ComfyUI platform integration (Phases 2–5): model store, capability picker, installer wiring, V2 pane#878
Merged
Merged
Conversation
_comfyui_models_dir() now returns model_store_root()/comfyui/models/<subdir> instead of hardcoded /var/lib/hal0/comfyui/models/<subdir>. Aligns with per-install model-store config and deployment expectations. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- Vendor 5 kyuz0 scripts (set_extra_paths, get_qwen_image, get_wan22, get_hunyuan15, get_ltx2) — adapted MODEL_DIR → /mnt/ai-models/comfyui/models - New get_sdxl.sh: SDXL base + SDXL-Lightning 8-step LoRA + sdxl-vae-fp16-fix; supports --precision and --dry-run - New get_esrgan.sh: 4x-UltraSharp + RealESRGAN_x4plus → upscale_models/; supports --dry-run - TDD: tests/comfyui/test_fetch_scripts.py 24/24 pass (bash -n, exec bit, yaml 8 keys + base_path, dry-run subdir assertions) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…lasses Adds src/hal0/comfyui/capabilities.py: CAPABILITIES dict (5 caps × 14 variants) with Capability/ModelVariant dataclasses and default_variant() helper. Tests in tests/comfyui/test_capabilities.py (8 tests, all green). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Task 2.4: fetch_model(variant) -> job_id via subprocess.Popen; get_job/cancel_job with module-level registry. esrgan (precision=None) skips --precision arg. 11 TDD tests (mocked Popen), no real downloads. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add two image-gen entries to the curated catalogue: - sdxl-lightning: ByteDance SDXL + 8-step Lightning LoRA (checkpoints/) - esrgan-4x: xinntao RealESRGAN x4plus upscaler (upscale_models/, bundle_only) TDD: tests/registry/test_curated_comfyui.py (8 tests, all green). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
TDD: tests/install/test_comfyui_scripts_shipped.py (6/6 pass).
- Add installer/comfyui/scripts/{comfy-up,comfy-down,comfy-logs,
comfy-postinstall,start-inference,stop-inference}.sh — verbatim from
the working CT105 copies (digest-pinned kyuz0 image kept intentionally).
- installer/install.sh: new "ComfyUI control scripts" block installs
scripts to /opt/comfyui/ (fixed path; comfy-up.sh self-references it),
creates /mnt/ai-models/comfyui/{models,output,input,user,custom_nodes},
and places extra_model_paths.yaml if absent. Dev-mode skipped.
No script-name mismatch: comfyui.py does not shell out to any
/opt/comfyui/*.sh (API comment confirms — scripts are manual-ops only).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds installer/comfyui/extra_model_paths.yaml with in-container bind path /root/comfy-models and all standard ComfyUI model-type keys. TDD tests verify file existence, YAML validity, base_path, and required keys. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds ComfyUI to EXTENSIONS (kind=app, default_enabled=True). install_extension branch starts the container via /opt/comfyui/comfy-up.sh if a container runtime is present, skipping silently otherwise. ComfyUI is NOT a systemd unit. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds ComfyUI to GET /api/install/services (probed via _container_active() using podman/docker inspect). Repair is a special case before the systemd allowlist: calls /opt/comfyui/comfy-up.sh directly, not systemctl. Adds _container_active() helper. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…stall hook - src/hal0/comfyui/selection.py: auto_selections() and variant_for() - POST /api/comfyui/models/fetch: auto or explicit selections → 202 + job ids - Selections.comfyui_defaults: (cap_id, family) pairs recorded at install, no pull - build_auto_selections() populates comfyui_defaults from CAPABILITIES defaults - tests: 15 new TDD tests (7 selection + 8 fetch route), all green Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add POST /render/cancel, POST /restart, GET /logs, POST /workflows/{name}/launch,
GET /preview to the ComfyUI API router. Switchover gate untouched. Telemetry
util field gated to null when no running job (gpu_busy_percent forced-high
artifact). it/s, eta, step absent by design (websocket-only, skip noted).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Port Design/design_handoff_comfyui_imagegen into comfyui-pane.jsx: render-hero (preview frame + step pips + progress + it/s + eta), ordered queue rows (running ldot.generating + pending), iGPU GTT radial gauge + RAM spark + 2x2 device grid, workflows strip (6 flows), models-on-share inventory, container footer with sctrl controls. CSS scoped under .comfy-page (blue --comfy accent). Empty-queue state in-flow with min-height (not position:absolute overlay — PR #845 guard). Reduced-motion rule freezes pulse + shimmer animations. Mock data via window.__comfyuiV2MockOverride seam for e2e injection. 8/8 imagegen-v2 e2e tests pass. comfyui-arbiter-v3 (4 tests) breaks as expected: those specs test the old switchover/pin UI removed in V2; will be updated in Task 5.2 live wiring. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Poll /api/comfyui/status at 900ms; transform server payload into the V2
pane's RUN/QUEUE/GTT/RAM data shape via transformComfyuiStatus (with safe
placeholders for absent per-job detail). Control handlers: Cancel →
POST /render/cancel, Restart (footer) → POST /restart, Logs →
GET /logs, workflow chip → POST /workflows/{name}/launch; Open ComfyUI
link uses live endpoint field. Preview <img> src → /api/comfyui/preview.
window.__comfyuiV2MockOverride seam preserved (override wins, bypasses
poll). Rewrote comfyui-arbiter-v3.spec.ts: 8 live-wired tests (status
renders into hero/queue/telemetry, cancel/launch/restart fire correct
endpoints, idle + degrade paths). imagegen-v2.spec.ts updated for
button/a duality. 16/16 pass; build clean.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The old fetch.py called scripts as `<script> --precision <p>` but all kyuz0 scripts except get_sdxl.sh take POSITIONAL args and require multiple invocations (dependency chain). Auto-fetch silently no-oped for 4/5 families. - capabilities.py: add `fetch_steps: tuple[tuple[str,...],...]` to ModelVariant; populate all 12 variants with exact argv sequences per script contract (positional for ltx2/wan22/hunyuan15/qwen, flag for sdxl, empty for esrgan). - fetch.py: fetch_model now iterates fetch_steps sequentially with `bash <script> *step_args`; stops on first nonzero exit; drop old `--precision` construction. Synchronous (steps run inline). - tests/comfyui/test_fetch.py: replace wrong-contract mock tests with recording _PopenRecorder asserting correct call count, argv order, and early-stop on failure. - tests/comfyui/test_fetch_dryrun.py: real subprocess smoke for flag-form scripts (esrgan --dry-run, sdxl --precision fp16 --dry-run) verifying argv is real-script-valid without network. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Follow-up to 5d3fd15. The synchronous step loop blocked the caller for the whole multi-hour download, breaking the async contract of POST /api/comfyui/models/fetch (returns 202, jobs run in background) and serializing {auto:true} multi-family fetches. - fetch.py: factor step loop into private worker `_run_sequence(rec, ...)`; fetch_model creates rec(status=running), starts it in a daemon Thread stored as rec["_thread"], and returns job_id IMMEDIATELY. get_job reflects live status; cancel_job terminates rec["_proc"] (current step). Stop-on- first-nonzero + cancelled handling stay in the worker. - test_fetch.py: add _wait_done() poll helper; assert terminal status only after worker completes; read _PopenRecorder after completion. New test_non_blocking_returns_while_step_running proves fetch_model returns with status=running while a slow step is still in wait(). Cancel test now cancels a genuinely in-flight (blocked) step. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Promotes ComfyUI toward a first-class hal0 platform component. Built subagent-driven from
docs/superpowers/plans/2026-06-16-comfyui-platform-integration.md(grilled + planned this session).fetch.pyuses--precisionflag but vendored kyuz0 scripts take positional args. The "pick a model per capability" path returns 202 then silently no-ops for qwen/ltx/wan/hunyuan. Do not assume model auto-pull works./statustelemetry incomplete (util/temp/clock/it_s/eta) — UI device grid shows 0.comfy-up.shstandalonedocker runmay clash with arbiter podman slot on :8188 — needs CT105 verify.ghcr.io/hal0ai/comfyuiimage) DEFERRED — keeps consuming kyuz0's digest-pinned image this week (per decision).What works (tested: 130 pytest + 16 Playwright e2e green, ruff clean)
/mnt/ai-models/comfyui/models; capability→model matrix (benchmark defaults: Qwen-Image 4-step, LTX-2 video, +SDXL-Lightning, +ESRGAN upscale); vendoredget_*.sh+ 2 new; async fetch wrapper (see ComfyUI [C1]: deferred model fetch broken for 4/5 families (positional vs --precision) #872); curated SDXL/ESRGAN.extra_model_paths.yaml;comfyuiExtensions-registry entry; installer services step + repair; capability picker records defaults (deferred pull)./render/cancel,/restart,/logs,/workflows/{name}/launch,/previewadded to existing/status+/switchover(gated)+/pin.comfyui-pane.jsx) replacing V1 — render hero + queue + GTT/RAM gauges + workflows strip + footer; live-wired to the API; empty-state lockup (PR fix(memory): empty-bank graph placeholder no longer locks up the dashboard #845) avoided; reduced-motion respected.Decisions (grill 2026-06-16)
Keep kyuz0 digest-pinned image (not build) this week · ComfyUI stays slot/arbiter-managed (not systemd unit) · "official" = deterministic provisioning · switchover implicit + gated · deferred pulls.
Security review: clean (path-traversal router-blocked, fetch input validated, no injection, no secrets).
🤖 Generated with Claude Code
Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com