Identify the requested output, input material, destination and existing run. For a review or plan-only request, return findings without creating files or calling generation tools. Reuse the user's approved choices.
Preserve explicit and saved image-provider/model choices, including Gemini or Vertex configuration. The agent host does not select the image provider. If no choice exists, use the configured backend; ask only if an actual choice is missing. Historical examples do not establish a current model ranking.
Locate the user-selected Python backend/environment. Check the relevant command's --help before using version-dependent flags. Do not scan another agent's private directories or read credentials into the conversation.
Separate native PowerPoint editing from raster image generation. A screenshot or embedded SVG is not a native editable text box, table or chart.

Choose the command

Intent	Command	Required material
Method, architecture or workflow figure	`generate`	Source text/PDF and communicative intent
Statistical plot	`plot`	Actual CSV/JSON data and plot intent
One image slide	`slide`	Slide prompt
Several image slides	`slide-batch`	Prompt directory
Compare an image with a reference	`evaluate`	Both images, source context and caption
Dataset information/download/removal	`data`	Requested operation; removal needs its own scope
Provider setup	`setup`	A configuration request; use the host's credential workflow

See command examples for invocation patterns. Use the selected checkout's help as the CLI authority. Code-only plotting requests can use the host's plotting tools without an image-model call.

Image-provider flags use factory names such as google_imagen, openai_imagen, bedrock_imagen and openrouter_imagen. VLM flags use names such as gemini, openai and anthropic. Verify names against the installed version.

Generate or revise

Keep supplied labels, arrows, quantities and scientific meaning traceable to the input. Mark invented example data as illustrative. Do not repair missing research facts by making them up. For mixed-language figures, carry an explicit list of required labels into the prompt and check them in the returned image.
Write inline input into a task-owned text file when execution is authorized. Pass file paths as quoted arguments; do not interpolate user prose into shell commands. Keep the original input and references intact.
Use the authorized provider and iteration limits. Do not add paid optimization, a provider comparison, a large dataset download or an unbounded retry loop to a request that did not authorize it. Reuse existing authorization within scope.
For edits, identify the exact prior run/image and reference files first. Use that command's supported continuation/edit mechanism. If the backend lacks reference editing, explain the limitation before substituting regeneration.
For a single changed slide, regenerate that slide and reassemble the deck; preserve unrelated pages and source prompts. A failed batch does not authorize discarding successful pages or regenerating the entire deck.

Recover without changing the task

Observed result	Next action
Rate limit or temporary service error	Use the backend's bounded retry policy within the approved limit; report exhausted attempts
Authentication/configuration error	Stop that call and report the required setup; do not print keys or try another account
Timeout with uncertain generation outcome	Inspect the run and existing output before retrying, to avoid duplicate work/charges
Requested provider unavailable	Keep the selection; present a concrete alternative only if a new choice is needed
Critic timeout, malformed output or unavailable review	Preserve the image and mark it `UNREVIEWED`; never treat missing feedback as acceptance
Partial batch	List successful/missing pages and resume only the missing or requested pages using supported commands

Do not switch providers merely because a fallback credential exists. A retry or fallback must preserve the user's routing and spending choices. Continuation flags are command-specific: generate --continue-run is not evidence that slide-batch --continue exists.

Inspect and deliver

Open the actual output when the host has an image viewer. Check requested labels, connections, numbers, cropping and legibility. If it cannot be inspected, say so. Report the artifact, run identifier when available, provider/model only as actually known, and review state separately. A successful API call, a Critic score and scientific acceptance are different claims.

When revisions are needed, identify the concrete defect and revise the affected input/run within scope. Do not repeatedly rerun generation without a new diagnosis. Keep failures and unreviewed pages visible in the final report.

paperbanana

PaperBanana

Establish the task