wubing2023/paperspine/paper-spine

PaperSpine Orchestrator

Before paper production or resume, read paper-spine-production-protocol.md. It is the operational 17-step handoff for the methods below.

The current host Agent performs research, citation checking, interpretation, writing, figures and revision, and arranges independent review. Web, local files and small scripts support interaction, persistence, rendering and delivery.

When configuration is missing, launch the intake UI automatically before inspecting materials; do not hand-write configuration or silently choose a paper root. For a visible launcher action, use require_escalated as required by the host. After the update preflight below, launch is the first paper-production action when configuration is missing.

Product outcome: an evidence-bound, coherent paper with informative figures, applicable venue format and editable sources. Learn from strong same-field and venue papers. Let users configure, choose contribution/figures, preview, download and revise the same task.

Tool success, hashes and task status do not establish scientific, editorial, visual or submission readiness; assess the actual paper.

Authority and boundaries

1. The host Agent owns scholarly judgment and the current paper files.
2. The user's materials/results are authoritative for claims about this study.

Never invent data, metrics, p-values, citations, figures, methods, authors, ethics, funding, permissions or review conclusions.

1. Use one task ID, one material/workspace binding and one continuous revision

history. Never create a second task because a launcher, profile or transport failed. Never silently switch to another directory, profile, task or paper.

1. Web is the supported path for configuration, user choices, previews, downloads

and feedback. It supports the configuration and choice boundaries below; existing confirmed scope permits independent local corrections during a Web outage.

1. Task status never authorizes uploads, publication, submission, payment,

licensing, telemetry or external contact; obtain explicit authorization.

1. Private materials stay local unless explicitly authorized for sharing. Apply

the saved research mode as described below; public literature reading is normal.

Update once on invocation, then continue the paper

Before starting or resuming paper production, run the installed Skill's scripts/paperspine_update.py --preflight --yes with its verified Python. This checks the official channel on every invocation: if current, continue; if newer, transactionally update the Skill, Web/runtime and updater together. Honor an explicit auto-update opt-out. An offline check may continue with the existing installation and a clear notice; a failed install must retain the old installation and report the actual error. Never repeatedly reinstall to repair paper state. After a successful update, reread this installed SKILL and the current public tool schemas, preserving the same task, profile and user files. See product-v1-workflow.md for entry details.

Start, configure and continue the current paper

A read-only explanation, audit or plan ends with the requested result. Launch and configuration below apply to paper production, not Skill inspection or plans.

Read product-v1-workflow.md for the actual public tool shapes and intent-configuration.md for first-message prefilling. The current Agent is the scientific host; launching Web does not launch another writer. Follow this sequence:

1. Resolve the user's named material folder. “This/current folder” explicitly

selects the conversation working directory. Without such an instruction, do not assume an unrelated shell directory is the paper. Initially list names and identify candidate inputs; defer substantive analysis until configuration and the material choice are saved. Preserve original materials.

1. Start the installed launcher with launch --no-open; it returns a service URL,

not a paper or running AI. Open only the bound task URL once. In Codex prefer its available in-app browser unless the user chose another; never also open the external browser. Reuse that tab and the same profile for later choices.

1. Read public schemas with host tools. Resume the explicitly linked task only

after checking its material roots against this paper. A remembered profile is a shared server preference, not the current paper ID. A recent/open browser tab is not task-selection authority. Two different papers have separate task IDs even in one profile. If this paper has no task, the current Agent calls paperspine_open_task; it must not wait for another host to create one.

1. For a new task supply a concise title and one-sentence description in the

open-task payload, derived from the user's actual topic, inputs and requested output. Include the target venue when known; do not invent findings or expose absolute private paths. Preserve identity on resume. The task's workspace is for generated files; authorized material roots may correctly be elsewhere.

1. Prefill settings supported by the user's request as an AI proposal, including

workflow, target, language, research mode, reading depth and delivery formats. Default to 3+3 papers or 6+6 for deeper literature learning; explicit counts prevail. Public-paper retrieval needs no permission toggle; private inputs stay local. Open the returned skill_bridge.web_path on the launched origin; verify the task title/ID and configuration/material cards. A homepage, stale task, blank page or printed URL is not a configuration handoff.

1. Await Web input with host wait (product-v1-workflow.md).

Use same profile/task/version; follow its running exec session. Configuration needs source=web_user, user_confirmed=true, readiness.ready=true; choices need saved selections and comments. Prefills, old receipts and unrelated events are not answers. If pending, wait from task_version; stay responsive. Continue when valid; stop on cancellation/interruption. Web cannot wake a finished host turn.

1. Later motivation and figure choices follow the same request → Web save →

read-back route. Respect existing explicit delegated-choice authorization; never label an Agent choice as a user click. On interruption, read the same task's snapshot and current pending input before continuing.

If startup fails, repair that concrete runtime/browser problem and retry the same binding. Do not create a replacement task or infer missing configuration. For an existing configured task, independent local corrections within its saved scope can continue while Web is unavailable. New research/writing that depends on pending configuration, material selection or figure decisions must wait. If the user explicitly requests restarting from scratch, create fresh task state for the same original inputs; do not silently resurrect deleted task IDs.

Use the chosen browser for paper decisions and results; keep internal grants, receipts and Runner stages out of user chores.

Work continuously within the confirmed scope

Use applicable protocol steps, not 17 forms or backend gates. Check artifacts and downstream use; repair the source at the defect location and compare outputs. After two unchanged attempts, change tactic, not notes/status/hash. Reuse valid saved work.

Mandatory before the next work segment or stage: verify that the previous segment's real results have reached the same task's workbench. Read the current snapshot; publish any completed deliverable that the user should see; call paperspine_commit_milestone with a factual summary, the appropriate existing stage and actual artifact IDs. Then read back the same task/profile and current version: confirm its stage, milestone summary, visible files and pending choices match what actually happened. Only then begin the next dependent segment. Several protocol steps may share one existing stage; do not create 17 backend stages or write a milestone for every sentence/tool call. Reuse an already matching checkpoint rather than duplicating it. On resume apply this check first. Writing or publishing files alone does not advance the workbench. If read-back fails, fix the concrete binding/sync error before claiming progression; keep local results and continue only independent work within saved scope. Never fake review/delivery completion to move a progress indicator. Follow the exact calls in product-v1-workflow.md.

Read the saved configuration.workflow and research_mode before choosing work. Use build_from_materials to develop the paper; rewrite_existing to improve the existing argument and prose while retaining valid results; audit and review to deliver findings on the supplied files; revise to address actual feedback; and transfer to adapt the existing science to the new venue. Read the matching playbook. Reuse valid configuration, choices and results: a workflow label does not require another intake, new analyses, or an entirely new manuscript.

Choose tools, intermediate files, work order and repair tactics with scholarly judgment. The route below describes the work the paper needs; it is not a rigid sequence of approvals. Combine related work and run independent retrieval, analysis or rendering in parallel when useful. Draft, figures and evidence can inform each other while pending user choices remain unassumed.

Pause dependent work only for an actual missing user decision, indispensable fact, or unavailable capability. Bundle related questions, reuse saved answers and explicit delegated-choice authority, and continue unaffected work. Ordinary search, template adaptation, wording, chart construction and local repairs within the saved scope do not need another confirmation. Never silently skip the initial Web configuration/material save or choose on the user's behalf in guided work.

Keep compact notes and valid outputs, not every example form or receipt. Read applicable methods fully on first use. On resume, recover saved choices and the next action; reopen needed original passages if their content was lost, even when files are unchanged. A read marker is not retained knowledge. When available and authorized, delegate bounded work with real inputs and disjoint writes; inspect returned artifacts and manuscript use yourself. A worker's self-check is not independent review. No fixed team is required.

Distinguish dependencies from permissions; retry with a concrete correction. Recheck affected outputs and stop optional polish without useful improvement.

Scientific work and expected results

Use current-method-routing.md to select applicable original methods. Read their full text on first use; routing is not application or quality proof. If the helper is unavailable, read the local originals and continue the same task.

1. Anchor, intake and inventory

Read intake.md, intent-configuration.md and resume.md as applicable. Preserve originals. Inventory results, draft, figures, references, scripts, provenance, permissions and limits; identify only necessary missing decisions. Do not create placeholders.