Implements the Replicate API backend (FLUX schnell / FLUX dev) per ImaGen
issue #3:
- internal/backend/replicate.go — Backend adapter. Supports model
refs as "owner/name" (uses /v1/models/{owner}/{name}/predictions) and
"owner/name:hash" (uses /v1/predictions with explicit version). Polls
/v1/predictions/{id} every 500ms with model-aware timeout (60s schnell,
120s dev). Resilience: 401 names api_token_env, 429 with exp backoff
up to 3 retries (honours Retry-After), 5xx retries once, image
download retries once on transient failure.
- internal/backend/replicate_pricing.go — hardcoded per-image USD rates
for known FLUX models, snapshotted from replicate.com/pricing with a
refresh TODO.
- internal/backend/replicate_test.go — mocked-HTTP unit tests covering
happy path (model + version-pinned), 401, 429 retry policy, failed
prediction, poll timeout, image-download retry, ctx cancel, BackendOpts
passthrough, default_steps, aspect-ratio reduction, sha256 prompt hash.
- internal/usage/usage.go — Supabase REST sink + read-side query for
mai.imagen_usage. Adapter writes are best-effort: failures warn but
the image still lands.
- cmd/imagen/usage.go — `imagen usage [--since DATE] [--raw]` reads
the table and prints a tab-aligned grouped or raw table with totals.
- cmd/imagen/backends.go — instances of type=replicate now report
"ok" or "not configured (set REPLICATE_API_TOKEN)" depending on env.
- internal/config/config.go — sample adds flux-schnell-replicate +
flux-dev-replicate; default_backend stays flux-schnell-local.
- Supabase migration mai.imagen_usage (id, created_at, backend, model,
seed, prompt_hash, latency_ms, cost_usd_estimate, caller) + indexes
on (created_at DESC) and (caller). The raw prompt is never stored.
Caller identity resolves from MAI_FROM_ID, then the tmux pane's
@mai-name option, mirroring the maimcp identity logic. Prompt hash is
sha256 of the user-facing prompt; raw prompt never reaches the table.
Adds an optional `imagen generate` post-step that opens a sibling
tmux window running tmux-img --hold <path>.
- internal/preview: Mode (auto|on|off), Resolve, and a Spawner that
shells out to tmux new-window. Typed errors for missing tmux,
missing tmux-img, and "preview forced on outside $TMUX".
- cmd/imagen/generate: --preview / --no-preview flags plus
$IMAGEN_PREVIEW. Resolution chain: config -> env -> flag.
auto requires both stdout-is-tty and $TMUX. Failures are
warnings - the image is already on disk.
- internal/config: output.preview field, validated to auto|on|off,
threaded into the sample.
- Tests for ParseMode, Resolve, Spawn argv (incl. shell quoting of
paths with apostrophes), missing-binary errors, and the CLI
resolution table.
- Docs (usage + architecture) updated.
/imagine SKILL.md edit lives in dotfiles - deferred to coordinate
with #4.
internal/backend/comfyui.go implements the Backend interface against
ComfyUI's /prompt + /history + /view HTTP API. Workflow is the canonical
FLUX.1 schnell shape — UNETLoader + DualCLIPLoader (clip_l + t5xxl fp8) +
VAELoader + ModelSamplingFlux + KSampler — assembled as a Go map per
request so Width / Height / Seed / Steps / sampler / scheduler all flow
into the right node inputs.
Resilience: one retry on /prompt 5xx and transient network errors, no
retry on 4xx. Connection-refused / timeouts surface a 'boot-whitetower
mrock' hint. node_errors mentioning a missing unet point users at
docs/setup-comfyui-mrock.md (matches both the 4xx and 200-with-errors
shapes ComfyUI uses across versions).
Result.Metadata carries model, seed_used, latency_ms, steps, sampler,
scheduler, width, height, prompt_id, client_id, plus best-effort
vram_used_mib pulled from /system_stats post-gen.
Tests use httptest with poll interval squashed to 1ms — no real mRock
dependency. Coverage: happy path, defaults, retry-once on 5xx, give-up
after two 5xx, no-retry on 4xx, missing-model hint (both 4xx and
200+node_errors paths), history-error surfaced, /view 4xx, unreachable
host, ctx cancel during poll, workflow-shape assertion, registration.
Config sample: flux-schnell-local is now default_backend; the user-facing
block names the unet file by basename (the mapping into models/unet/ is
the server's convention, captured in docs/setup-comfyui-mrock.md from
phase 1).
Smoke verified end-to-end: imagen generate ... --backend
flux-schnell-local --size 1024x1024 --output /tmp/cat-via-cli.png on
mRock returned a 1024x1024 PNG of a cat in a fishbowl in 10.3s with a
sidecar carrying seed + latency_ms + the rest of the metadata.
Native systemd install (matches Ollama pattern on Arch — Docker on mRock
has no nvidia runtime; native venv via uv is the lighter path). The
Black-Forest-Labs FLUX.1-schnell HF repo is gated, so the download script
points at ungated mirrors (Comfy-Org/flux1-schnell + sirorable/flux-ae-vae)
that ship the same Apache-2.0 weights.
First image — cat in a fishbowl, 1024x1024, 4 steps — generated end-to-end
in 9.79s via curl + workflow JSON; stored at
/home/m/dev/ImaGen/poc/first-image.png on mRiver (not committed; transient
PoC artefact). Go adapter is phase 2.
