named under `옛 방의 씨앗`. - **Name the author in 히스토리.** Every 히스토리 line names who wrote that pass. Raw voice stays `@junghan`. Agent mend/re-explanation names the **model with version** (e.g. `sonnet5`, `opus`, `gpt-5.6`, `grok-4.5`) so the 해설본's stance is legible later — harness-only stamps like bare `pi` are not enough. Read the live session model (`PI_MODEL` / `PI_AGENT_ID`); never invent one. Do not hard-code one sigil grammar; follow the pattern already used in nearby recent 히스토리 lines, and keep the version visible inside whatever form that is. Never merge a human edit and an agent edit into one untagged line. ## Image contract - Before drafting, offset-read the GLGMAN common block and the nearest `:PROMPT:` headings only; do not fork the visual language, and do not load the full character sheet or world bible. - Keep image link, generation parameters, and complete prompt together in the autholog. - Compare the actual image with the prompt. Reversed gestures, missing objects, or ambiguity are provenance. - New mends normally close with an image. Existing image-less authologs are a slow backfill queue unless GLG explicitly asks for generation now. ## Context budget / retrieval gate Concentrate force on three points only: **(A) ops context + raw restore**, **(B) room judgment**, **(C) present re-explanation + anchors + image**. Everything else stays thin and decisive. ### Hard stops - **Raw once.** Read the source raw fully exactly once. Treat later `exact-match` verification as the final gate, not a second interpretive pass. - **Outline-first candidates.** After semantic/`denotecli` search, run `denotecli read ID --outline` only. Reject immediately when outline/frontmatter already shows a living independent raw, a finished autholog with another piece, or a long reading note. Do **not** full-body `denotecli read ID` on candidates. - **Body read is narrow.** Open body only for the chosen room, or for a still-undecided candidate, and only the needed `offset` window (~40–80 lines). Stop the candidate loop the moment one true empty/thin room is confirmed. - **Cap 2–5 candidates.** Never widen the set to “be thorough.” Graph/backlink walks run only after room choice, and only when the selected room or raw actually needs link repair — never per candidate. - **Re-explanation anchors: 2–4.** Read only the parts the raw itself calls. Label each use as fact / association / judgment. Do not preload adjacent garden lore. - **Image SSOT is offset-thin.** Read the GLGMAN common block and the nearest `:PROMPT:` via offset. Do not read the full character sheet or world bible. Generation: skill `glg-image` — **speed first** (autholog often retries 2–3×): default **Gemini Flash Lite @ 1K**, escalate Flash @ 1K, optional **grok**; filename `__brand_<geminilite|geminiflash|grok|…>` required. No auto-2K/Pro. - **No dblock refresh/eval.** Never run org dblock update / agent-server dblock eval to “reflect” a new magnet while writing. GLG’s garden-export script refreshes magnets. Conservative **regexp edits** on an existing dblock definition are allowed when the pattern itself is wrong (`\|` etc.). Do not bulk-edit reverse-link indexes. On the new autholog, repair only its own outgoing `관련메타` / `관련노트` links (and move a displaced seed when Contract requires it). - **tag-mend is optional and separate.** Invoke only for new/suspicious tags or a filename-boundary problem. Standard authologs that already carry established magnet tags get a filename-budget check only. - **Final verify is focused.** Raw exact match, focused diff of touched files, filename budget, image exists (+ pixels vs prompt when generated). No whole-repo `status` archaeology, no full diff dump, no meta-index rebuild. ### Allowed deep reads | Target | Depth | |--------|--------| | Source raw | Full, once | | Candidate rooms | `--outline` first; body only if still needed, 40–80 lines | | Chosen room | Enough to remodel safely | | Re-explanation anchors | 2–4 notes, needed sections only | | GLGMAN / image | Common block + adjacent prompt offsets | | Meta notes | Outgoing link targets of the new autholog only; never their dblocks | ## Workflow 1. **Stop before writing** when room choice is ambiguous. 2. **Find 2–5 candidates** with `denotecli search`, `search-content`, `knowledge_search`, then `read --outline` only — obey the retrieval gate; drop living/long notes at outline. 3. **Choose one true room:** stale/thin/outdated, conceptually resonant, and not a complete autholog with another raw piece. Confirm → stop searching. 4. **Undo wrong placement first.** Restore a displaced topic before reusing another room. 5. **Remodel:** front matter, abstract, 히스토리 (agent line = model with version; follow nearby stamp pattern), outgoing meta/notes links, concise explanation (fact/association/judgment), exact raw block, optional follow-up raw, and old-room seed. 6. **Outgoing links only.** Same magnet rule as Contract: point `관련메타` / `관련노트` at useful magnets; move a displaced seed when needed. No dblock refresh/eval; no unverified LLM citations or claims. 7. **Handle the image.** Offset-read common block + nearest prompt; generate via skill `glg-image` (**lite@1K first**, then flash@1K, optional grok). Save `~/screenshot/YYYYMMDDTHHMMSS--slug__brand_<geminilite|geminiflash|grok>.jpg`. Keep complete prompt under `:PROMPT:`; or record explicit text-first deferral without claiming completion. 8. **Rename through Emacs:** `agent-denote-set-front-matter ... :rename t`; never raw `mv`. 9. **Verify (focused):** `denotecli read ID --outline`, raw exact match, actual pixels vs prompt, image path exists, filename budget, focused diff of touched files only. 10. **Report context spent.** After the mend closes, give a short spend note for this pass only: roughly what was read (raw / outlines / chosen room / anchors / image SSOT), what was written, and approximate bulk (e.g. candidate full-body avoided). No whole-session dump — just enough that the next reader sees where force went. ## Standard shape ```org #+title: @힣: ... #+filetags: :autholog:...: #+description: ... #+begin_quote [!abstract] 이 노트에 대하여 ... #+end_quote * 히스토리 * 관련메타 * 관련노트 * 한 줄 * <core axes and misreading boundary> * 이미지 — <scene and actual reading> ** 생성 파라미터 ** 프롬프트 전문 :PROMPT: * 원문 보존 — <surface> 날것 #+begin_quote [!danger] ...exact raw... #+end_quote * 후속 댓글 보존 — ... * 옛 방의 씨앗 ```