papr-pipeline

name	papr-pipeline
description	Runs an autonomous multi-round paper improvement loop on a LaTeX project, chaining /papr-panel -> /papr-experiment -> /papr-write -> /papr-review per round and stopping at score 9+/10. Use when the user says "improve my paper", "review pipeline", "run papr", or asks for iterative paper improvement. Heavy side effects: long compute, latex compiles, external Codex calls.
argument-hint	[rounds] [paper_dir]
disable-model-invocation	true
allowed-tools	["Read","Write","Edit","Grep","Glob","Bash","Agent","WebSearch","WebFetch"]

Paper Improvement Pipeline

Fully autonomous. Do NOT ask the user between rounds. Do NOT pause for confirmation.

Each round: /papr-panel -> /papr-experiment -> /papr-write -> /papr-review (Codex plugin, latest GPT)

IMPORTANT: Experiments are NOT optional. Text reframing alone rarely gets a paper accepted at a top venue. If the panel identifies weak experimental evidence, missing ablations, or unsupported claims, experiments MUST be designed and run. Do not skip experiments just because they take time.

On invocation

1. paper_dir = second argument. If missing, ask once and wait.
2. N = first argument (default 3).
3. Setup directories and state. The unique timestamp ensures isolation from previous runs. Each pipeline invocation gets its own directory. Output goes under papr-runs/ (NOT .claude/) because writes to .claude/ are gated by Claude Code permission prompts even with --dangerously-skip-permissions.

mkdir -p papr-runs
BASE=papr-runs/papr-$(date +%m-%d-%H-%M-%S)
mkdir -p $BASE/round-{1..$N}
ln -sfn round-1 $BASE/latest
ln -sfn $(basename $BASE) papr-runs/latest-run

4. Write papr-runs/latest-run/latest/ROUND_STATE.md:

# Pipeline State
## Paper: [paper_dir]
## Rounds: [N]
## Started: [date]

5. Write empty papr-runs/latest-run/latest/DISCUSSION_THREAD.md.
6. Start round loop.

Round loop

Each round MUST complete ALL 4 phases before starting the next.

CRITICAL ISOLATION: Panel and external review MUST run without any prior-round context. Spawn them as isolated Agents so they get a completely fresh context window with zero knowledge of previous rounds. This prevents score inflation.

for round in 1..N:
    ln -sfn round-{round} [BASE]/latest

    # Phase 1: Panel (ISOLATED -- fresh context, no prior scores)
    Agent(prompt: "Run /papr-panel [paper_dir]. This is a fresh first-time
      review. You have no knowledge of previous rounds or scores.")

    # Phase 2: Experiments (ALWAYS run, not optional)
    # Read action list. Even if no items say "requires new experiment",
    # check if any reviewer raised weak evidence, missing ablations,
    # or unsupported claims. Those NEED experiments.
    read ROUND_STATE.md for action list
    invoke /papr-experiment [paper_dir]

    # Phase 3: Write + humanize + compile
    invoke /papr-write [paper_dir]

    # Phase 4: External review (ISOLATED -- fresh context)
    Agent(prompt: "Run /papr-review [paper_dir]. This paper has never been
      reviewed before. Do not read any pipeline state before constructing
      the review prompt.")

    # Verify BOTH scores exist before proceeding
    read ROUND_STATE.md -> confirm panel score + external score (or "skipped")
    if combined score >= 9: stop, print "Paper ready."

# Print final table: Round | Internal | External

Panel and review are invoked via Agent tool (fresh context), NOT via Skill tool (which would inherit the pipeline's contaminated context). Experiment and write CAN use Skill tool since they need prior round context.

Output

papr-runs/papr-[MM-DD-HH-MM-SS]/
├── round-1/   ROUND_STATE.md, DISCUSSION_THREAD.md
├── round-2/   ...
└── latest -> round-N/
papr-runs/latest-run -> papr-[MM-DD-HH-MM-SS]/

Add papr-runs/ to .gitignore so run artifacts stay out of your paper repo.