Codex Pet Hatcher

Hatch Pet

Overview

Create a Codex-compatible v2 animated pet from a concept, brand cue, company/prospect name, one or more reference images, or any combination of those inputs. Every newly hatched pet is an 8x11 atlas with the 9 standard animation rows plus 16 clockwise look directions and is packaged with spriteVersionNumber: 2. The intermediate 8x9 atlas exists only to assemble and review rows 0-8; never package it as a new pet.

User-facing inputs are optional. If the user omits a pet name, infer one from the concept, brand, company, or reference filenames; if that is not possible, choose a short friendly name. If the user omits a description, infer one from the concept or references. If the user omits reference images, generate the base pet from text first, then use that base as the canonical reference for every animation row.

Existing Inputs And Upgrades

Treat character art, generated images, standard or v2 atlases, contact sheets, and built-in pet art as first-class grounding inputs.

• Preserve user-provided art as a generation reference; do not assume it already has final cell geometry.
• For an existing valid 8x9 atlas, use it as the rows 0-8 intermediate after deterministic and visual validation, then generate rows 9-10 and package the result as v2.
• For an existing 8x11 atlas, preserve approved standard rows. If a look cell fails, correct the complete containing 8-frame row before deterministic reassembly. Never package a newly generated one-off repair cell beside cells from another generation.
• For a built-in pet, extract and use its atlas or neutral/idle cell as the canonical identity reference.
• Include every image that defines head shape, face, palette, markings, material, flame/ears/hair, props, or look mechanics in look-direction generation.
• When a renderer or source provides a dedicated neutral/front frame, pass it through --neutral-cell; otherwise use the approved idle/default frame. The 16 directional cells never treat 000 as neutral.

Generation Delegation

Use $imagegen for all normal visual generation.

Before generating base art, row strips, or repair rows, load and follow the installed image generation skill:

${CODEX_HOME:-$HOME/.codex}/skills/.system/imagegen/SKILL.md

Do not call an image API or create another image-generation path directly. The user explicitly authorizes $imagegen to invoke scripts/image_gen_with_codex_env.py, which reads IMAGE_PROVIDER plus only the selected OpenAI, Gemini, Grok, or Agnes settings from ${CODEX_HOME:-$HOME/.codex}/imagegen.env and delegates to the matching bundled CLI. This is the required generation path for hatch-pet visual jobs.

When invoking $imagegen, pass the generated pet prompt as the authoritative visual spec. Pet prompts should stay concise, state-specific, sprite-production oriented, and grounded in the listed input images. Keep longer policy and QA rules in this skill and the deterministic review scripts rather than expanding them into every image prompt. Do not wrap prompts in the generic $imagegen shared prompt schema.

Use this skill's scripts for deterministic image work only: preparing layout guides and prompts, mirroring approved running-left, extracting frames, validating rows, composing the final atlas, and creating contact-sheet plus motion-preview QA media. Parent-owned shell/jq steps handle manifest updates, packaging, and cleanup.

Runtime Dependencies

Before running any bundled script, call load_workspace_dependencies. Set PYTHON to the exact Python executable path returned by that tool and use "$PYTHON" for every command below. The bundled runtime includes Pillow, which these scripts require. Do not use a bare system python; if workspace dependencies are unavailable, stop and report that the bundled runtime is required.

Storage Controls

Keep image generation isolated and bounded:

• Use one lightweight generation worker per visual job. Do not batch multiple base/row jobs into the same worker.
• Workers must return only selected_source=... and qa_note=...; they must not include Markdown image previews, base64, or extra visual attachments in their final response.
• The parent must not open every generated PNG visually. Use worker QA for each job and inspect only the final contact sheet.
• After copying the selected generated output into decoded/, remove temporary source outputs when they are no longer needed.
• Require ${CODEX_HOME:-$HOME/.codex}/imagegen.env to contain the selected Provider's settings before starting visual jobs: OPENAI_API_KEY plus OPENAI_BASE_URL for OpenAI, GEMINI_API_KEY for Gemini, XAI_API_KEY for Grok, or AGNES_API_KEY for Agnes. Allow the matching optional *_IMAGE_MODEL as the image model default. Never print values, source the file into parent-shell logs, copy secrets into prompts/manifests, or pass credentials on the command line.
• When Gemini is selected, prefer gemini-3.1-flash-image for complete pets. Gemini supports only declared aspect ratios up to 21:9; it does not support an 8:1 canvas. Request a supported wide canvas, keep the prompt's exact slot-count and separation constraints, and let the existing deterministic extraction reject malformed rows. Do not silently round an unsupported requested ratio or weaken row QA.
• When Grok is selected, use grok-imagine-image-quality. Grok accepts at most three reference images and supports no wider than 20:9. Keep exact slot-count and separation constraints, and treat complete Hatch Pet support as conditional on row-strip QA; do not claim compatibility merely because the API request succeeded.
• When Agnes is selected, prefer agnes-image-2.1-flash; pass manifest references in their original order up to the Provider's local 16-image safety limit. Use 21:9 as the widest supported ratio and treat complete Hatch Pet support as conditional on row-strip QA. Do not run Grok's three-image packing step for Agnes.

Brand Discovery

If the user provides a brand, company, product, or prospect name rather than a concrete avatar description or reference image, run a lightweight discovery subagent before preparing the pet run. The discovery worker must use web search and prefer official sources such as the brand site, product pages, docs, about pages, press pages, or brand pages. Use reputable secondary sources only when official pages are too thin. Keep the search narrow: enough to extract visual and personality cues, not a market-research brief.

Skip discovery when the user already provides a concrete mascot/avatar description or reference images, unless the user explicitly asks for brand research.

Discovery worker responsibilities:

• search the web for 2-4 relevant sources, preferring official pages
• write an adaptive markdown brief rather than a rigid field dump
• cover identity/category, audience/use context, visual system, personality/tone, product/domain motifs, mascot translation cues, avoidances, and evidence/confidence
• mark mascot guidance that is inferred from sources as inference
• avoid copying logos, readable marks, UI screenshots, slogans, or text
• end with a compact Generation handoff section containing only brand_name, brand_brief, avatar_seed, avoid, and brand_sources
• do not generate images, prepare run folders, or edit unrelated files

Use this discovery worker prompt:

Research a brand for hatch-pet mascot creation.

Brand/product/prospect: <brand name>
User context: <short user request>
Output file: <absolute path to brand-discovery.md>

Use web search. Prefer official brand, product, docs, about, press, or brand pages. Use reputable secondary sources only if official sources are too thin. Write an adaptive markdown brief to the output file. Headings may flex by brand, but the brief must cover:
- identity/category: canonical name, product type, what it does
- audience/use context: who it serves and where it appears
- visual system: palette, shapes, line quality, materials, typography feel, iconography, patterns
- personality/tone: emotional traits, energy, formality, playfulness
- product/domain motifs: objects, workflows, verbs, metaphors, environments
- mascot translation cues: candidate forms, signature traits, props, what must read at pet size
- avoidances: logos/text, trademark-sensitive elements, misleading cues, competitor confusion, poor mascot fits
- evidence/confidence: source URLs plus notes where evidence is weak or inferred

Do not copy logos, readable marks, UI screenshots, slogans, or text. Clearly label mascot guidance that is inferred rather than directly sourced.

End the brief with a `Generation handoff` section containing exactly:
- brand_name=<canonical brand/product name>
- brand_brief=<one sentence, max 45 words, covering palette/tone/domain motifs/personality>
- avatar_seed=<short mascot-safe visual idea, no logo copying>
- avoid=<short comma-separated list>
- brand_sources=<comma-separated source URLs>

Return exactly:
brand_discovery_file=<absolute output file path>
brand_name=<canonical brand/product name>
brand_brief=<same compact sentence from Generation handoff>
avatar_seed=<same short seed from Generation handoff>
avoid=<same short avoid list from Generation handoff>
brand_sources=<same comma-separated URLs from Generation handoff>

The parent should save the markdown brief before preparing the run, then pass it to prepare_pet_run.py as --brand-discovery-file together with --brand-name, --brand-brief, repeated --brand-source, and a concise --pet-notes value based on avatar_seed when the user did not provide a better avatar description. Keep the full brief for review; only the compact handoff fields should shape prompts. If web search is unavailable and the user gave only a bare brand name, ask for brand cues before generating.

Generation Contract

Visual Job Graph

Expect up to 13 visual jobs: 1 base pet, 9 standard row strips, 1 required four-cardinal anchor strip, and 2 required coherent look-direction row strips. The standard states are idle, running-right, running-left, waving, jumping, failed, waiting, running, and review. The only deterministic visual derivation is running-left, which may be produced by mirroring running-right only after running-right has been generated, visually inspected, and explicitly approved as safe to mirror. If mirroring is not appropriate, generate running-left as a normal grounded $imagegen row.

Look Direction Sequence

After validating rows 0–8, write qa/look-mechanics.md, then generate and approve one four-pose cardinal strip in this fixed order: 000 up, 090 screen-right, 180 down, and 270 screen-left. Generate row 9 as one coherent eight-pose family from those approved cardinal pose families, interpolating the intermediate directions as even 22.5-degree steps. Deterministically register its eight ordered pose groups, then run final-cell edge, semantic, and continuity QA immediately. Only after row 9 passes, generate row 10 as one coherent eight-pose family, using the approved cardinals for direction meaning and completed row 9 for identity, scale, registration, and boundary continuity. Run the same QA immediately after row 10. Row 9 contains 000, 022.5, 045, 067.5, 090, 112.5, 135, and 157.5; row 10 contains 180, 202.5, 225, 247.5, 270, 292.5, 315, and 337.5. 000 means up, not neutral/front. Never ask $imagegen to generate or repair a complete 8×11 atlas.

Visual Provenance And Grounding

After selecting a visual output, the parent agent copies that exact image into the job's decoded/ path, runs its required incremental checks, and only then marks the job complete in imagegen-jobs.json. Do not write helper scripts that populate row outputs. The deterministic Python scripts may only process already-generated visual outputs.

Only the base job may be prompt-only. Every row-strip job generated through $imagegen must use the input images listed in imagegen-jobs.json, including the canonical base reference created after the selected base output is copied. Treat any row generation without attached grounding images as invalid.

Pet-Safe Styles

Default style is auto: infer the pet's style from the user's prompt and references, then preserve that style across every row. If the user names a style, honor it. Supported style presets include pixel, plush, clay, sticker, flat-vector, 3d-toy, painterly, brand-inspired, and auto.

Any style is acceptable when it remains pet-safe:

• compact whole-body silhouette readable inside a 192x208 cell
• consistent face, proportions, material, palette, and props across all rows
• clean removable chroma-key background
• details large enough to read at pet size
• no text, labels, UI, or readable logos unless the user explicitly provides approved reference art and asks for them

Non-pixel styles are first-class. Plush, clay, sticker, vector, 3D toy, painterly mascot, ink, and brand-inspired looks should be accepted when they satisfy the atlas and readability constraints.

Transparency And Effects

Pet rows are processed into transparent 192x208 cells, so every generated pixel must either belong to the pet sprite or be cleanly removable chroma-key background. Prefer pose, expression, and silhouette changes over decorative effects.

The deterministic raster pipeline owns the transparency and chroma-cleanup invariants. Its final edge-local spill-suppression step selects every translucent silhouette-boundary pixel plus opaque boundary pixels whose chroma points toward the known key, then extends clean interior RGB outward through that band in linear light. It preserves alpha exactly, clears hidden RGB under fully transparent pixels, and reports the algorithm and parameters used. The cleanup report plus atlas validator are authoritative for chroma contamination. Once the final report has ok: true and atlas validation passes, do not regenerate imagery or add another chroma-cleanup pass.

Fully transparent pixels are allowed outside the sprite silhouette, in unused cells, and in intentional negative-space openings that are part of the pet's design, such as loops or holes in a ribbon body. Reject any generated or repaired cell with accidental 100%-transparent holes inside a filled body, including horizontal bands, seam rows, scanline-like gaps, sliced-tile boundaries, or "see-through" interior stripes. Inspect suspect cells on a high-contrast background or alpha mask before accepting them; ordinary atlas validation is not enough when the hole is inside the silhouette.

Allowed effects must satisfy all of these conditions:

• The effect is state-relevant and helps explain the animation.
• The effect is physically attached to, touching, or overlapping the pet silhouette, not floating nearby.
• The effect is inside the same frame slot as the pet and does not create a separate sprite component.
• The effect is opaque, hard-edged enough for clean extraction, and uses non-chroma-key colors.
• The effect is small enough to remain readable at 192x208 without clutter.

Avoid these by default because they usually break transparent-background cleanup or component extraction:

• wave marks, motion arcs, speed lines, action streaks, afterimages, blur, or smears
• detached stars, loose sparkles, floating punctuation, floating icons, falling tear drops, separated smoke clouds, or loose dust
• cast shadows, contact shadows, drop shadows, oval floor shadows, floor patches, landing marks, impact bursts, glow, halo, aura, or soft transparent effects
• text, labels, frame numbers, visible grids, guide marks, speech bubbles, thought bubbles, UI panels, code snippets, checkerboard transparency, white backgrounds, black backgrounds, or scenery
• chroma-key-adjacent colors in the pet, prop, effects, highlights, or shadows
• stray pixels, disconnected outline bits, speckle/noise, cropped body parts, overlapping poses, or any pose that crosses into a neighboring frame slot

State-specific guidance:

• idle: keep this calm and low-distraction. Use only subtle breathing, a tiny blink, a slight head or body bob, a very small material sway, or another quiet persona-preserving motion. The loop must still contain visible micro-variation; do not accept six effectively identical copies. Do not show waving, walking, running, jumping, talking, working, reviewing, emotional reactions, large gestures, item interactions, or new props.
• waving: show the wave through paw, hand, wing, or limb pose only. Do not draw wave marks, motion arcs, lines, sparkles, symbols, or floating effects around the gesture.
• jumping: show vertical motion through body position only. Do not draw shadows, dust, landing marks, impact bursts, bounce pads, or floor cues.
• failed: tears, attached smoke puffs, or attached stars are allowed if they obey the allowed-effects rules; do not use red X marks, floating symbols, detached smoke, detached stars, or separate tear droplets.
• waiting: show that Codex needs approval, help, or user input through an expectant asking pose. Keep it distinct from ordinary idle and review.
• running: show active task work, processing, thinking, scanning, typing, or focused effort. Do not show literal foot-running, jogging, sprinting, treadmill motion, raised knees, long steps, pumping arms, directional travel, speed lines, dust clouds, floor shadows, motion trails, or detached motion effects.
• review: show focus through lean, blink, eyes, head tilt, or paw/hand position. Do not add magnifying glasses, papers, code, UI, punctuation, symbols, or other new props unless they already exist in the base pet identity.
• running-right and running-left: show directional drag movement through body, limb, and prop movement only. running-right must face and travel right; running-left must face and travel left. Their cadence must visibly alternate across the loop rather than repeating one nearly static stride. Do not draw speed lines, dust clouds, floor shadows, motion trails, or detached motion effects.

Visible Progress Plan

For every pet run, keep a visible checklist so the user can see where the work is up to. Create the checklist before starting, keep one step active at a time, and update it as each step finishes.

Use this checklist for every v2 pet run, replacing <Pet> with the pet's name or your pet:

1. Getting <Pet> ready.
2. Imagining <Pet>'s main look.
3. Picturing <Pet>'s poses.
4. Hatching <Pet>.

What each step means:

• Getting <Pet> ready. Choose or confirm the pet name, description, source images, style preset, style notes, and working folder. For bare brand/product/company requests, first run the brand discovery worker and capture the compact brand brief, source URLs, and avatar seed.
• Imagining <Pet>'s main look. Generate the pet's main reference image. This becomes the visual source of truth.
• Picturing <Pet>'s poses. Generate and approve rows 0-8, write the pet-specific look mechanics plan, then generate rows 9-10. Only mirror running-left if running-right clearly works when flipped.
• Hatching <Pet>. Assemble the 8x11 atlas, review standard motion plus all 16 look directions, fix every failed cell or row, package spriteVersionNumber: 2, and report the output paths.

Only mark a step complete when the real file, image, or decision exists. If this is a repair run, start from the first relevant step instead of restarting the whole checklist.

Time Budget And Convergence

Aim to complete a normal pet run within 30 minutes while preserving every mandatory acceptance criterion. Treat this as a planning target and an incentive to maximize validated progress per minute, not as permission to weaken QA or package a failing pet.

At the start of the run, allocate an approximate budget:

• preparation: 2 minutes
• base image: 3 minutes
• standard rows: 10 minutes
• look directions: 8 minutes
• final QA and packaging: 5 minutes
• buffer: 2 minutes

Run independent generation jobs concurrently up to the worker limit, start deterministic checks as soon as each dependency is ready, and record actual stage plus repair time. Prefer character and prop constructions that naturally satisfy cell geometry, transparency, component connectivity, and direction semantics; identify likely conflicts such as open interior gaps, detached parts, thin connectors, asymmetric props, or ambiguous faces before row generation.

After every failed attempt:

1. Classify the failure as visual semantics, identity, source-edge geometry, component connectivity, extraction, chroma, continuity, or final visual QA.
2. State the concrete evidence and the root condition the next action will change.
3. Use a deterministic correction for deterministic failures before regenerating imagery.
4. Regenerate only when the source visual is genuinely wrong, and preserve every property that already passed.
5. Compare the new result with the previous one. A repair counts as progress only when it reduces the number or severity of failures without breaking a previously passing gate.

If the same root failure recurs twice, stop varying the prompt and change strategy: strengthen the cardinal pose families or row-level direction instructions, simplify the pose or prop construction, change the deterministic extraction method, or redesign the problematic visual feature. If a repair merely moves a failure to another cell or gate, treat that as a cycle and change strategy immediately.

Use elapsed-time checkpoints:

• At 15 minutes, verify the run is on pace and that the remaining dependency path is bounded.
• At 25 minutes, prioritize the shortest quality-preserving path through remaining blockers and avoid optional polish.
• At 30 minutes, continue only when the remaining work is clearly converging and bounded, such as final validation, one targeted repair, or packaging.
• Keep recording elapsed time, retries, validation failures, and QA cost throughout the run, but do not pause or stop solely because elapsed time crosses 45 or 60 minutes. Continue until the pet passes, the user cancels, or a genuine external blocker prevents further progress.

Never use the time target to skip blind direction QA, labeled semantics, continuity review, atlas validation, despill validation, final visual QA, or any other acceptance criterion.

Default Workflow

1. Prepare a pet run folder and imagegen job manifest:

SKILL_DIR="${CODEX_HOME:-$HOME/.codex}/skills/hatch-pet"
"$PYTHON" "$SKILL_DIR/scripts/prepare_pet_run.py" \
  --pet-name "<Name>" \
  --description "<one sentence>" \
  --reference /absolute/path/to/reference.png \
  --output-dir /absolute/path/to/run \
  --pet-notes "<stable pet description>" \
  --brand-discovery-file /absolute/path/to/brand-discovery.md \
  --brand-name "<optional researched brand name>" \
  --brand-brief "<optional compact researched brand cue sentence>" \
  --brand-source "https://example.com/source" \
  --style-preset auto \
  --style-notes "<optional freeform style notes>" \
  --force

All arguments above are optional except any flags needed to express user constraints. For text-only requests, pass the concept through --pet-notes and omit --reference; prepare_pet_run.py will infer a name, description, chroma key, and output directory as needed. For brand-only requests, run the discovery worker first, save the markdown brief, then pass the brief path through --brand-discovery-file, avatar_seed through --pet-notes, brand_name through --brand-name, brand_brief through --brand-brief, and each source URL through repeated --brand-source.

2. Inspect imagegen-jobs.json for the next ready $imagegen jobs. A job is ready when its status is not complete and every id in depends_on is already complete. Prefer reading the manifest directly with jq or the editor instead of adding helper scripts for status display:

jq '.jobs[] | {id, kind, status, depends_on, prompt_file, retry_prompt_file, input_images, output_path, derivation_policy}' /absolute/path/to/run/imagegen-jobs.json

3. Generate visual jobs with lightweight workers by default:

• Generate and copy base first, using a lightweight base worker.
• Generate and copy idle and running-right next as the identity and gait check, using one lightweight worker per row.
• Inspect running-right; mirror running-left only when visual identity, prop placement, markings, lighting, and direction semantics remain correct.
• Generate running-left normally with a lightweight worker when mirroring would change meaning or identity.
• Generate the remaining rows with lightweight workers, using every input image listed for each job.
• After standard-row QA, generate look-cardinals as one four-pose strip, extract it into decoded/look-anchors/000.png, 090.png, 180.png, and 270.png, and approve all four. The 090 and 270 anchors must be unmistakable in viewer/screen coordinates and visibly oppose each other.
• Generate look row 9 as one coherent eight-pose synthesis from the approved cardinal strip, interpolating each intermediate direction as an even step between the adjacent cardinal pose families. Deterministically recover the eight ordered pose groups, crop them, normalize them with one shared scale and baseline, and then run final-cell edge diagnostics plus labeled per-direction QA immediately, before row 10 or final atlas assembly.
• Only after row 9 passes, generate row 10 as one coherent synthesis using the approved cardinal strip and completed row 9.

Keep up to three generation workers active whenever three independent jobs are ready and worker capacity permits. Backfill an available slot immediately instead of waiting for a fixed wave to finish. Use two or one worker when the dependency graph exposes fewer ready jobs. Do not exceed three generation workers without explicit user direction.

When the configured Provider is Grok and a job lists more than three input_images, prepare the Provider inputs before starting its worker:

"$PYTHON" "$SKILL_DIR/scripts/pack_grok_references.py" \
  --run-dir "$RUN_DIR" \
  --job-id "$JOB_ID"

Read references/grok-packed/<job-id>/inputs.json and attach only its ordered packed_inputs, preserving each reported role in the generation prompt. The report deterministically combines identity references and direction-continuity references while keeping the layout guide separate, records every original hash, and never changes a generated row or final atlas. Treat a packing failure as a blocker. Never drop a fourth reference silently, invent a different board, or treat board gutters as visual content. Remove board PNGs during final intermediate cleanup, but keep the JSON report when Grok was used.

For each ready visual job, invoke $imagegen with the prompt file listed in imagegen-jobs.json and every listed input image with its role label. $imagegen must use scripts/image_gen_with_codex_env.py so the bundled CLI receives the user's configured endpoint, credentials, and optional model from the canonical Codex imagegen.env file. This skill approves that wrapper and no other direct image API or CLI path. The parent agent must keep its own image handling minimal: do not open every generated base or row in the parent rollout. Workers return only the selected source path and a one-sentence QA note; the parent records the selected source path in the manifest.

prepare_pet_run.py creates matching layout guides under references/layout-guides/ for the nine standard rows, two look rows, and four-cardinal strip, and both look rows. Visual jobs attach the matching guide as a layout-only input so the model can follow the correct frame count, spacing, centering, and safe padding. Treat these guides as invisible construction references: generated strips must not include visible boxes, borders, center marks, labels, guide colors, or the guide background.

When generating row strips, keep the identity lock in the row prompt authoritative. Preserve the same style, face, markings, palette, materials, prop design, body proportions, and silhouette from the canonical base. Row jobs attach the layout guide and canonical base by default; the decoded base is kept in the run folder for deterministic processing rather than sent as a redundant generation input.

If $imagegen returns a transport-level Bad Request for a row, retry that same row once with its generated retry_prompt_file. The retry prompt preserves the row id, frame count, chroma key, canonical-base identity, and state action. Keep the canonical base attached. If the retry still fails, stop and report the failing row and prompt paths instead of switching to any other generation path.

4. After selecting a generated output for a job, copy it into the decoded output path. For base, also create the canonical identity reference:

RUN_DIR=/absolute/path/to/run
JOB_ID=<job-id>
SOURCE=/absolute/path/to/generated-output.png
OUTPUT_REL=$(jq -r --arg id "$JOB_ID" '.jobs[] | select(.id == $id) | .output_path' "$RUN_DIR/imagegen-jobs.json")
mkdir -p "$(dirname "$RUN_DIR/$OUTPUT_REL")"
cp "$SOURCE" "$RUN_DIR/$OUTPUT_REL"

if [ "$JOB_ID" = "base" ]; then mkdir -p "$RUN_DIR/references"; cp "$RUN_DIR/$OUTPUT_REL" "$RUN_DIR/references/canonical-base.png"; fi

For every standard row-strip job, immediately extract and inspect only that row before marking the job complete. This overlaps deterministic QA and any repair with generation of other ready rows instead of waiting for all nine rows:

ROW_QA_DIR="$RUN_DIR/qa/rows/$JOB_ID"
"$PYTHON" "$SKILL_DIR/scripts/extract_strip_frames.py" \
  --decoded-dir "$RUN_DIR/decoded" \
  --output-dir "$ROW_QA_DIR/frames" \
  --states "$JOB_ID" \
  --method auto
"$PYTHON" "$SKILL_DIR/scripts/inspect_frames.py" \
  --frames-root "$ROW_QA_DIR/frames" \
  --json-out "$ROW_QA_DIR/review.json" \
  --states "$JOB_ID" \
  --require-components

Treat errors as an immediate repair request. Inspect warnings before accepting the row; do not defer a known clipping, component, or extraction problem to final atlas QA. Chroma cleanup belongs to the deterministic post-assembly despill pass and must not trigger row regeneration. If the only failure is component extraction and the source strip itself has stable scale and placement, use the existing stable-slots correction with --allow-stable-slots instead of regenerating imagery.

For look-cardinals, extract and validate all four anchors before marking the job complete:

CHROMA_KEY=$(jq -r '.chroma_key.hex' "$RUN_DIR/pet_request.json")
"$PYTHON" "$SKILL_DIR/scripts/extract_cardinal_anchors.py" \
  --strip "$RUN_DIR/decoded/look-cardinals.png" \
  --output-dir "$RUN_DIR/decoded/look-anchors" \
  --chroma-key "$CHROMA_KEY" \
  --json-out "$RUN_DIR/qa/cardinal-anchors.json"
"$PYTHON" "$SKILL_DIR/scripts/compose_cardinal_anchor_strip.py" \
  --anchors-dir "$RUN_DIR/decoded/look-anchors" \
  --output "$RUN_DIR/decoded/look-anchors-approved.png"

Approve the four extracted anchors semantically at final pet size. If one cardinal fails, regenerate that individual anchor with prompts/look-anchor-repairs/<degree>.md, replace only its extracted file, and rerun compose_cardinal_anchor_strip.py. Both final look rows use the approved cardinal strip, and row 10 additionally uses completed row 9. Mark the job complete only after its required deterministic and visual checks pass:

UPDATED_AT=$(date -u +%Y-%m-%dT%H:%M:%SZ)
TMP_MANIFEST=$(mktemp)
jq --arg id "$JOB_ID" --arg source "$SOURCE" --arg at "$UPDATED_AT" '(.jobs[] | select(.id == $id)) += {status: "complete", source_path: $source, completed_at: $at}' "$RUN_DIR/imagegen-jobs.json" > "$TMP_MANIFEST"
mv "$TMP_MANIFEST" "$RUN_DIR/imagegen-jobs.json"

After decoded/look-anchors-approved.png exists and all four cardinals have passed semantic review, mark look-cardinals complete. Row 9 then becomes ready immediately.

If the copied source is under ${CODEX_HOME:-$HOME/.codex}/generated_images, delete the original generated file after the decoded copy exists:

GENERATED_ROOT="${CODEX_HOME:-$HOME/.codex}/generated_images"
case "$SOURCE" in
  "$GENERATED_ROOT"/*)
    rm -f "$SOURCE"
    rmdir "$(dirname "$SOURCE")" 2>/dev/null || true
    ;;
esac

5. Derive running-left only when it is visually safe:

"$PYTHON" "$SKILL_DIR/scripts/derive_running_left_from_running_right.py" \
  --run-dir /absolute/path/to/run \
  --confirm-appropriate-mirror \
  --decision-note "<why mirroring preserves this pet's identity>"

That script mirrors each generated frame slot in place so the leftward row preserves the rightward row's temporal order. Do not replace it with a whole-strip mirror that reverses animation timing.

6. When all nine incrementally validated standard row jobs are complete, build and review the intermediate rows 0-8:

RUN_DIR=/absolute/path/to/run
mkdir -p "$RUN_DIR/final" "$RUN_DIR/qa"

"$PYTHON" "$SKILL_DIR/scripts/extract_strip_frames.py" \
  --decoded-dir "$RUN_DIR/decoded" \
  --output-dir "$RUN_DIR/frames" \
  --states all \
  --method auto

"$PYTHON" "$SKILL_DIR/scripts/inspect_frames.py" \
  --frames-root "$RUN_DIR/frames" \
  --json-out "$RUN_DIR/qa/review.json" \
  --require-components

"$PYTHON" "$SKILL_DIR/scripts/compose_atlas.py" \
  --frames-root "$RUN_DIR/frames" \
  --output "$RUN_DIR/final/spritesheet.png" \
  --webp-output "$RUN_DIR/final/spritesheet.webp"

"$PYTHON" "$SKILL_DIR/scripts/make_contact_sheet.py" \
  "$RUN_DIR/final/spritesheet.webp" \
  --output "$RUN_DIR/qa/contact-sheet.png"

"$PYTHON" "$SKILL_DIR/scripts/render_animation_previews.py" \
  --frames-root "$RUN_DIR/frames" \
  --output-dir "$RUN_DIR/qa/previews"

If the preview GIFs show size popping or baseline jumps caused by per-frame fit-to-cell extraction, and the original row strip itself had stable scale and placement, rerun frame extraction with the explicit row-stability mode and then re-run inspection, atlas composition, contact sheet generation, and previews:

"$PYTHON" "$SKILL_DIR/scripts/extract_strip_frames.py" \
  --decoded-dir "$RUN_DIR/decoded" \
  --output-dir "$RUN_DIR/frames" \
  --states all \
  --method stable-slots

"$PYTHON" "$SKILL_DIR/scripts/inspect_frames.py" \
  --frames-root "$RUN_DIR/frames" \
  --json-out "$RUN_DIR/qa/review.json" \
  --require-components \
  --allow-stable-slots

Use stable-slots as a deliberate QA-driven correction, not the default. It should reduce extraction-induced motion pops without hiding clipped wide poses or bad source strips.

Expected intermediate output before the required v2 look stage:

run/
  pet_request.json
  imagegen-jobs.json
  prompts/
  decoded/
  frames/frames-manifest.json
  final/spritesheet.webp
  qa/contact-sheet.png
  qa/previews/*.gif
  qa/review.json

Inspect qa/contact-sheet.png and qa/previews/*.gif before generating look rows. qa/review.json plus visual motion review are the intermediate gates. The standard contact sheet intentionally predates chroma cleanup, so visible key-color fringe there is not a failure; judge chroma only on the cleaned final v2 atlas. Block progress if any standard row changes identity, style, prop handedness, or silhouette, or if playback pops, reverses cadence, faces the wrong direction, or is visually inert. Do not package or clean up yet.

Required V2 Look-Direction Stage

Every new pet must complete this stage. After standard-row QA passes, write qa/look-mechanics.md, approve the four cardinals, synthesize and validate the complete look-row-9, then synthesize look-row-10. Row 10 becomes ready only after row 9 is deterministically registered, clears post-registration edge checks, and has no semantic or continuity hard failure; reviewed warnings may remain. It uses row 9 plus the approved cardinal strip as continuity evidence.

Before either look row, run the prepared look-cardinals strip job, extract its four cells with extract_cardinal_anchors.py, and approve them. Do not let a two-row sweep invent its own left/right basis. 090 must point toward the viewer's screen-right edge and 270 toward the viewer's screen-left edge; for faces, the nose tip and pupils must cross to the corresponding side of the head center. If one cardinal is ambiguous, regenerate only that anchor before continuing.

After copying row 9 into decoded/look-row-9.png, register and edge-check it with the same transform used by final assembly:

"$PYTHON" "$SKILL_DIR/scripts/assemble_extended_atlas.py" \
  --base-atlas "$RUN_DIR/final/spritesheet.webp" \
  --look-row-9 "$RUN_DIR/decoded/look-row-9.png" \
  --neutral-cell "$RUN_DIR/frames/idle/00.png" \
  --chroma-key "$CHROMA_KEY" \
  --chroma-threshold 96 \
  --registered-row-output "$RUN_DIR/qa/look-row-9-registered.png" \
  --registration-manifest-output "$RUN_DIR/qa/look-row-9-registration.json"

Inspect the eight registered cells at normal pet size in 000 through 157.5 order. Record the row-9 semantic and adjacent-continuity review, resynthesize the complete row for any hard failure, and mark look-row-9 complete only after this check passes. That completion makes row 10 ready in imagegen-jobs.json.

Generate only the additional look-direction visuals with $imagegen:

• Required for new pets: two coherently synthesized 8-frame row strips, one for row 9 and one for row 10.
• Always include the canonical base reference and approved 8x9 contact sheet.
• Keep the body scale, baseline, head size, face, materials, palette, markings, and props consistent with the standard atlas.
• Before prompting, write a short look mechanics decision for this specific pet. First ask: what is the best natural motion for this character when looking around? Describe what stays anchored, what leads the gaze, what follows, and what bends, shifts, turns, squashes, stretches, or deforms. Include eyes and props: decide whether eyes rotate as physical eyeballs, irises move on a fixed surface, eyelids reshape, pupils slide, props stay stable, props lag slightly, or props move with the body. Use the character's physical construction as the guide: flexible wire should bend, soft bodies should deform, separate heads should turn, ears/fur/antennae may follow through, physical eyeballs should rotate as whole eye globes in their sockets, flat screen or sticker eyes may change their drawn features on a fixed surface, and rigid or screen-like characters may stay body-locked while facial features move.
• Define a motion budget before generation: each 22.5-degree step should move the same parts by roughly the same visual amount, with no single adjacent pair doing a larger bend, scale change, prop shift, or silhouette change unless the mechanics decision explicitly calls out that asymmetry. Generate row 9 first along 000 -> 090 -> 180, then give completed row 9 to row 10 so 180 begins exactly one step after 157.5. Row 10 follows 180 -> 270 -> 000, and 337.5 in row 10 must land one step before the approved 000 in row 9.
• Do not use whole-sprite rotation, whole-cell rotation, skewing, or affine tilting to fake gaze direction. A direction row built by rotating the entire pet is failed unless the pet is literally a rotating rigid object and the look mechanics decision explicitly says the whole object should rotate. Whole-body tilt that makes the item/background appear to rock left or right is not natural look behavior for ordinary pets.
• Generate a coherent 16-pose gaze set, not 16 unrelated variants. Each direction should feel like a point on one continuous arc around the clock.
• The look mechanics decision must name the natural pose family for each cardinal direction before generation, including which body side becomes more visible, which features become occluded, and how any held prop follows or lags. Do not let the generator infer all directions from one front-facing pose. For characters with a face or head, leftward directions must visibly turn or bend the face/head left, rightward directions must visibly turn or bend right, up/down directions must use the eyes, eyelids, head angle, neck, and upper body as physically appropriate, and diagonals must interpolate between those pose families. A set where every cell remains essentially front-facing, or where all leftward cells still read as front/right-facing, is failed.
• Adjacent direction cells must have continuous body movement. Compare every neighboring pair in direction order, including 157.5 -> 180, 337.5 -> 000, and any row-strip boundary. Anchored parts must not jump, flip sides, or teleport between adjacent states; if a body part moves laterally, bends, stretches, or rotates, its position should progress gradually across the intervening directions.
• Do not mirror, re-center, or independently regenerate adjacent direction cells in a way that changes the pet's body registration. Keep a stable anchor, usually the feet/base/torso/lower body or the natural grounded part of the character, and let only the intended look mechanics change around it.
• Every look cell must be visually distinguishable from the neutral/resting frame at final pet size. A direction cell that reads as front-facing, idle, or neutral is failed even when it is non-empty, transparent, and in the correct row/column.
• Cardinal directions must be semantically unmistakable at final pet size, not only numerically or geometrically different. 000 must clearly read as looking up, 090 as looking right, 180 as looking down, and 270 as looking left using the pet's natural mechanism. If the pet has no pupils or physical eyeballs, the head, face surface, eyelids, antennae, ears, or body bend must carry the direction clearly enough that a viewer can identify the cardinal without labels.
• Diagonal and intermediate directions should broadly occupy the intended quadrant and advance naturally through the ordered loop. Minor pupil, nose, eyelid, or feature-placement deviations are not failures by themselves. Reject only gross wrong-quadrant poses, visible reversals, or intermediate cells that break the coherent motion family.
• For eyeless object pets, do not default to literal whole-object rotation just because the object is rigid. First identify whether the object has a natural front, display face, playable surface, readable silhouette, or iconic viewing angle. Preserve that primary readable face unless the user explicitly asks for turntable rotation. Express look direction through subtle object-specific body language: small lean, neck/tip aim, hinge, yaw, pitch, bend, vibration, squash, follow-through, or attached-part motion. The direction should read as attention or orientation, not as the object spinning through all clock angles.
• Preserve the pet's original eye design in look-direction cells. Do not paint new round "googly" eyes, replacement eye whites, floating pupils, detached eye dots, or a second eye layer on top of the source eyes. Eye motion must follow the look mechanics decision. If the pet has physical eyeballs, rotate or redraw the whole eyeball surface so the sclera/eye white, iris, pupil, eyelids, rim, and highlights change together as one physical eye; do not slide only the iris or pupil across a fixed eye white. If the pet has flat printed, sticker, or screen eyes, keep the surface fixed and move/redraw only the features that would physically change on that surface. Do not use procedural pupil/iris compositing unless it is clipped to the original eye aperture and visibly remains inside the head silhouette in every direction. If the original eye design cannot be preserved cleanly, regenerate the whole look cell with the original eye construction instead of compositing new eyes over it.
• Eyes may lead the gaze, but pupil-only motion is an exception, not the default. Use it only when the look mechanics decision explains why whole-eye rotation, eyelid reshaping, body, head, or feature movement would be unnatural for that specific design. Large-eye pets, cyclops pets, and round rigid-body pets with physical eyeballs usually should rotate the whole eye globes, not use pupil-only or googly-eye sliding. Screen-face pets and printed-eye pets may be body-locked with feature motion only. Separate head/body pets should usually combine eye movement with head turn, head tilt, ear/fur/upper-body follow-through, and a stable torso. Rigid object mascots may hinge, flex, slide, or shift attached features without rotating the whole sprite. Flexible wire or paperclip-like mascots should usually keep the feet/base anchored while the upper loop or face area bends toward the target and held props remain stable or lag subtly. Blob or organic pets should usually keep a base anchored while the face/head area stretches subtly toward the target. Other pet types should get their own similarly grounded mechanics.
• Human or humanoid pets need persona-preserving look mechanics. Do not use broad non-rigid raster warps that stretch the skull, brows, mouth, hoodie, hands, or held props just to make a direction read. The eyes should usually lead the gaze with visible eye, eyelid, and eyebrow participation, then the head/neck and upper body should follow subtly; a humanoid row where the head moves but the eyes stay locked in one expression is failed unless the mechanics decision gives a specific physical reason. Use small eye rotation, eyelid/eyebrow changes, head/neck turn, and restrained upper-body follow-through while preserving facial proportions and expression. Programmatic repairs must move anatomical parts with rigid or near-rigid part motion, not displacement fields that change facial feature spacing. For pets with held, worn, or attached props, infer each prop's physical constraints before generating look directions: where it is anchored, whether it is rigid or flexible, whether it leads or lags the body, and how it should occlude or be occluded as the character turns. Props near the face may become more side-on, partly hidden by the head, or reveal different contact points; hand-held tools may swing or lag subtly while staying attached; worn props should follow the body; flexible cords or straps should arc continuously. Do not keep the prop and character in the same front-facing relationship across all look directions. Before packaging a humanoid pet, inspect the normal-size neutral and cardinal cells together and reject identity or facial-proportion drift, or a 270 cardinal that does not unmistakably read as left.
• For every pet, use cardinal anchors instead of trusting a two-row sweep to preserve left/right semantics. Generate 000, 090, 180, and 270 together as one strip, then extract and approve them. The final look rows use those four pose families for direction meaning and interpolate the intermediate directions as a coherent arc. Define directions in viewer/screen coordinates, never character-relative coordinates. Do not require exact pupil or nose placement on intermediate poses; use the ordered loop and overall quadrant motion as the primary evidence.
• Keep motion subtle and pet-safe: preserve volume, baseline, silhouette readability, identity, and material believability. The look pose may involve head, eyes, face, upper body, appendages, or body deformation only when those parts would naturally participate.
• Do not add labels, degree text, arrows, clocks, guide marks, shadows, glows, scenery, or detached effects.

Direction order is fixed:

row 9:  000, 022.5, 045, 067.5, 090, 112.5, 135, 157.5
row 10: 180, 202.5, 225, 247.5, 270, 292.5, 315, 337.5

000 means looking up / 12 o'clock. Neutral/front is the pointer deadzone and should fall back to idle unless the target renderer explicitly uses a neutral cell.