Template

Files

yuhaiming 4a4aabf19e docs: plan Omen cat Codex pet

2026-07-16 14:23:35 +08:00

30 KiB

Raw Blame History

Omen Cat Codex Pet Implementation Plan

For agentic workers: REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (- [ ]) syntax for tracking.

Goal: Build, validate, and locally install a complete Codex v2 animated pet named 幽喵, grounded in the approved Omen-cat reference and dark character design.

Architecture: The installed hatch-pet scripts own deterministic layout, extraction, atlas assembly, chroma cleanup, and validation. The installed imagegen skill is the only visual generation layer; each visual job runs in an isolated worker and returns one selected source path. The parent copies approved outputs into a durable workspace run, updates the structured job manifest, runs every deterministic gate, and installs only the validated v2 WebP and manifest.

Tech Stack: Codex hatch-pet skill, Codex imagegen skill, bundled Python 3 with Pillow, PowerShell, JSON manifests, PNG/WebP/GIF assets.

Global Constraints

Design spec: D:/code/water/docs/superpowers/specs/2026-07-16-omen-cat-pet-design.md.
Canonical source: C:/Users/admin/AppData/Local/Temp/codex-clipboard-ecce7d4b-e0a8-495f-98ba-9bf2f2d030b4.png.
Durable work root: D:/code/water/tmp/hatch-pet/omen-cat.
Run directory: D:/code/water/tmp/hatch-pet/omen-cat/run.
Installed skill directory: C:/Users/admin/.codex/skills/hatch-pet.
Bundled Python: C:/Users/admin/.cache/codex-runtimes/codex-primary-runtime/dependencies/python/python.exe.
Pet id: youmiao; display name: 幽喵; sprite contract: spriteVersionNumber: 2.
Style: dark flat illustration with restrained cel shading; preserve purple cat-eared hood, black void face, three cyan face marks, compact paws, long attached tail, broad armor, and bandage wraps.
Never add text, logos, weapons, scenery, floor shadows, smoke, particles, detached effects, thin fragments, blur, or bloom.
Use imagegen for every normal visual job. Do not replace missing generated rows with local drawing, tiling, or procedural synthesis.
The final atlas must be exactly 1536x2288, based on 192x208 cells, and pass all v2 deterministic and visual gates.
The dirty repository worktree is user-owned. Stage or commit only the design and plan documents after Git write approval succeeds.

Task 1: Preserve The Reference And Prepare The Run

Files:

Create: D:/code/water/tmp/hatch-pet/omen-cat/reference/omen-cat-source.png
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/pet_request.json
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/imagegen-jobs.json
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/prompts/**
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/references/layout-guides/**

Interfaces:

Consumes: the approved design spec and the clipboard PNG.
Produces: a durable source image, prepared pet request, prompt set, layout guides, and dependency-aware visual job manifest.
Step 1: Verify the temporary source still exists

Run:

Test-Path -LiteralPath 'C:/Users/admin/AppData/Local/Temp/codex-clipboard-ecce7d4b-e0a8-495f-98ba-9bf2f2d030b4.png'

Expected: True. If it is False, stop and ask the user to attach the image again; do not substitute the web-search image.

Step 2: Copy the source into the durable work root

Run:

New-Item -ItemType Directory -Force -Path 'D:/code/water/tmp/hatch-pet/omen-cat/reference'
Copy-Item -LiteralPath 'C:/Users/admin/AppData/Local/Temp/codex-clipboard-ecce7d4b-e0a8-495f-98ba-9bf2f2d030b4.png' -Destination 'D:/code/water/tmp/hatch-pet/omen-cat/reference/omen-cat-source.png'
Get-FileHash -Algorithm SHA256 -LiteralPath 'D:/code/water/tmp/hatch-pet/omen-cat/reference/omen-cat-source.png'

Expected: the destination exists and Get-FileHash prints one SHA-256 value.

Step 3: Prepare the hatch-pet run

Run:

$PYTHON = 'C:/Users/admin/.cache/codex-runtimes/codex-primary-runtime/dependencies/python/python.exe'
$SKILL_DIR = 'C:/Users/admin/.codex/skills/hatch-pet'
& $PYTHON "$SKILL_DIR/scripts/prepare_pet_run.py" --pet-name '幽喵' --pet-id 'youmiao' --display-name '幽喵' --description '一只披着暗紫幽影兜帽、以冷青面纹观察任务的沉静猫形伙伴。' --reference 'D:/code/water/tmp/hatch-pet/omen-cat/reference/omen-cat-source.png' --output-dir 'D:/code/water/tmp/hatch-pet/omen-cat/run' --pet-notes 'Dark Omen-inspired cartoon cat; oversized angular purple cat-eared hood; black void face with three cyan energy marks; compact feline body; broad charcoal armor and bandage wraps; long attached tail; no weapons, logos, text, scenery, shadows, smoke, particles, detached effects, blur, or bloom.' --style-preset 'flat-vector' --style-notes 'Dark flat illustration with restrained cel shading, broad readable armor shapes, crisp opaque cyan marks, compact sprite-safe silhouette.' --force

Expected: exit code 0; the run contains pet_request.json, imagegen-jobs.json, prompts, and layout guides.

Step 4: Verify prepared identity and v2 job graph

Run:

$request = Get-Content -Raw -LiteralPath 'D:/code/water/tmp/hatch-pet/omen-cat/run/pet_request.json' | ConvertFrom-Json
$jobs = Get-Content -Raw -LiteralPath 'D:/code/water/tmp/hatch-pet/omen-cat/run/imagegen-jobs.json' | ConvertFrom-Json
$request | Select-Object pet_id, display_name, description, chroma_key
$jobs.jobs | Select-Object id, kind, status, depends_on, output_path

Expected: pet_id is youmiao; display name is 幽喵; jobs include base, all nine standard states, look-cardinals, look-row-9, and look-row-10.

Step 5: Checkpoint the design documents when Git approval is available

Run only after Git write approval succeeds:

git add -- docs/superpowers/specs/2026-07-16-omen-cat-pet-design.md docs/superpowers/plans/2026-07-16-omen-cat-pet.md
git commit -m "docs: plan Omen cat Codex pet"

Expected: one commit containing only the two documentation files. If the approval service still returns 404, record the environment blocker and continue the asset run without staging unrelated files.

Task 2: Generate And Approve The Canonical Base

Files:

Consume: D:/code/water/tmp/hatch-pet/omen-cat/run/prompts/base-pet.md
Consume: D:/code/water/tmp/hatch-pet/omen-cat/run/imagegen-jobs.json
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/decoded/base.png
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/references/canonical-base.png

Interfaces:

Consumes: the durable Omen-cat source with the role recorded in the base job.
Produces: one approved full-body chroma-key base that every row worker uses as the identity source of truth.
Step 1: Dispatch one isolated base worker

Give the worker exactly this task, with every input_images entry from the base manifest job attached and role-labeled:

Generate the hatch-pet base image.

Run dir: D:/code/water/tmp/hatch-pet/omen-cat/run
Job id: base
Prompt file: D:/code/water/tmp/hatch-pet/omen-cat/run/prompts/base-pet.md

Use imagegen only. Preserve the canonical reference's purple cat-eared hood, black void face, three cyan marks, compact feline paws, and attached tail. Apply the approved darker layered mantle and broad bandage armor. Return one centered full-body pet on the exact flat chroma background with no text, scenery, shadows, smoke, particles, detached effects, blur, or bloom.

Return exactly two lines. The first line starts with `selected_source=` followed by the absolute PNG path produced by imagegen. The second starts with `qa_note=` followed by one sentence confirming the visible identity and background checks.

Expected: exactly two output lines and a visually coherent single pet.

Step 2: Copy the selected file and establish canonical identity

Copy the exact selected_source returned by the worker to both paths below; do not transform it:

D:/code/water/tmp/hatch-pet/omen-cat/run/decoded/base.png
D:/code/water/tmp/hatch-pet/omen-cat/run/references/canonical-base.png

Expected: both files exist and have identical SHA-256 hashes.

Step 3: Validate the base visually before marking it complete

Require all of the following: one centered whole body, both cat ears readable, three cyan face marks in the correct arrangement, complete paws and tail, broad attached armor, flat uniform key background, and no forbidden effects. Reject identity drift or a cropped tail.

Expected: a pass note grounded in visible identity landmarks.

Step 4: Mark only base complete in the JSON manifest

Update the base object in imagegen-jobs.json with status: "complete", the exact source_path, and an ISO-8601 UTC completed_at. Preserve every other job and field.

Expected: idle and running-right become dependency-ready; no other visual job is marked complete.

Task 3: Generate And Incrementally Validate Standard Rows

Files:

Consume: D:/code/water/tmp/hatch-pet/omen-cat/run/prompts/rows/*.md
Consume: D:/code/water/tmp/hatch-pet/omen-cat/run/references/layout-guides/*.png
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/decoded/{idle,running-right,running-left,waving,jumping,failed,waiting,running,review}.png
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/qa/rows/*/review.json

Interfaces:

Consumes: the canonical base and each job's exact manifest-listed inputs.
Produces: nine separately approved row strips with deterministic frame-extraction evidence.
Step 1: Generate idle and running-right in parallel isolated workers

Use one worker per job. Each worker reads its exact prompt and retry prompt, attaches every manifest input image, and returns only selected_source plus qa_note.

Expected: idle has visible micro-variation; running-right faces screen-right and visibly alternates its gait without dust, speed lines, or shadow.

Step 2: Copy and validate each selected row before manifest completion

For each copied row, run:

$PYTHON = 'C:/Users/admin/.cache/codex-runtimes/codex-primary-runtime/dependencies/python/python.exe'
$SKILL_DIR = 'C:/Users/admin/.codex/skills/hatch-pet'
$RUN_DIR = 'D:/code/water/tmp/hatch-pet/omen-cat/run'
$JOB_ID = 'idle'
& $PYTHON "$SKILL_DIR/scripts/extract_strip_frames.py" --decoded-dir "$RUN_DIR/decoded" --output-dir "$RUN_DIR/qa/rows/$JOB_ID/frames" --states $JOB_ID --method auto
& $PYTHON "$SKILL_DIR/scripts/inspect_frames.py" --frames-root "$RUN_DIR/qa/rows/$JOB_ID/frames" --json-out "$RUN_DIR/qa/rows/$JOB_ID/review.json" --states $JOB_ID --require-components

Repeat with $JOB_ID = 'running-right', then later with each remaining standard row id.

Expected: both commands exit 0; review JSON has no errors. Inspect warnings before acceptance.

Step 3: Decide whether running-left can be mirrored

Mirror only when hood, cyan marks, armor, bandage wraps, tail construction, and cel lighting are symmetrical and retain their meaning after flipping. When safe, run:

& 'C:/Users/admin/.cache/codex-runtimes/codex-primary-runtime/dependencies/python/python.exe' 'C:/Users/admin/.codex/skills/hatch-pet/scripts/derive_running_left_from_running_right.py' --run-dir 'D:/code/water/tmp/hatch-pet/omen-cat/run' --confirm-appropriate-mirror --decision-note 'The approved base and running-right row use symmetric face marks, armor, wraps, tail construction, and cel lighting; per-frame mirroring preserves identity and temporal order.'

If any listed feature is asymmetric, dispatch a normal running-left imagegen worker instead.

Expected: running-left faces screen-left, preserves frame timing, and passes its incremental frame inspection.

Step 4: Generate the remaining six rows with up to three workers active

Jobs: waving, jumping, failed, waiting, running, and review. Backfill a worker slot as soon as another job becomes dependency-ready.

Expected state gates:

waving: forepaw gesture only; no wave marks.
jumping: body-height motion only; no floor cue.
failed: lowered ears, dimmed cyan marks, inward tail; no floating symbols.
waiting: expectant, alert pose distinct from idle.
running: focused task activity without literal locomotion.
review: narrowed marks and deliberate head tilt without new props.
Step 5: Copy, inspect, and complete each remaining row individually

After each source is copied, run the same extract_strip_frames.py and inspect_frames.py pair with that row's exact id. Mark a job complete only after deterministic inspection and visual state review pass.

Expected: all nine standard jobs are complete; every per-row review JSON has no errors.

Task 4: Assemble And Review The Standard Atlas

Files:

Create: D:/code/water/tmp/hatch-pet/omen-cat/run/frames/**
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/qa/review.json
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/final/spritesheet.webp
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/qa/contact-sheet.png
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/qa/previews/*.gif

Interfaces:

Consumes: nine approved standard strips.
Produces: the reviewed intermediate 8x9 atlas and motion evidence required before look-direction generation.
Step 1: Extract all standard frames and inspect them together

Run:

$PYTHON = 'C:/Users/admin/.cache/codex-runtimes/codex-primary-runtime/dependencies/python/python.exe'
$SKILL_DIR = 'C:/Users/admin/.codex/skills/hatch-pet'
$RUN_DIR = 'D:/code/water/tmp/hatch-pet/omen-cat/run'
New-Item -ItemType Directory -Force -Path "$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

Expected: exit code 0 and no errors in qa/review.json.

Step 2: Compose the intermediate atlas, contact sheet, and previews

Run:

& $PYTHON "$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"

Expected: an intermediate 1536x1872 atlas, contact sheet, and one preview per standard state.

Step 3: Run standard visual review

Inspect the contact sheet and all GIFs for identity drift, cropped parts, reversed or static gait, wrong facing, size popping, and incorrect state semantics. Key-color fringe is not a failure at this stage.

Expected: all nine rows pass. If extraction alone causes popping and source strips are stable, rerun extraction with --method stable-slots, then rerun inspection with --allow-stable-slots, atlas composition, contact sheet, and previews.

Task 5: Define Look Mechanics And Approve Cardinal Anchors

Files:

Create: D:/code/water/tmp/hatch-pet/omen-cat/run/qa/look-mechanics.md
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/decoded/look-cardinals.png
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/decoded/look-anchors/{000,090,180,270}.png
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/decoded/look-anchors-approved.png
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/qa/cardinal-anchors.json

Interfaces:

Consumes: approved standard contact sheet, canonical base, and design look mechanics.
Produces: four semantically approved direction anchors in fixed order: up, screen-right, down, screen-left.
Step 1: Write the pet-specific look mechanics note

Use this exact decision:

# Look Mechanics

The paws and lower torso remain anchored. The three cyan face marks lead the gaze inside the black face plane; the hood opening and head turn follow. Cat ears tilt with restrained follow-through. The layered mantle and attached tail lag by one small visual step without changing attachment points. Screen-left and screen-right require the hood opening and face plane to turn visibly, not only the cyan marks. Up and down combine face-mark placement, hood pitch, ear angle, and upper-body compression. The complete sprite never rotates, skews, or tilts to fake direction.

Expected: qa/look-mechanics.md exists before cardinal generation.

Step 2: Generate the four-cardinal strip in one isolated worker

The worker reads the cardinal prompt and qa/look-mechanics.md, attaches all manifest inputs, and returns only selected_source plus a landmark-based qa_note covering 000, 090, 180, and 270.

Expected: four separated poses in that order; 090 unmistakably points to the viewer's right and 270 to the viewer's left.

Step 3: Extract and compose approved anchors

Run:

$request = Get-Content -Raw -LiteralPath "$RUN_DIR/pet_request.json" | ConvertFrom-Json
$CHROMA_KEY = $request.chroma_key.hex
& $PYTHON "$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"

Expected: extraction report passes and all four anchor files are complete and unclipped.

Step 4: Approve cardinal semantics before row 9

Inspect all four at normal pet size. For left/right, record the cyan marks and hood opening relative to head center; for up/down, record vertical mark placement, hood pitch, and ear angle.

Expected: all four pass. If one fails, regenerate only that cardinal anchor with its prepared repair prompt, recompose the approved strip, and repeat semantic review.

Task 6: Generate, Register, And Approve Look Row 9

Files:

Create: D:/code/water/tmp/hatch-pet/omen-cat/run/decoded/look-row-9.png
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/qa/look-row-9-registered.png
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/qa/look-row-9-registration.json

Interfaces:

Consumes: approved cardinal strip, canonical base, standard contact sheet, layout guide, and look mechanics.
Produces: eight registered poses for 000 through 157.5 with one shared scale and baseline.
Step 1: Generate row 9 as one coherent eight-pose family

The isolated worker uses the manifest prompt and retry prompt plus every listed input. It must synthesize exactly 000, 022.5, 045, 067.5, 090, 112.5, 135, 157.5 in order.

Expected: eight separated pose groups, no overlap or outer clipping, stable identity, and evenly progressing head, marks, ears, mantle, and tail.

Step 2: Register row 9 using the final assembly transform

Run:

& $PYTHON "$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"

Expected: deterministic pose recovery, registration, and post-registration edge checks pass.

Step 3: Review row-9 semantics and adjacent continuity

Inspect all eight registered cells at normal pet size. Require correct quadrants, stable paws and lower torso, continuous hood and face-mark motion, and attached mantle and tail.

Expected: no hard semantic or continuity failure. Regenerate the complete row for a wrong quadrant, reversal, conspicuous snap, identity drift, or broken attachment.

Step 4: Mark look-row-9 complete

Update only the row-9 manifest object with its exact source path and completion timestamp.

Expected: look-row-10 becomes ready only after this completion.

Task 7: Generate Row 10 And Assemble The V2 Atlas

Files:

Create: D:/code/water/tmp/hatch-pet/omen-cat/run/decoded/look-row-10.png
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/final/spritesheet-extended.png
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/final/spritesheet-extended.webp
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/final/spritesheet-extended.json
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/qa/chroma-despill-extended.json
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/final/validation-extended.json

Interfaces:

Consumes: approved cardinals, completed row 9, row-9 registration, and standard atlas.
Produces: the cleaned, validated 1536x2288 v2 atlas.
Step 1: Generate row 10 as one coherent eight-pose family

The isolated worker must attach the approved cardinal strip and completed row 9 and generate exactly 180, 202.5, 225, 247.5, 270, 292.5, 315, 337.5 in order.

Expected: 180 continues one even step after 157.5, and 337.5 lands one even step before the row-9 000 pose.

Step 2: Assemble the complete v2 atlas

Run:

& $PYTHON "$SKILL_DIR/scripts/assemble_extended_atlas.py" --base-atlas "$RUN_DIR/final/spritesheet.webp" --registered-row-9 "$RUN_DIR/qa/look-row-9-registered.png" --row-9-registration "$RUN_DIR/qa/look-row-9-registration.json" --look-row-10 "$RUN_DIR/decoded/look-row-10.png" --neutral-cell "$RUN_DIR/frames/idle/00.png" --chroma-key $CHROMA_KEY --chroma-threshold 96 --output "$RUN_DIR/final/spritesheet-extended.png" --webp-output "$RUN_DIR/final/spritesheet-extended.webp" --manifest-output "$RUN_DIR/final/spritesheet-extended.json"

Expected: eight row-10 pose groups are recovered under the persisted row-9 scale and all normalized-cell edge checks pass.

Step 3: Run the single authoritative chroma cleanup pass

Run exactly once:

& $PYTHON "$SKILL_DIR/scripts/despill_chroma_edges.py" "$RUN_DIR/final/spritesheet-extended.png" --output "$RUN_DIR/final/spritesheet-extended.png" --webp-output "$RUN_DIR/final/spritesheet-extended.webp" --chroma-key $CHROMA_KEY --json-out "$RUN_DIR/qa/chroma-despill-extended.json"

Expected: qa/chroma-despill-extended.json has ok: true. Do not run a second cleanup pass.

Step 4: Validate the v2 atlas contract

Run:

& $PYTHON "$SKILL_DIR/scripts/validate_atlas.py" "$RUN_DIR/final/spritesheet-extended.webp" --json-out "$RUN_DIR/final/validation-extended.json" --chroma-key $CHROMA_KEY --require-v2

Expected: exit code 0, dimensions 1536x2288, used cells non-empty, unused cells transparent, and no opaque chroma-key errors.

Task 8: Run Direction QA And Final Visual Review

Files:

Create: D:/code/water/tmp/hatch-pet/omen-cat/run/qa/contact-sheet-extended.png
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/qa/look-directions.png
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/qa/direction-blind-pairs.png
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/qa/direction-blind-answer-key.json
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/qa/direction-blind-verdicts-{1,2,3}.json
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/qa/direction-blind-verdicts.json
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/qa/direction-blind-validation.json
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/qa/look-continuity.json
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/qa/direction-semantics.json

Interfaces:

Consumes: the finalized v2 atlas plus standard previews and validation reports.
Produces: independent blind evidence, labeled per-direction semantics, continuity evidence, and final visual acceptance.
Step 1: Create final review media

Run:

& $PYTHON "$SKILL_DIR/scripts/make_contact_sheet.py" "$RUN_DIR/final/spritesheet-extended.webp" --output "$RUN_DIR/qa/contact-sheet-extended.png"
& $PYTHON "$SKILL_DIR/scripts/make_direction_qa_sheet.py" "$RUN_DIR/final/spritesheet-extended.webp" --output "$RUN_DIR/qa/look-directions.png"
& $PYTHON "$SKILL_DIR/scripts/make_direction_blind_qa_sheet.py" "$RUN_DIR/final/spritesheet-extended.webp" --output "$RUN_DIR/qa/direction-blind-pairs.png" --answer-key "$RUN_DIR/qa/direction-blind-answer-key.json"
& $PYTHON "$SKILL_DIR/scripts/measure_direction_continuity.py" "$RUN_DIR/final/spritesheet-extended.webp" --json-out "$RUN_DIR/qa/look-continuity.json"

Expected: all four commands exit 0 and create their declared outputs.

Step 2: Dispatch three fresh isolated blind reviewers

Each reviewer receives only qa/direction-blind-pairs.png, never the atlas, prompt, degree order, labeled sheet, hidden key, or another verdict. Each returns exactly one JSON object classifying every shown A/B pair on the named horizontal or vertical axis.

Expected: three separate valid verdict files with all pairs present.

Step 3: Combine and validate blind verdicts

Run:

& $PYTHON "$SKILL_DIR/scripts/combine_direction_blind_verdicts.py" --verdicts "$RUN_DIR/qa/direction-blind-verdicts-1.json" --verdicts "$RUN_DIR/qa/direction-blind-verdicts-2.json" --verdicts "$RUN_DIR/qa/direction-blind-verdicts-3.json" --json-out "$RUN_DIR/qa/direction-blind-verdicts.json"
& $PYTHON "$SKILL_DIR/scripts/validate_direction_blind_verdicts.py" --answer-key "$RUN_DIR/qa/direction-blind-answer-key.json" --verdicts "$RUN_DIR/qa/direction-blind-verdicts.json" --json-out "$RUN_DIR/qa/direction-blind-validation.json"

Expected: direction-blind-validation.json has ok: true; both cardinal pairs pass. Intermediate warnings proceed to labeled review.

Step 4: Run one independent final visual QA worker

The worker inspects the standard and extended contact sheets, all standard GIFs, focused direction sheet, blind validation, continuity JSON, frame review JSON, and v2 validation. It must report all 16 direction verdicts and concrete horizontal/vertical evidence.

Expected: visual_qa=pass, no repair rows, and no wrong or ambiguous cardinal.

Step 5: Persist all 16 labeled semantic verdicts

Write qa/direction-semantics.json as structured JSON with one entry for each exact label: 000, 022.5, 045, 067.5, 090, 112.5, 135, 157.5, 180, 202.5, 225, 247.5, 270, 292.5, 315, and 337.5. Every entry contains verdict, expected, observed, and reason from the independent worker.

Expected: no fail verdict. A warning is allowed only for an intermediate direction whose labeled loop remains coherent.

Step 6: Resolve failures at row scope

For a major failure, regenerate the smallest complete failing row, then repeat that row's deterministic checks and all downstream assembly and QA. Never patch an individual final look cell. For a minor intermediate blind disagreement, record an accepted evidence-based resolution in qa/blind-review-resolution.json only when labeled review confirms the correct quadrant and continuity.

Expected: all hard gates pass before packaging.

Task 9: Install The Pet And Retain QA Evidence

Files:

Create outside workspace: C:/Users/admin/.codex/pets/youmiao/pet.json
Create outside workspace: C:/Users/admin/.codex/pets/youmiao/spritesheet.webp
Create: D:/code/water/tmp/hatch-pet/omen-cat/run/qa/run-summary.json

Interfaces:

Consumes: the fully validated extended WebP and all required QA reports.
Produces: one locally installed Codex v2 pet and a durable run summary.
Step 1: Request the required out-of-workspace write approval

Request approval specifically for creating C:/Users/admin/.codex/pets/youmiao and writing only pet.json plus spritesheet.webp there.

Expected: explicit approval before any installation write.

Step 2: Install the validated WebP and manifest

After approval, run:

$PET_DIR = 'C:/Users/admin/.codex/pets/youmiao'
New-Item -ItemType Directory -Force -Path $PET_DIR
Copy-Item -LiteralPath 'D:/code/water/tmp/hatch-pet/omen-cat/run/final/spritesheet-extended.webp' -Destination "$PET_DIR/spritesheet.webp"
$petManifest = [ordered]@{ id = 'youmiao'; displayName = '幽喵'; description = '一只披着暗紫幽影兜帽、以冷青面纹观察任务的沉静猫形伙伴。'; spriteVersionNumber = 2; spritesheetPath = 'spritesheet.webp' }
$petManifest | ConvertTo-Json | Set-Content -Encoding utf8 -LiteralPath "$PET_DIR/pet.json"

Expected: both files exist together and pet.json declares spriteVersionNumber: 2.

Step 3: Validate the installed copy

Run:

& 'C:/Users/admin/.cache/codex-runtimes/codex-primary-runtime/dependencies/python/python.exe' 'C:/Users/admin/.codex/skills/hatch-pet/scripts/validate_atlas.py' 'C:/Users/admin/.codex/pets/youmiao/spritesheet.webp' --json-out 'D:/code/water/tmp/hatch-pet/omen-cat/run/final/validation-installed.json' --chroma-key $CHROMA_KEY --require-v2
Get-Content -Raw -LiteralPath 'C:/Users/admin/.codex/pets/youmiao/pet.json' | ConvertFrom-Json | Select-Object id, displayName, spriteVersionNumber, spritesheetPath

Expected: validation passes; printed values are youmiao, 幽喵, 2, and spritesheet.webp.

Step 4: Write the run summary and clean disposable intermediates

Write qa/run-summary.json with ok: true, spriteVersionNumber: 2, and absolute paths for the run, spritesheet, validation, despill report, contact sheet, direction sheet, semantics, blind validation, continuity, review, and installed package.

Keep the source copy, pet_request.json, final WebP, final and installed validation JSON, chroma report, extended contact sheet, focused direction sheet, direction semantics, blind sheets/verdicts/validation, continuity report, standard GIF previews, frame review, and run summary. Remove prompts, layout guides, decoded row strips, extracted frames, PNG intermediates, 8x9 atlas, and job manifest only after installation and summary validation pass.

Expected: the installed pet remains intact and the retained QA set is sufficient to audit every acceptance gate.

Step 5: Final completion report

Report the installed pet directory, retained run directory, final atlas dimensions, manifest version, deterministic validation result, blind cardinal result, final visual QA result, and any accepted intermediate warnings.

Expected: no completion claim without fresh passing evidence from the installed copy.

30 KiB Raw Blame History

Omen Cat Codex Pet Implementation Plan

Global Constraints

Task 1: Preserve The Reference And Prepare The Run

Task 2: Generate And Approve The Canonical Base

Task 3: Generate And Incrementally Validate Standard Rows

Task 4: Assemble And Review The Standard Atlas

Task 5: Define Look Mechanics And Approve Cardinal Anchors

Task 6: Generate, Register, And Approve Look Row 9

Task 7: Generate Row 10 And Assemble The V2 Atlas

Task 8: Run Direction QA And Final Visual Review

Task 9: Install The Pet And Retain QA Evidence

30 KiB

Raw Blame History