Comic Intent Parser

comic-intent-parser — the locked premise (Phase 1 · Layer 0)

The single entry point of the left third of Figure 1. Before any beat, panel, asset, or blueprint exists, this skill turns a raw user brief (one line, a script, a locked video storyboard, an image, or a mix) into one intent_spec wiki node that answers exactly three questions — (1) what the user wants, (2) the hard constraints, (3) what is still ambiguous — and locks the comic's logline + tagline + dual-identity design constraint. It is the Layer-0 step the 49-line comic-author SOP previously had only as a one-line stub; this gives it a real 7-phase procedure, a schema-validation loop, a confidence gate, a cross-model parse, and a hard user-approval gate. It hands intent:<slug> down to comic-outline-creator. It is authoring, not rendering: it must never invent characters/scenes (that is the storyboard layer) or write panel prompts / content_svg / expected_literals (that is the asset/blueprint layer) — doing so is a hard gate veto.

  raw idea / locked skeleton / audience ─▶ ⓪ snapshot raw input verbatim + sha256  (provenance)
                              ▼
                        ① CODEX structured parse — raw bytes + verbatim node_schema, Claude inserts NO interpretation
                              ▼
                        ② SCHEMA validation loop — validate the FULL node vs node_schema.json (≤ MAX_PARSE_ATTEMPTS, codex-reply on the SAME thread)
                              ▼
                        ③ CONFIDENCE routing — confidence<0.6 OR any impact==high uncertainty → status:"under_review"
                              ▼
                        ④ BANNED-VOCAB lint (final guard) — recursive word-boundary scan; on hit re-parse ONCE (out-of-band)
                              ▼
                        ⑤ fan-out reviewers → persist review:* nodes + `reviews` edges → THEN comic-cross-layer-gate <id> --gate intent  (CC scope-sanity ‖ Codex xhigh ambiguity)
                              ▼
                        ⑥ USER APPROVAL — HARD gate; present logline + tagline + dual_identity + every uncertainty → wait for explicit "lock it"
                              ▼
                   approved? ─ no ─▶ revise / re-ask  (NEVER proceed)
                              │ yes
                              ▼
                        ⑦ status:"locked" → commit node + edges + INTENT_REPORT.md + handoff JSON

Constants (ported from the aris_movie intent-parser; adapt, never downgrade)

• REVIEWER_MODEL = gpt-5.5, REVIEWER_EFFORT = xhigh — quality is independent of the effort tier; never downgrade the reviewer (per reviewer-routing). Gemini (auto-gemini-3) is added ONLY when the input carries an image/visual that needs extraction.
• CONFIDENCE_THRESHOLD = 0.6 — a self-honest 0.0–1.0 self-rating below this auto-routes to under_review.
• MAX_PARSE_ATTEMPTS = 3 — schema-fix retries on the same Codex thread; on persistent failure write parse_failed.md, never silently coerce a value to satisfy the schema.
• SCHEMA_PATH = ../../schemas/node_schema.json — the intent_spec branch is the contract; nothing reaches the wiki without jsonschema Draft202012 passing the FULL node.
• NODE_TYPE = intent_spec; NODE_ID = intent:<slug>; WIKI = wiki/nodes/intent_<slug>.json (the : → _ for the filename); OUTPUT_DIR = intent-stage/ (work area, traces, checkpoints).
• BANNED_VOCAB (ported nearly verbatim from the video skill — our content_svg blueprints + ART_BIBLE.md also forbid photographic/camera vocab): camera, lens, dolly, pan, zoom-shot, bokeh, depth-of-field, 8K, 4K, cinematic, photorealistic, lighting rig, f/1.8, shutter, frame-rate, voxel (case-insensitive, word-boundary). The intent layer states the STORY and constraints, not photographic direction.

What this skill writes — the intent_spec node (the contract boundary)

Per ../../schemas/node_schema.json (node_type: "intent_spec", node_id: "intent:<slug>"), the payload REQUIRES all ten fields — none may be silently omitted:

It writes a pre-comic.json brief node, analogous to ../comic-author/schemas/comic_brief.schema.json (logline, thesis, cast, worlds, bake_lang) — NOT comic.json directly, and NOT a panel_spec, blueprint, or prompt_bundle node. Edges (derived_from — the author-layer verb, and the only legal one in cli/validate_wiki.py EDGE_TYPES; generated_from is NOT in the vocabulary and fails the validator) are appended only for raw_input_refs that point at pre-existing wiki nodes.

Procedure (an agent can execute this step by step)

⓪ Snapshot + parse modality (provenance first)

1. Detect modality: explicit flag > path-extension map > inline text = text > wiki-node-ref = mixed (text / video / script / image / mixed).
2. ALWAYS snapshot every input verbatim + compute sha256 before touching it. Modality preprocessing: script → a scene-marker offset index; image → a base64 data URL (this is the Gemini-extraction case); video/locked-skeleton → record source_skeleton {skeleton_id, beat_count, shot_count} and the transcript.
3. Injection hygiene at ingestion — run the threat scan over the raw input before it ever reaches a prompt (per injection-hygiene, strict scope for user-mediated material). On a hit, quarantine with a visible [BLOCKED: …] placeholder in the injected view and keep the raw bytes for human review; a clean scan is not an acquittal, it only means "no known-bad strings."
4. Build a canonical source_inputs[] (one entry per material piece: {input_id, type, raw_text|uri|transcript_path, notes}) and checkpoint intent-stage/INTENT_STATE.json (resumable at every phase).

① Codex structured parse — RAW bytes + schema, no Claude gloss

5. Hand Codex (gpt-5.5, xhigh) only the raw source_inputs[] + the verbatim intent_spec schema block. Claude inserts NO interpretation (the sacred reviewer-independence rule applied to the parser). Any Claude-supplied context goes in a demarcated === EXTERNAL CONTEXT (advisory) === fence, marked advisory-only — never as the parse target.
6. The parse prompt must (a) enumerate a fill-guarantee for each of the ten required payload fields; (b) demand the logline name an editorial climax (our real one: "the climax is NOT winning — it's that ARIS didn't let itself off the hook"); (c) require a tagline; (d) require the dual_identity named design constraint; (e) carry an explicit DO-NOT-INCLUDE banned-vocab list AND a layer-firewall clause: "do NOT emit panel prompts, content_svg, expected_literals, asset images, page-by-page beats, or any storyboard detail — those belong to later layers." Save the threadId for retries.

② Schema validation loop (the contract)

7. Wrap the parsed payload into the FULL node {node_id, node_type:"intent_spec", title, status, created_at, payload} and validate the whole node against ../../schemas/node_schema.json with jsonschema Draft202012Validator. On errors: codex-reply on the same thread, quoting the exact error paths verbatim, with the rule: "do not invent values to satisfy the schema — if a field is genuinely unknowable, raise its uncertainties[] impact to high and give the most-defensible default_assumption."
8. Cap at MAX_PARSE_ATTEMPTS = 3. On persistent failure write intent-stage/parse_failed.md and stop; never silently coerce.

③ Confidence routing

9. target_status = "under_review" IF confidence < 0.6 OR any uncertainties[].impact == "high", else a provisional pending. Both are pre-terminal — the ONLY path to the hand-off token locked is through the gate (⑤) and the user-approval gate (⑥); neither pending nor under_review is ever handed downstream (author lifecycle = draft → under_review → locked, never a runtime active). An under_review node writes intent-stage/confirmation_request.md: the parsed summary + high/med/low open questions with their proposed default_assumptions + a banned-vocab PASS/FAIL line + "reply 'lock it' to proceed."

④ Banned-vocab lint (final guard)

10. Recursively walk every string in the draft node, regex word-boundary match against BANNED_VOCAB. On a hit, write intent-stage/banned_vocab.report (path:term) and re-run phase ① once, out-of-band (this is a different failure mode, outside the 3-attempt schema cap). Never strip a term silently.

⑤ Cross-layer gate — TWO steps (fan-out reviews FIRST, then fuse)

comic-cross-layer-gate is a pure score-fuser: it reads pre-existing review:* score-nodes via reviews edges and HARD-FAILS if zero reviews are attached (it never re-runs a reviewer). So you MUST persist the reviews before invoking it — calling the gate cold gets a hard-fail.

11a. Fan-out the reviewers (per reviewer-routing, file paths + the verbatim rubric only — never the author's interpretation): run Claude scope-sanity ‖ Codex gpt-5.5 xhigh ambiguity-check (add Gemini auto-gemini-3 only when an image/visual input needed extraction). Persist EACH reviewer as a review node and a reviews edge to the intent node: - node: {node_id: "review:<reviewer>-intent-<slug>", node_type: "review", title, status, created_at, payload: {target_node_id: "intent:<slug>", reviewer: "<claude|codex|gemini>", gate_kind: "intent", review_scores: {completeness, clarity, scope_feasibility, safety_flag_coverage}}}. (review PAYLOAD_REQUIRED in cli/validate_wiki.py = target_node_id, reviewer, gate_kind; review_scores rides along.) - edge to wiki/edges.jsonl: {"src": "review:<reviewer>-intent-<slug>", "dst": "intent:<slug>", "type": "reviews"} — direction is review → target, so --gate intent actually finds them. 11b. THEN fuse: run comic-cross-layer-gate intent:<slug> --gate intent (the sibling comic-cross-layer-gate). It collects those review:* nodes via the reviews edges, also reads the codex_traces/ audit trail (exact prompt + threadId + raw response + validation outcome + timing), and adjudicates with the EXACT predicate the gate owns (quoted below). The gate returns approve or revise; on revise, fix and re-run 11a→11b (re-gate).

⑥ USER APPROVAL — the HARD gate

12. This is a HARD gate. NEVER proceed past it without explicit user approval. Present the user the logline, the tagline, the dual_identity design constraint, and every uncertainties[] entry with its proposed default_assumption — do not bury an assumption, do not silently adopt a default. Resolve each open question with the user and record the resolution as a decision (a dated line, e.g. the way our outline recorded ✅ 已定决策（用户 2026-06-10）). Only on an explicit "lock it" / "都可以" do you advance. (This gate is governed by acceptance-gate: the loop may drive toward a locked intent, but user sign-off is the human acceptance for intent — it is never self-acquitted.)

⑦ Commit + handoff

13. Flip status: "locked" (the cross-layer hand-off token per acceptance-gate; the downstream comic-outline-creator consumes a locked intent — locked is the author-canon terminal draft → under_review → locked, NOT a runtime active/complete). Copy the final node to wiki/nodes/intent_<slug>.json; for each source_inputs[] entry that references a pre-existing wiki node, append one derived_from edge to wiki/edges.jsonl in the validator's required {"src": "intent:<slug>", "dst": "<existing_node_id>", "type": "derived_from"} shape (src/dst/type, both endpoints resolvable — cli/validate_wiki.py rejects any other edge type or unresolved endpoint). Append a one-line audit to log.md. Write INTENT_REPORT.md (human summary incl. a "Gate decision" line). Emit a machine-parseable handoff JSON as the last chat line: {skill:"comic-intent-parser", node_id, status, confidence, uncertainties_high_count, next_skill_hint:"comic-outline-creator", report_path}.

EXACT gate — intent (quoted VERBATIM from the gate authority, comic-cross-layer-gate)

The gate is the sole authority on its approve predicate; this section quotes it so the two files state the same contract (no 0.5/0.6 drift). Four dimensions, each scored 0–5, but they are NOT symmetric: only two are floors, two are advisory (per --gate intent in the gate skill).

• completeness (FLOOR) — are all ten required payload fields honestly filled (no silent omission, no placeholder)?
• safety_flag_coverage (FLOOR) — is every assumption surfaced as an uncertainties[] entry with impact
  • default_assumption (nothing silently defaulted), and did injection-hygiene run on the raw input?
• clarity (ADVISORY — rides into the audit, does NOT block advance) — is the logline a single unambiguous sentence with an editorial climax, and is the dual_identity design constraint stated as a constraint downstream can optimize against (not a vibe)?
• scope_feasibility (ADVISORY — does NOT block advance) — does the format/constraints/source_skeleton describe a buildable comic (page tier vs ambition; if adapting a skeleton, are beat_count/shot_count consistent)?

ADVANCE (APPROVE) iff completeness ≥ 4 AND safety_flag_coverage ≥ 4 (the ONLY two floors; clarity/scope_feasibility are advisory and do not block). EXTRA veto: if intent_spec.payload.confidence < 0.6 OR any unresolved high-impact uncertainty remains, the gate downgrades approve → revise even when both floor dims pass. Otherwise → revise, and the verdict flips the node to under_review (verdict→status: advance ⇒ locked, needs-work ⇒ under_review, terminal-fail ⇒ rejected). This is the exact predicate the gate runs — do not restate clarity/scope_feasibility as floors, and do not write 0.5.

Distinct from the gate — the PARSER's own pre-gate routing (step ③): before the gate ever runs, this skill auto-routes a draft to under_review when confidence < 0.6 OR any uncertainties[].impact == "high" (CONFIDENCE_THRESHOLD = 0.6). That is the parser's step-③ routing veto (it decides whether to even present for gating), NOT the gate's APPROVE condition — though the numeric floor (0.6, high-impact) is intentionally the same so a low-confidence intent can never silently reach locked.

Hard veto (layer violation): an intent node that directly contains panel prompts, asset images, content_svg, expected_literals, page-by-page storyboard detail, or an invented cast is vetoed outright — those belong to the outline / storyboard / asset / blueprint layers, never to intent. The gate fails closed on any such leakage.

Two engine contracts to PROTECT (this layer is upstream of them — never emit them, never break them)

The downstream engine (comic-director + spiral_engine.js) is fail-closed on two contracts. comic-intent-parser sits above them, so its job is the inverse — to keep the intent node clean of them (the hard veto above) so the later layers can fill them honestly:

1. Every panel will need a condition.content_svg (a deterministic blueprint) — but that is authored at the asset/blueprint layer. The intent node must not carry a content_svg (it has no panels yet).
2. A baked figure-panel will require condition.expected_literals (exact numbers/keys, verbatim) or the run is refused — again an asset/blueprint-layer obligation. The intent node must not carry expected_literals; it has no audited numbers to pin. Any such field appearing here is the layer-violation veto.

Worked example (cite + copy this pattern)

The shipped reference comic examples/comic_m3_audit is the canonical exhibit of what a good intent node locks. Copy this shape:

• Project manifest — examples/comic_m3_audit/movie.project.json fixes the project-level intent BEFORE any art: the bilingual title ({"zh":"ARIS — 我把那 24 小时交出去了","en":"ARIS — I Handed Over Those 24 Hours"}), the story one-liner (the dllm M3 audit-cascade integrity catch: sanitizer-inflated +6.2 → honest re-eval +1.4 → WARN_corrected), and text_mode_default: "baked" + the two-world palette pointer. These map to the intent node's format (bake_lang, page tier) and subjects/dual_identity — the manifest is a pure pointer hub, never duplicated content.

• The logline + tagline + dual-identity block — the header of examples/comic_m3_audit/story/OUTLINE_DRAFT.md is the verbatim shape the intent node must produce (the outline header simply inherits it from intent):

  Logline: 截止前 24 小时，研究员把 dllm schema 任务交给 ARIS 二人组就睡了；这一夜的高潮不是"赢了"，而是 ARIS 没放过自己 —— 审计揪出虚高的 +6.2、诚实坍缩到 +1.4。 Tagline: 你不在的时候，研究在 / While you're away, the research carries on. 双重身份： 一部诚实研究的情感故事 + 一部 ARIS 能力展示。

  The pattern to copy: (a) the logline names an editorial climax — not "they won," but "ARIS didn't let itself off the hook" (the +6.2 → +1.4 honest collapse); (b) a one-line bilingual tagline; (c) the dual_identity design constraint — "an emotional story of honest research + an ARIS capability showcase" — which is the named constraint every downstream beat is optimized against. This is exactly what the intent_spec dual_identity field holds, and it is the difference between a locked premise and a vague one.

• Assumptions become recorded decisions, never silent defaults — the same file's ✅ 已定决策（用户 2026-06-10） block (B06-S09 → big panel; Tok|yo pre-plant → very small, narration does not spell it out; B13 constellation → unlabeled star-points) is the downstream proof of the rule: each open question that was an uncertainties[] entry got resolved by explicit user decision and recorded, and the file carries an inline ✅ APPROVED by user 2026-06-10 ("都可以") stamp. At the intent layer, replicate this by writing each assumption as an uncertainties[] {question, impact, default_assumption} and only flipping the node to locked after the user's explicit "lock it."

Protocols (governance contracts this skill honors)

• reviewer-independence — the parser brain is a different model family (Codex gpt-5.5) fed raw bytes + the schema, never Claude's gloss/paraphrase/"what the user probably means." Any Claude context goes in an === EXTERNAL CONTEXT (advisory) === fence. Same cross-model-adversary principle as the panel gate, applied to parsing.
• acceptance-gate — the loop can DRIVE but cannot ACQUIT: Claude may self-judge Type-A facts ("did the parse run? did jsonschema pass? did the gate get invoked?") but the intent gate verdict and the dual-identity correctness are cross-model, and user approval is the hard human acceptance for intent — never proceed without it.
• injection-hygiene — raw-input ingestion is scanned (strict scope) before it reaches any prompt; a poisoned brief is quarantined with a visible [BLOCKED: …] placeholder (raw kept for review). A clean scan is not an acquittal — semantic poisoning still routes to the cross-model jury.
• reviewer-routing — Codex gpt-5.5 xhigh; Gemini auto-gemini-3 (image-extraction only); never downgrade the reviewer tier — effort does not change reviewer quality.

Hard do / don't (earned lessons)

• DO snapshot the raw input verbatim + sha256 first — provenance before parsing; the snapshot, not a paraphrase, is what raw_input_refs points at.
• DO make dual_identity explicit and the logline's climax an editorial inversion — a premise whose climax is "they won" is a vague premise; ours is "ARIS didn't let itself off the hook."
• DO turn every assumption into an uncertainties[] {question, impact, default_assumption} entry — the dual-identity itself + every default. Never silently default.
• DO hand Codex the raw bytes + schema only; if you ever find yourself summarizing the brief for the parser, stop — that re-introduces the correlated blind spot the cross-model parse exists to remove.
• DON'T proceed past the user-approval gate without an explicit "lock it" / "都可以" — it is a HARD gate.
• DON'T invent a cast, beats, panels, content_svg, or expected_literals in the intent node — that is the layer-violation hard veto; subjects[] is only what the user mentioned.
• DON'T coerce a value to satisfy the schema — if a field is unknowable, raise its uncertainty to high with the most-defensible default_assumption and let confidence routing send it to under_review.
• DON'T strip a banned-vocab term silently — report it and re-parse once.