# Reported issues for PaperBanana

Pod holds 10 of 10 problems reported by people outside the maintainer team. Issues filed by the project's own owners, members and collaborators are excluded entirely — a maintainer's release checklist is not a warning to a prospective user.

Back to [PaperBanana](/mcp/paperbanana).

## Most discussed

### Version inconsistency between pyproject.toml (0.1.2) and server.json (0.1.1)

## Summary

`pyproject.toml:7` declares `version = "0.1.2"`. `server.json:6` declares `"version": "0.1.1"`. The two should match (or both derive from a single source of truth).

## Reproduction

```
$ grep -A1 '^\[project\]' pyproject.toml | grep version
version = "0.1.2"

$ grep '"version"' server.json
  "version": "0.1.1",
```

## Why this matters

Downstream consumers pinning a paperbanana version (e.g. `paperbanana>=0.1.2` in an integration that needs a recent fix) will see one number in pac

[Read the thread](https://github.com/llmsresearch/paperbanana/issues/215) · 2026-05-18 · closed · external user · 1 comment

### [Feature]: MCP: expose diagram run continuation (continue_run + optional user feedback)

### Problem or motivation

## Motivation
PaperBanana already supports a strong human-in-the-loop workflow from the CLI: paperbanana generate --continue / --continue-run <run_id> with optional --feedback, backed by load_resume_state and PaperBananaPipeline.continue_run in paperbanana/core/resume.py and paperbanana/core/pipeline.py. The README also documents this path and the Python API for continue_run.

Users driving the tool from Cursor / Claude Code via MCP only get generate_diagram, which alw

[Read the thread](https://github.com/llmsresearch/paperbanana/issues/189) · 2026-04-23 · closed · outside contributor · 0 comments

### [Feature]: feat(mcp): expose orchestrate and batch workflows as MCP tools

### Problem or motivation

## Problem
The CLI and Studio support publication-scale workflows (orchestrate, batch, plot-batch, sweep + reports), but the MCP server is still oriented around single diagram/plot generation and evaluation. Users who drive PaperBanana from Cursor / Claude Code / other MCP clients cannot kick off a full figure package or a manifest-based batch without leaving the agent context or shelling out manually.

## Goal
Extend the MCP server so IDE agents can invoke the same or

[Read the thread](https://github.com/llmsresearch/paperbanana/issues/179) · 2026-04-21 · closed · outside contributor · 0 comments

### [Feature]: Add native comparative evaluation for statistical plots

### Problem or motivation

## Problem statement
PaperBanana supports comparative VLM-as-judge evaluation for methodology diagrams, but there is no equivalent first-class workflow for statistical plots. This creates a gap where plot generation exists but plot quality cannot be systematically evaluated with the same 4-dimension protocol and interfaces (CLI/Studio/MCP).

## Why this matters
- Prevents regression testing for plot outputs.
- Makes prompt/model tuning for plot generation harder to val

[Read the thread](https://github.com/llmsresearch/paperbanana/issues/150) · 2026-04-13 · closed · outside contributor · 0 comments

### [Feature]: Support PDF files as methodology / evaluation context (with optional page ranges)

### Problem or motivation

## Motivation
Researchers usually work from PDF papers, not hand-maintained .txt excerpts. Today, using PaperBanana means manually copying the methods section (or other relevant pages) into a text file. That is extra friction, easy to get wrong (wrong section, truncated context), and does not match how people actually write and revise papers.

Adding native PDF input lowers the barrier to trying the tool on real work, improves reproducibility of "what text was in conte

[Read the thread](https://github.com/llmsresearch/paperbanana/issues/108) · 2026-03-23 · closed · outside contributor · 0 comments

### [Feature]: Add structured progress streaming across pipeline, CLI, and MCP

### Problem or motivation

## Background / Motivation
Right now, PaperBanana’s progress reporting is:
- Human-oriented only: rich console messages in the CLI and generic logs from structlog.
- Not machine-readable: UIs and higher-level tools cannot easily track fine‑grained status (phases, iterations, timings).
- Not exposed via MCP: MCP clients (e.g., Cursor, Claude Code) only see the final image; they can’t surface live pipeline progress to users.
-
This limits:
- Building custom UIs (e.g., we

[Read the thread](https://github.com/llmsresearch/paperbanana/issues/103) · 2026-03-16 · closed · outside contributor · 2 comments

### [Feature]: Add save_prompts run artifacts to persist formatted agent prompts

### Problem or motivation

## Problem
- PaperBanana has a config knob output.save_prompts: true in configs/config.yaml, but prompts are not actually saved anywhere.
- When debugging quality regressions or provider differences, users can’t easily reproduce what was sent to the VLM/image generator without digging through logs (and logs may not include full prompts).
- 
## Motivation
- Reproducibility: When a diagram/plot looks wrong, the most important debug artifact is the exact prompt that was s

[Read the thread](https://github.com/llmsresearch/paperbanana/issues/76) · 2026-03-06 · closed · outside contributor · 0 comments

### [Bug]: MCP `generate_diagram` returns JPEG data with `image/png` media type

### What happened?

When calling the `generate_diagram` MCP tool, the tool result includes a base64-encoded image declared as `image/png`, but the actual image data is JPEG. This causes the Anthropic API to reject the tool result with a validation error:

```
Error: messages.N.content.0.tool_result.content.1.image.source.base64:
The image was specified using the image/png media type, but the image appears to be a image/jpeg image
```

This error halts the agent after the first diagram, preventin

[Read the thread](https://github.com/llmsresearch/paperbanana/issues/58) · 2026-02-26 · closed · external user · 1 comment

## Most recent

### [Feature]: MCP: add tools to continue a prior diagram or plot run (CLI / Studio parity)

### Problem or motivation

## Problem
Authors often want to extend refinement after a first generation: more visualizer–critic rounds, or new rounds guided by short natural-language feedback, without redoing retrieval / planner / stylist from scratch.

The codebase already supports this for local workflows:

- CLI: paperbanana generate --continue / --continue-run and related flags
- Core: paperbanana.core.resume.load_resume_state and pipeline APIs that accept resume state
- Studio: continue work

[Read the thread](https://github.com/llmsresearch/paperbanana/issues/187) · 2026-04-23 · closed · outside contributor · 0 comments

### [Feature]: Auto-generate figure captions from the final pipeline output

### Problem or motivation

## Problem / Motivation

`paperbanana generate` takes a user-supplied `--caption` (called `communicative_intent` in `GenerationInput`) as **input** to the pipeline. 
This caption seeds the planner and stylist but is never transformed into a publication-ready figure caption — the kind that appears under a figure in a paper: *"Figure 3: Overview of our proposed method. (a) The encoder maps..."*.
After the pipeline finishes, users still need to manually write this caption

[Read the thread](https://github.com/llmsresearch/paperbanana/issues/98) · 2026-03-12 · closed · outside contributor · 0 comments

The remaining reports are on [the project's issue tracker](https://github.com/llmsresearch/paperbanana/issues).