m/ImaGen
1
0
Fork 0
Code Issues 10 Pull Requests Actions Packages Projects Releases Wiki Activity

ImaGen #4: /imagine skill — single entry point for all mai agents #4

New Issue
Open
opened 2026-05-08 12:30:05 +00:00 by mAi · 5 comments
mAi commented 2026-05-08 12:30:05 +00:00
Collaborator
Copy Link

Goal

Ship the /imagine skill — the canonical entry point for any mai agent (otto/head, otto-wa, lex-research workers, mai-writer, m's own /imagine on the CLI) to request an image. Wraps the ImaGen CLI + HTTP API with a thin skill body that codifies usage patterns.

Prerequisites: ImaGen#1 (framework), and at least one real adapter — #2 (ComfyUI) OR #3 (Replicate) — must be live.

Scope

1. Skill file

Create ~/.dotfiles/claude/.claude/skills/imagine/SKILL.md (top-level skill, no mai- prefix per m's preference 2026-05-08 14:24):

---
name: imagine
description: Generate images via the ImaGen framework — local FLUX on mRock or hosted APIs (Replicate, etc.). Use when an agent needs an image for a blog post, presentation, mockup, social post, or any visual artifact. Trigger on m's "imagine X", "Bild von X", "generate image", "make a picture of X", or whenever a worker's output benefits from an image.
allowed-tools: Bash
---

# imagine

Single entry point for image generation across all mai agents.

## Quick path

`imagen generate "<prompt>" --backend <name> --output <path>` is the only command you need.

Default backend is read from `~/.config/imagen.yaml`. Don't override unless the situation demands it (m's machine offline → switch to Replicate; expensive batch → switch to local).

## Backend choice (decision rules)

- **Default**: whatever `imagen config get default_backend` returns. Today that's `flux-schnell-local` on mRock. Don't second-guess.
- **mRock offline** (Tailscale ping fails, ImaGen returns connection error): fall back to `flux-schnell-replicate`. Note in your output that the fallback was used.
- **High-quality / hero image** (blog header, landing-page art, presentation cover): use `flux-dev-replicate` if available. Costs more (~$0.025/img vs $0.003).
- **Batch / bulk** (10+ images for a job): always local. Don't burn API budget on bulk.

## Style presets

`--style photo|illustration|diagram|sketch|blog-header` — see `m/ImaGen/internal/prompt/styles.yaml` for the full list. Append the style preset to your prompt by passing the flag; don't paste the preset text manually.

## Prompt-writing tips for FLUX

- FLUX prefers natural-language descriptions over keyword soup. Write a sentence, not a Stable-Diffusion-style tag list.
- Specify camera/lens/lighting for photo style: "shot on 35mm film, soft window light, shallow depth of field".
- For text-in-image, describe the text explicitly with quotes: "a sign that reads 'Open'". FLUX handles short text reasonably well; for heavy text use a dedicated backend (Ideogram — separate adapter, not shipped yet).

## Output handling

- Default output goes to `~/Pictures/imagen/<date>-<slug>-<seed>.png` plus a `.json` sidecar with the full request + metadata.
- Pass `--output /custom/path.png` to override.
- For PWA replies / WhatsApp images / blog uploads, generate to a temp path, then upload from there. Don't pass the temp path to m's PWA — render the image inline.

## Cost awareness

Every API-backend generation logs a row in `mai.imagen_usage` (cost-estimate, caller, latency). m sees weekly spend via `imagen usage --since 7d`. Don't burn budget on speculative-batch work — confirm the prompt with m first if you'd be running >10 generations.

## Don't

- Don't use Midjourney (no real API, m skipped it).
- Don't generate images of real people without m's explicit go-ahead — privacy / portrait rights are real concerns even for AI-generated likenesses.
- Don't generate copyrighted characters, logos, or trademarks without an explicit creative-license context (m's blog post analysing them is fine, casual reuse is not).
- Don't store generated images in mBrian or memory — they belong in a file directory or a service like the eventual lex.msbls.de page surface.

## Examples

```bash
# Blog header for a post about UPC patent law
imagine generate "a marble courthouse interior bathed in late-afternoon light, a single open book on a wooden table, blog-header style" \
  --style blog-header \
  --output /tmp/upc-courthouse.png

# Quick mockup for a slide
imagine generate "isometric diagram of three servers connected by glowing lines on a dark background, minimal, no text" \
  --style diagram \
  --size 1920x1080

# Fallback when mRock is offline
imagen generate "your prompt" --backend flux-schnell-replicate

When to skip imagine

Not every visual ask needs gen-AI. Skip in favour of:

  • mexdraw — for technical diagrams with controllable geometry (Excalidraw)
  • A Lucide / Heroicons icon — for UI icons (FLUX overshoots)
  • A real photo from m's library — when m has the actual photo

See also

  • m/ImaGen — repo / framework / adapters
  • mai-mexdraw — programmatic Excalidraw diagrams (different shape, different use case)

### 2. Skill registration

Add `imagine` to the mai-* skill catalogue (whatever index file lists available skills) so it surfaces in agent skill listings. Confirm `/imagine` works from a fresh Claude session.

### 3. Worker integration smoke

Pick one existing worker pattern to demonstrate the skill in action:

- `mai-writer` (drafts internal communications) — wire it so when m requests "schreib eine Einladung mit Bild für Y", the writer can call `/imagine` for the header image without leaving its skill scope.
- Or simpler: a one-shot test where otto/head calls `/imagine` from a PWA reply scenario.

Pick whichever is lower-friction.

### 4. Image delivery via PWA

m's main consumption surface is the PWA. Verify the existing `m pwa reply` flow can attach an image (or document the pattern if it can't yet — separate follow-up to add `--attach-image` if missing).

If `m pwa reply --attach-image <path>` already works: document the pattern in the skill body. If not: file a follow-up issue on `m/mAi` for that and skip the demo.

## Acceptance criteria

1. `/imagine` is invokable from any Claude pane after the skill file lands.
2. The skill body covers the decision rules from §1 (backend choice, style, prompt tips, cost, "don't" list, "when to skip").
3. A live demo: m says "imagine a fishbowl with a cat", the agent calls the CLI, an image lands at `~/Pictures/imagen/...png`, m sees it (PWA inline if delivery works, file-link otherwise).
4. Skill file passes whatever lint exists in `~/.dotfiles/claude/` for skill metadata.

## Out of scope

- Image post-processing in the skill (cropping, filters) — caller can do that with ImageMagick.
- Multi-image generation per call — for v0 always one image per `imagine generate`.
- Automatic style learning / m's preference profile — too early.
- Inline preview in tmux panes — `tmux-img` already exists for that, separate concern.

## Refs

- ImaGen framework: ImaGen#1 (depends-on)
- ComfyUI adapter: ImaGen#2 (one of #2/#3 depends-on)
- Replicate adapter: ImaGen#3 (one of #2/#3 depends-on)
- m's preference 2026-05-08 14:24 ("skill kann imagine sein")

## Workflow

Coder role. **Blocked on #1 + at least one of #2/#3.** Assign mAi only after the framework + one real adapter are live and smoke-tested.
## Goal Ship the `/imagine` skill — the canonical entry point for **any mai agent** (otto/head, otto-wa, lex-research workers, mai-writer, m's own /imagine on the CLI) to request an image. Wraps the ImaGen CLI + HTTP API with a thin skill body that codifies usage patterns. Prerequisites: ImaGen#1 (framework), and at least one real adapter — #2 (ComfyUI) OR #3 (Replicate) — must be live. ## Scope ### 1. Skill file Create `~/.dotfiles/claude/.claude/skills/imagine/SKILL.md` (top-level skill, no `mai-` prefix per m's preference 2026-05-08 14:24): ```markdown --- name: imagine description: Generate images via the ImaGen framework — local FLUX on mRock or hosted APIs (Replicate, etc.). Use when an agent needs an image for a blog post, presentation, mockup, social post, or any visual artifact. Trigger on m's "imagine X", "Bild von X", "generate image", "make a picture of X", or whenever a worker's output benefits from an image. allowed-tools: Bash --- # imagine Single entry point for image generation across all mai agents. ## Quick path `imagen generate "<prompt>" --backend <name> --output <path>` is the only command you need. Default backend is read from `~/.config/imagen.yaml`. Don't override unless the situation demands it (m's machine offline → switch to Replicate; expensive batch → switch to local). ## Backend choice (decision rules) - **Default**: whatever `imagen config get default_backend` returns. Today that's `flux-schnell-local` on mRock. Don't second-guess. - **mRock offline** (Tailscale ping fails, ImaGen returns connection error): fall back to `flux-schnell-replicate`. Note in your output that the fallback was used. - **High-quality / hero image** (blog header, landing-page art, presentation cover): use `flux-dev-replicate` if available. Costs more (~$0.025/img vs $0.003). - **Batch / bulk** (10+ images for a job): always local. Don't burn API budget on bulk. ## Style presets `--style photo|illustration|diagram|sketch|blog-header` — see `m/ImaGen/internal/prompt/styles.yaml` for the full list. Append the style preset to your prompt by passing the flag; don't paste the preset text manually. ## Prompt-writing tips for FLUX - FLUX prefers natural-language descriptions over keyword soup. Write a sentence, not a Stable-Diffusion-style tag list. - Specify camera/lens/lighting for photo style: "shot on 35mm film, soft window light, shallow depth of field". - For text-in-image, describe the text explicitly with quotes: "a sign that reads 'Open'". FLUX handles short text reasonably well; for heavy text use a dedicated backend (Ideogram — separate adapter, not shipped yet). ## Output handling - Default output goes to `~/Pictures/imagen/<date>-<slug>-<seed>.png` plus a `.json` sidecar with the full request + metadata. - Pass `--output /custom/path.png` to override. - For PWA replies / WhatsApp images / blog uploads, generate to a temp path, then upload from there. Don't pass the temp path to m's PWA — render the image inline. ## Cost awareness Every API-backend generation logs a row in `mai.imagen_usage` (cost-estimate, caller, latency). m sees weekly spend via `imagen usage --since 7d`. Don't burn budget on speculative-batch work — confirm the prompt with m first if you'd be running >10 generations. ## Don't - Don't use Midjourney (no real API, m skipped it). - Don't generate images of real people without m's explicit go-ahead — privacy / portrait rights are real concerns even for AI-generated likenesses. - Don't generate copyrighted characters, logos, or trademarks without an explicit creative-license context (m's blog post analysing them is fine, casual reuse is not). - Don't store generated images in mBrian or memory — they belong in a file directory or a service like the eventual lex.msbls.de page surface. ## Examples ```bash # Blog header for a post about UPC patent law imagine generate "a marble courthouse interior bathed in late-afternoon light, a single open book on a wooden table, blog-header style" \ --style blog-header \ --output /tmp/upc-courthouse.png # Quick mockup for a slide imagine generate "isometric diagram of three servers connected by glowing lines on a dark background, minimal, no text" \ --style diagram \ --size 1920x1080 # Fallback when mRock is offline imagen generate "your prompt" --backend flux-schnell-replicate ``` ## When to skip imagine Not every visual ask needs gen-AI. Skip in favour of: - `mexdraw` — for technical diagrams with controllable geometry (Excalidraw) - A Lucide / Heroicons icon — for UI icons (FLUX overshoots) - A real photo from m's library — when m has the actual photo ## See also - `m/ImaGen` — repo / framework / adapters - `mai-mexdraw` — programmatic Excalidraw diagrams (different shape, different use case) ``` ### 2. Skill registration Add `imagine` to the mai-* skill catalogue (whatever index file lists available skills) so it surfaces in agent skill listings. Confirm `/imagine` works from a fresh Claude session. ### 3. Worker integration smoke Pick one existing worker pattern to demonstrate the skill in action: - `mai-writer` (drafts internal communications) — wire it so when m requests "schreib eine Einladung mit Bild für Y", the writer can call `/imagine` for the header image without leaving its skill scope. - Or simpler: a one-shot test where otto/head calls `/imagine` from a PWA reply scenario. Pick whichever is lower-friction. ### 4. Image delivery via PWA m's main consumption surface is the PWA. Verify the existing `m pwa reply` flow can attach an image (or document the pattern if it can't yet — separate follow-up to add `--attach-image` if missing). If `m pwa reply --attach-image <path>` already works: document the pattern in the skill body. If not: file a follow-up issue on `m/mAi` for that and skip the demo. ## Acceptance criteria 1. `/imagine` is invokable from any Claude pane after the skill file lands. 2. The skill body covers the decision rules from §1 (backend choice, style, prompt tips, cost, "don't" list, "when to skip"). 3. A live demo: m says "imagine a fishbowl with a cat", the agent calls the CLI, an image lands at `~/Pictures/imagen/...png`, m sees it (PWA inline if delivery works, file-link otherwise). 4. Skill file passes whatever lint exists in `~/.dotfiles/claude/` for skill metadata. ## Out of scope - Image post-processing in the skill (cropping, filters) — caller can do that with ImageMagick. - Multi-image generation per call — for v0 always one image per `imagine generate`. - Automatic style learning / m's preference profile — too early. - Inline preview in tmux panes — `tmux-img` already exists for that, separate concern. ## Refs - ImaGen framework: ImaGen#1 (depends-on) - ComfyUI adapter: ImaGen#2 (one of #2/#3 depends-on) - Replicate adapter: ImaGen#3 (one of #2/#3 depends-on) - m's preference 2026-05-08 14:24 ("skill kann imagine sein") ## Workflow Coder role. **Blocked on #1 + at least one of #2/#3.** Assign mAi only after the framework + one real adapter are live and smoke-tested.
mAi commented 2026-05-08 14:45:58 +00:00
Author
Collaborator
Copy Link

Shift-1 checkpoint (athena/coder)

Status: drafted + committed; live demo parked on #2.

What landed (commit 301fbf3 on mai/athena/issue-4-imagine-skill):

  • mai/.mai/skills/imagine/SKILL.md — top-level skill (no mai- prefix per m's preference 2026-05-08 14:24)
  • Body covers: backend choice rules, style presets (photo/illustration/diagram/sketch/blog-header), FLUX prompt tips, cost awareness, channel-aware delivery (terminal vs PWA vs WhatsApp vs Telegram), don't-list (Midjourney/real people/copyright/storage), when-to-skip (mexdraw/icons/real photos/screenshots), examples, see-also.

Path correction (worth flagging in #4 description): the issue says skills go to ~/.dotfiles/claude/.claude/skills/. That path is gitignored in dotfiles. The tracked location is mai/.mai/skills/<name>/SKILL.md; ~/.claude/skills is a symlink via ~/.mai/skills → ~/.dotfiles/mai/.mai/skills/. Verified against existing top-level skills (edit-article, show-md, seo-audit, lex-research — all under mai/.mai/skills/). Skill placed in the correct location; original draft path was a bug.

Spec-drift fixed in skill body vs issue draft:

  1. Issue examples wrote imagine generate — actual CLI is imagen generate. Fixed.
  2. Issue references imagen config get default_backend — that subcommand does not exist (only init|validate|path). Skill points at imagen config validate (which prints default=...) and ~/.config/imagen.yaml.
  3. Issue references imagen usage --since 7d — also does not exist yet. Skill notes it as a planned follow-up; until then jq -s 'map(.metadata.cost_estimate_usd // 0) | add' ~/Pictures/imagen/*.json.

PWA delivery: filed m/mAi#213 for m pwa reply --attach-image. Confirmed via m pwa reply --help that no image-attach flag exists today. Skill body documents the file-link fallback per channel.

Lint: no skill-metadata lint script exists in ~/.dotfiles/ (grepped). Frontmatter matches existing skills' pattern (name, description, allowed-tools). If stricter validation is wanted, that's a separate task.

Blocked on: ImaGen#2 (FLUX install + ComfyUI adapter on mRock — hermes). Once phase-1 image is in, athena runs /imagine "a fishbowl with a cat" --backend flux-schnell-local end-to-end and confirms acceptance #3.

Next steps:

  • (head) review 301fbf3 and merge into dotfiles/main once you're satisfied with the skill body
  • (head) ping athena when hermes's phase-1 lands → athena runs the live demo
  • (later) close #213 with the inline-attach work to upgrade PWA delivery from file-link to inline preview
## Shift-1 checkpoint (athena/coder) **Status:** drafted + committed; live demo parked on #2. **What landed (commit `301fbf3` on `mai/athena/issue-4-imagine-skill`):** - `mai/.mai/skills/imagine/SKILL.md` — top-level skill (no `mai-` prefix per m's preference 2026-05-08 14:24) - Body covers: backend choice rules, style presets (photo/illustration/diagram/sketch/blog-header), FLUX prompt tips, cost awareness, channel-aware delivery (terminal vs PWA vs WhatsApp vs Telegram), don't-list (Midjourney/real people/copyright/storage), when-to-skip (mexdraw/icons/real photos/screenshots), examples, see-also. **Path correction (worth flagging in #4 description):** the issue says skills go to `~/.dotfiles/claude/.claude/skills/`. That path is **gitignored** in dotfiles. The tracked location is `mai/.mai/skills/<name>/SKILL.md`; `~/.claude/skills` is a symlink via `~/.mai/skills → ~/.dotfiles/mai/.mai/skills/`. Verified against existing top-level skills (`edit-article`, `show-md`, `seo-audit`, `lex-research` — all under `mai/.mai/skills/`). Skill placed in the correct location; original draft path was a bug. **Spec-drift fixed in skill body vs issue draft:** 1. Issue examples wrote `imagine generate` — actual CLI is `imagen generate`. Fixed. 2. Issue references `imagen config get default_backend` — that subcommand does not exist (only `init|validate|path`). Skill points at `imagen config validate` (which prints `default=...`) and `~/.config/imagen.yaml`. 3. Issue references `imagen usage --since 7d` — also does not exist yet. Skill notes it as a planned follow-up; until then `jq -s 'map(.metadata.cost_estimate_usd // 0) | add' ~/Pictures/imagen/*.json`. **PWA delivery:** filed [m/mAi#213](https://mgit.msbls.de/m/mAi/issues/213) for `m pwa reply --attach-image`. Confirmed via `m pwa reply --help` that no image-attach flag exists today. Skill body documents the file-link fallback per channel. **Lint:** no skill-metadata lint script exists in `~/.dotfiles/` (grepped). Frontmatter matches existing skills' pattern (`name`, `description`, `allowed-tools`). If stricter validation is wanted, that's a separate task. **Blocked on:** ImaGen#2 (FLUX install + ComfyUI adapter on mRock — hermes). Once phase-1 image is in, athena runs `/imagine "a fishbowl with a cat" --backend flux-schnell-local` end-to-end and confirms acceptance #3. **Next steps:** - (head) review `301fbf3` and merge into `dotfiles/main` once you're satisfied with the skill body - (head) ping athena when hermes's phase-1 lands → athena runs the live demo - (later) close #213 with the inline-attach work to upgrade PWA delivery from file-link to inline preview
mAi commented 2026-05-08 14:47:10 +00:00
Author
Collaborator
Copy Link

Path correction (head note): the skill file lands at mai/.mai/skills/imagine/SKILL.md in m/dotfiles, NOT claude/.claude/skills/imagine/ as the issue body says. The latter path is gitignored in dotfiles; the canonical tracked layout uses mai/.mai/skills/ and ~/.claude/skills symlinks via ~/.mai/skills. Athena verified against existing top-level skills (edit-article, show-md, seo-audit, lex-research) before committing.

Draft committed at 301fbf3 on mai/athena/issue-4-imagine-skill. Acceptance #1, #2, #4 covered. Acceptance #3 (live demo) parked until ImaGen#2 lands a working FLUX backend on mRock. Filed m/mAi#213 for m pwa reply --attach-image follow-up; skill body documents file-link fallback in the meantime.

Path correction (head note): the skill file lands at `mai/.mai/skills/imagine/SKILL.md` in `m/dotfiles`, NOT `claude/.claude/skills/imagine/` as the issue body says. The latter path is gitignored in dotfiles; the canonical tracked layout uses `mai/.mai/skills/` and `~/.claude/skills` symlinks via `~/.mai/skills`. Athena verified against existing top-level skills (edit-article, show-md, seo-audit, lex-research) before committing. Draft committed at `301fbf3` on `mai/athena/issue-4-imagine-skill`. Acceptance #1, #2, #4 covered. Acceptance #3 (live demo) parked until ImaGen#2 lands a working FLUX backend on mRock. Filed m/mAi#213 for `m pwa reply --attach-image` follow-up; skill body documents file-link fallback in the meantime.
mAi commented 2026-05-08 15:05:41 +00:00
Author
Collaborator
Copy Link

Shift-2 checkpoint (athena/coder) — live demo PASSED

All four acceptance criteria met. Image at /home/m/dev/ImaGen/poc/third-image-via-skill.png (1024x1024 PNG, 1.3 MB).

Skill discoverability (#1): ~/.claude/skills/imagine resolves via the symlink chain into the committed file. Claude's skill listing shows imagine: Generate images via the ImaGen framework... alongside the rest of the top-level skills.

Skill body (#2): Commit 301fbf3 covers backend choice, style presets, FLUX prompt tips, cost awareness, channel-aware delivery, do not list, when-to-skip.

Live demo (#3):

  • Prompt: "a fishbowl with a cat staring out, photo, soft light"
  • Backend: flux-schnell-local (comfyui adapter → http://mrock:8188)
  • Model: flux1-schnell.safetensors
  • Seed: 6585474305228830100
  • Latency: 7839 ms
  • VRAM: 11067 MiB
  • Output: /home/m/Pictures/imagen/2026-05-08-a-fishbowl-with-a-cat-staring-out-photo-6585474305228830100.png + .json sidecar; copied to poc/third-image-via-skill.png for head/m
  • On-prompt: fishbowl on a wooden table, tabby kitten inside, soft golden-hour light. Visually verified.

Skill lint (#4): No skill-metadata lint script exists in ~/.dotfiles/. Frontmatter matches existing top-level skills (name + description + allowed-tools).

Live-run observation for a future tightening (not blocking): The skill body says cost-estimate is in the sidecar metadata "when the adapter supports it". The ComfyUI adapter does not emit cost_estimate_usd (no API billing — local). One-line edit: "for hosted backends only; local ones omit the field". Can land with the next imagen-related ticket; not committing to athena's branch since head wants a single clean commit.

Cleanup note for the merge: I created a temporary symlink at /home/m/.dotfiles/mai/.mai/skills/imagine → /home/m/.dotfiles/.worktrees/athena/mai/.mai/skills/imagine so the demo could find the skill. Before merging the worktree branch into dotfiles main, rm /home/m/.dotfiles/mai/.mai/skills/imagine to clear the path; then git merge produces the real directory in its place.

Branch: mai/athena/issue-4-imagine-skill
Commit: 301fbf3 feat(skills): add /imagine skill — single entry point for image generation

Ready for head's merge into dotfiles main and the commit-link stamp on this issue.

## Shift-2 checkpoint (athena/coder) — live demo PASSED All four acceptance criteria met. Image at `/home/m/dev/ImaGen/poc/third-image-via-skill.png` (1024x1024 PNG, 1.3 MB). **Skill discoverability (#1):** `~/.claude/skills/imagine` resolves via the symlink chain into the committed file. Claude's skill listing shows `imagine: Generate images via the ImaGen framework...` alongside the rest of the top-level skills. **Skill body (#2):** Commit `301fbf3` covers backend choice, style presets, FLUX prompt tips, cost awareness, channel-aware delivery, do not list, when-to-skip. **Live demo (#3):** - Prompt: "a fishbowl with a cat staring out, photo, soft light" - Backend: `flux-schnell-local` (comfyui adapter → `http://mrock:8188`) - Model: `flux1-schnell.safetensors` - Seed: `6585474305228830100` - Latency: **7839 ms** - VRAM: 11067 MiB - Output: `/home/m/Pictures/imagen/2026-05-08-a-fishbowl-with-a-cat-staring-out-photo-6585474305228830100.png` + `.json` sidecar; copied to `poc/third-image-via-skill.png` for head/m - On-prompt: fishbowl on a wooden table, tabby kitten inside, soft golden-hour light. Visually verified. **Skill lint (#4):** No skill-metadata lint script exists in `~/.dotfiles/`. Frontmatter matches existing top-level skills (name + description + allowed-tools). **Live-run observation for a future tightening (not blocking):** The skill body says cost-estimate is in the sidecar metadata "when the adapter supports it". The ComfyUI adapter does not emit `cost_estimate_usd` (no API billing — local). One-line edit: "for hosted backends only; local ones omit the field". Can land with the next imagen-related ticket; not committing to athena's branch since head wants a single clean commit. **Cleanup note for the merge:** I created a temporary symlink at `/home/m/.dotfiles/mai/.mai/skills/imagine → /home/m/.dotfiles/.worktrees/athena/mai/.mai/skills/imagine` so the demo could find the skill. Before merging the worktree branch into dotfiles main, `rm /home/m/.dotfiles/mai/.mai/skills/imagine` to clear the path; then `git merge` produces the real directory in its place. **Branch:** `mai/athena/issue-4-imagine-skill` **Commit:** `301fbf3 feat(skills): add /imagine skill — single entry point for image generation` Ready for head's merge into dotfiles main and the commit-link stamp on this issue.
mAi commented 2026-05-08 15:06:34 +00:00
Author
Collaborator
Copy Link

Live demo PASSED — all four acceptance criteria green

Run

  • Prompt: a fishbowl with a cat staring out, photo, soft light
  • Backend: flux-schnell-local (comfyui adapter, http://mrock:8188)
  • Model: flux1-schnell.safetensors
  • Steps / Latency: 4 / 7839 ms (under the <8s target)
  • Seed: 6585474305228830100
  • VRAM peak: 11067 MiB on the 4070 Ti SUPER
  • Image: /home/m/dev/ImaGen/poc/third-image-via-skill.png (1024×1024, 1.3 MB PNG)
  • Sidecar: .json next to the image, contains seed + latency + backend + prompt

Acceptance

# Check Result
1 /imagine invokable from any Claude pane ok - skill listing shows imagine: Generate images via the ImaGen framework...
2 Skill body covers decision rules (backend choice / style / prompt tips / cost / don't-list / when-to-skip) ok - committed in dotfiles 301fbf3
3 Live demo: agent calls CLI, image lands at ~/Pictures/imagen/...png, m sees it ok - file-link delivery (PWA inline pending m/mAi#213)
4 Skill file passes whatever lint exists in ~/.dotfiles/claude/ ok - frontmatter matches existing top-level skills; no skill-lint script exists in dotfiles

Skill body observations from the live run (one-line follow-up worth filing)

Athena flagged: the skill body says cost-estimate appears in the sidecar "when the adapter supports it" - that's true for hosted backends (Replicate, OpenAI - landing in #3) but the local ComfyUI adapter doesn't emit cost_estimate_usd. One-line clarification - will fold into a follow-up commit on the next imagen-side touch.

Merge status

Branch mai/athena/issue-4-imagine-skill (commit 301fbf3) is ready to merge into dotfiles main. Held by head pending an unrelated unresolved state on m's local dotfiles working tree (UU on mai/.mai/skills/lex-research/SKILL.md from a prior merge; a few uncommitted modifications). Will merge in one shot once that's clear, then stamp the merge commit hash + apply done label here.

## Live demo PASSED — all four acceptance criteria green ### Run - **Prompt**: `a fishbowl with a cat staring out, photo, soft light` - **Backend**: `flux-schnell-local` (comfyui adapter, http://mrock:8188) - **Model**: `flux1-schnell.safetensors` - **Steps / Latency**: 4 / **7839 ms** (under the <8s target) - **Seed**: `6585474305228830100` - **VRAM peak**: 11067 MiB on the 4070 Ti SUPER - **Image**: `/home/m/dev/ImaGen/poc/third-image-via-skill.png` (1024×1024, 1.3 MB PNG) - **Sidecar**: `.json` next to the image, contains seed + latency + backend + prompt ### Acceptance | # | Check | Result | |---|-------|--------| | 1 | `/imagine` invokable from any Claude pane | ok - skill listing shows `imagine: Generate images via the ImaGen framework...` | | 2 | Skill body covers decision rules (backend choice / style / prompt tips / cost / don't-list / when-to-skip) | ok - committed in dotfiles `301fbf3` | | 3 | Live demo: agent calls CLI, image lands at `~/Pictures/imagen/...png`, m sees it | ok - file-link delivery (PWA inline pending m/mAi#213) | | 4 | Skill file passes whatever lint exists in `~/.dotfiles/claude/` | ok - frontmatter matches existing top-level skills; no skill-lint script exists in dotfiles | ### Skill body observations from the live run (one-line follow-up worth filing) Athena flagged: the skill body says cost-estimate appears in the sidecar "when the adapter supports it" - that's true for hosted backends (Replicate, OpenAI - landing in #3) but the local ComfyUI adapter doesn't emit `cost_estimate_usd`. One-line clarification - will fold into a follow-up commit on the next imagen-side touch. ### Merge status Branch `mai/athena/issue-4-imagine-skill` (commit `301fbf3`) is **ready to merge into dotfiles main**. Held by head pending an unrelated unresolved state on m's local dotfiles working tree (UU on `mai/.mai/skills/lex-research/SKILL.md` from a prior merge; a few uncommitted modifications). Will merge in one shot once that's clear, then stamp the merge commit hash + apply `done` label here.
mAi commented 2026-05-08 17:41:15 +00:00
Author
Collaborator
Copy Link

Merged into dotfiles main

Branch mai/athena/issue-4-imagine-skill merged via --no-ff into m/dotfiles main at 7c841fdbe369168a1dc99a5397a6c0dc91598b98. Pushed to origin/main.

  • Implementation commit (athena): 301fbf3 — mai/.mai/skills/imagine/SKILL.md (151 lines)
  • Merge commit (head): 7c841fdbe369168a1dc99a5397a6c0dc91598b98

All four acceptance criteria green

# Check Result
1 /imagine invokable from any Claude pane ok — skill listing includes imagine: Generate images via the ImaGen framework...
2 Skill body covers decision rules (backend / style / prompt tips / cost / don't / when-to-skip) ok
3 Live demo: agent calls CLI, image lands at ~/Pictures/imagen/...png, m sees it ok — /home/m/dev/ImaGen/poc/third-image-via-skill.png, 7.84 s end-to-end
4 Skill metadata frontmatter matches existing top-level skills ok

Filed during the work

  • m/mAi#213 — m pwa reply --attach-image follow-up (skill body documents file-link fallback)
  • Three spec-drift fixes athena caught (imagen generate typo, missing imagen config get, missing imagen usage) — all reflected in the skill body
  • Path correction: skill file lives at mai/.mai/skills/imagine/, NOT claude/.claude/skills/imagine/ (latter is gitignored in dotfiles); saved to memory under imagen group so future workers don't repeat the trap

Open follow-ups (not blocking close)

  • One-line cost-estimate clarification in the skill body (athena flagged that the cost-estimate metadata note doesn't apply to the local ComfyUI backend, only hosted backends like Replicate). Trivial; can fold into the next /imagine SKILL.md touch.
  • ImaGen#5 acceptance #5 (the /imagine SKILL.md preview-section update) was deferred during #5 to avoid colliding with this merge — now unblocked. Will pick up next.
## Merged into dotfiles main Branch `mai/athena/issue-4-imagine-skill` merged via `--no-ff` into m/dotfiles main at `7c841fdbe369168a1dc99a5397a6c0dc91598b98`. Pushed to origin/main. - Implementation commit (athena): `301fbf3` — `mai/.mai/skills/imagine/SKILL.md` (151 lines) - Merge commit (head): `7c841fdbe369168a1dc99a5397a6c0dc91598b98` ### All four acceptance criteria green | # | Check | Result | |---|-------|--------| | 1 | `/imagine` invokable from any Claude pane | ok — skill listing includes `imagine: Generate images via the ImaGen framework...` | | 2 | Skill body covers decision rules (backend / style / prompt tips / cost / don't / when-to-skip) | ok | | 3 | Live demo: agent calls CLI, image lands at `~/Pictures/imagen/...png`, m sees it | ok — `/home/m/dev/ImaGen/poc/third-image-via-skill.png`, 7.84 s end-to-end | | 4 | Skill metadata frontmatter matches existing top-level skills | ok | ### Filed during the work - m/mAi#213 — `m pwa reply --attach-image` follow-up (skill body documents file-link fallback) - Three spec-drift fixes athena caught (`imagen generate` typo, missing `imagen config get`, missing `imagen usage`) — all reflected in the skill body - Path correction: skill file lives at `mai/.mai/skills/imagine/`, NOT `claude/.claude/skills/imagine/` (latter is gitignored in dotfiles); saved to memory under `imagen` group so future workers don't repeat the trap ### Open follow-ups (not blocking close) - One-line cost-estimate clarification in the skill body (athena flagged that the cost-estimate metadata note doesn't apply to the local ComfyUI backend, only hosted backends like Replicate). Trivial; can fold into the next `/imagine` SKILL.md touch. - ImaGen#5 acceptance #5 (the `/imagine` SKILL.md preview-section update) was deferred during #5 to avoid colliding with this merge — now unblocked. Will pick up next.
mAi added the
done
 label 2026-05-08 17:41:16 +00:00
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
done
Work complete, awaiting m close
needs-review
Awaiting m review
needs-smoke
Code merged; real-API smoke still pending
No Label
done
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
mAi m
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: m/ImaGen#4
No description provided.
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?