Skip to content
neuroflow0.2.31 · docs

← neuroflow

Agents

Agents are autonomous subprocesses launched by commands when deeper, focused work is needed.

Unlike commands (which interact conversationally with you), agents operate semi-autonomously on a specific task — searching a database, auditing a folder, running a check. A command invokes an agent, the agent does the work, and the result flows back to you.

Available agents

Nine agents ship with neuroflow: scholar, sentinel, sentinel-dev, literature-review, paper-writer, paper-critic, poster-critic, autoresearch and flowie. Their full definitions are under Agents in the navigation.

scholar

Academic literature research specialist.

Searches PubMed and bioRxiv in parallel for a given topic — with CrossRef / Semantic Scholar / arXiv fallbacks through the same MCP server — and returns a clean, structured list of results.

Invoked by: ad-hoc literature searches outside the ideation workflow, and /neuroflow (journal recommendation step). Note: /neuroflow:ideation runs its literature searches inline — it does not spawn this agent.

Full definition: scholar

What it does:

  1. Runs PubMed and bioRxiv in parallel, then the fallbacks one at a time when bioRxiv returns little
  2. If results are thin, generates 2–3 alternative queries (synonyms, narrower/broader terms)
  3. Deduplicates results across sources
  4. Downloads only open-access copies (preprint server, Semantic Scholar open-access link, publisher open access) in batches of 2, never Sci-Hub; papers without an open copy keep a metadata stub and a note to get them through the library
  5. Returns results in a structured format with markers:
  6. ⚠️ PREPRINT — bioRxiv papers that have not been peer-reviewed
  7. 🔒 PAYWALLED — papers without open-access full text
  8. Ends every reply with a [REPORT downloaded=... files=... stubs=...] line

Output format:

PubMed results
──────────────
**N2 and P300 in auditory attention** (2023) — Smith et al.
*NeuroImage* | DOI: 10.1016/j.neuroimage.2023.001
Shows P300 amplitude reduces with cognitive load in selective attention tasks.

bioRxiv results
───────────────
**Noise effects on ERP** (2024) — Jones et al.
*bioRxiv* | DOI: 10.1101/2024.001
⚠️ PREPRINT — White noise as stressor reduces P300 in healthy adults.

Summary: P300 attenuation under high cognitive load is consistent in the
literature. White noise as a specific stressor is understudied — a gap exists.

Follow-up actions after results:

Action What happens
"download" Fetch full text for open-access papers (skips paywalled)
"save" / "md" Save as literature-[topic]-[date].md in .neuroflow/ideation/
"summarize" Deeper synthesis: main findings, methodological patterns, contradictions

Rules: - Never fabricates papers, authors, or DOIs - Each DOI is labelled with the check actually done (doi_check: from an API record, resolves on a date, does not resolve), never "verified" - PubMed and bioRxiv results are always presented separately

No credentials required

PubMed and bioRxiv search is handled by the paper-search-mcp-nodejs server — no credentials needed.

sentinel

Project coherence guard.

Audits .neuroflow/ for internal consistency and drift. Called by the /neuroflow:sentinel command when in a project repository (not a plugin repository).

Invoked by: /neuroflow:sentinel (when .neuroflow/ exists)

Full definition: sentinel

What it checks: it runs nf_check.py first — the deterministic checks NF1–NF8 — then the judgement checks a script cannot make:

  • flow.md completeness — files listed vs files on disk
  • Timestamp drift — stale flow.md vs recent file activity
  • Reasoning logs (reasoning/*.jsonl)
  • Phase consistency — active phase vs session logs vs folder activity
  • Preregistration drift — planned analyses vs what was actually done
  • Plugin version — an older or missing plugin_version in project_config.md is reported with a pointer to /neuroflow:migrate; sentinel never edits the field
  • Subfolder name validation — no unrecognized or skill-named folders
  • CLAUDE.md neuroflow reference check

See /sentinel for full documentation.

sentinel-dev

Plugin development coherence guard.

Audits the neuroflow plugin repository itself for structural consistency. Called by the /neuroflow:sentinel command when run inside the plugin repo (where .claude-plugin/plugin.json exists).

Invoked by: /neuroflow:sentinel (when .claude-plugin/plugin.json exists)

Full definition: sentinel-dev

What it checks: it runs scripts/automation/validate_pr.py first — V1–V15, the same checks CI runs on every pull request — then the judgement checks (hooks documentation, real names and institutions, guards versus prose):

  • Command folder names vs frontmatter name: fields
  • Skill folder names vs SKILL.md frontmatter
  • README tables vs actual files
  • Version sync between plugin.json and all references
  • Dead references (links to files that don't exist)
  • Command frontmatter completeness

literature-review

Structured literature review of downloaded papers.

Runs 12 analytical protocols on the papers in .neuroflow/ideation/papers/, from landscape mapping to a future research agenda. Each protocol is checked by a rubric critic pass inside the agent (a self-check) and saved as a resumable checkpoint in .neuroflow/ideation/literature-review-[date]/; /ideation can add an independent critic.

Invoked by: /neuroflow:ideation, after papers are retrieved

Full definition: literature-review

paper-writer

Manuscript writer for the unified paper phase.

Drafts neuroscience manuscript sections from upstream project memory, the target journal's guidelines and the analysis results, citing only from the project library. It works inside the write→critique loop with paper-critic and revises until the critic approves or the loop runs out. It returns outlines, drafts and open questions instead of waiting for answers.

Invoked by: /neuroflow:paper

Full definition: paper-writer

paper-critic

Hyper-critical manuscript reviewer.

Applies the full eight-area neuroflow:review-neuro methodology to every section draft and returns [STATUS: APPROVED] or [STATUS: REJECTED] with specific, actionable feedback. It never writes content itself.

Invoked by: /neuroflow:paper, inside the write→critique loop

Full definition: paper-critic

poster-critic

Conference poster reviewer.

Judges the poster's LaTeX source, its compile log, a rendered preview and, when it exists, the compiled PDF against design, content and communication standards, including the QR code and LaTeX correctness. Returns [STATUS: APPROVED] or [STATUS: REJECTED] with actionable feedback.

Invoked by: /neuroflow:poster, in its worker-critic loop (at most 3 cycles)

Full definition: poster-critic

autoresearch

The single managing agent of the improvement loop.

Makes one focused change to a research artifact per iteration, judges it against the current best itself (no worker or evaluator fan-out), and keeps or reverts it. Its memory is a per-loop wiki it reads before and writes after every move. It never stops on its own judgement: it stops at the caps set in the loop's config, when you stop it, or after repeated errors. The protocol lives in the neuroflow:autoresearch-protocol skill.

Invoked by: /neuroflow:autoresearch

Full definition: autoresearch

flowie

Personal identity agent.

Reads your flowie profile (~/.neuroflow/flowie/profile.md) and shapes the work to who you are: your research stances, writing style and methodological preferences. At the start of a session it surfaces your active tasks for the current project and notes a stale sync. Profile data never appears verbatim in external-facing outputs such as manuscripts, grants or reports.

Invoked by: Claude, when a linked flowie profile should shape the work; see Your profile

Full definition: flowie

Per-phase agents were removed

Earlier versions listed one agent per research phase. Those agent files were never spawned by commands and have been removed: each command follows its phase skill directly.

How agents differ from commands

Commands Agents
Invoked by You (directly) Commands (programmatically)
Interaction style Conversational Autonomous task execution
Scope Phase-level work Focused sub-task
Output Phase subfolder in .neuroflow/ Result returned to calling command