Food Pipeline

Food-Pipeline — Master Research-to-Publication Orchestrator

The top-level conductor. It does not do research, writing, or review itself — it routes the project to the specialist skills (each a multi-subagent system), enforces quality gates between stages, and manages the review→revise loop. Original work.

Skills it orchestrates (each brings its own subagent team)

• journal-selector — target-journal constraints (structure, limits, reference style, figure spec).
• food-research — literature/evidence synthesis (quick brief / full review / systematic PRISMA + OHAT). Use food-deep-research instead for an open-ended, source-validated deep dive or a standalone literature review.
• food-paper — whole-process manuscript system (field → questions → data/stats → figures → argument → draft → polish → self-review).
• food-figure — submission-grade figures at the journal spec (invoked within food-paper).
• food-review — multi-reviewer peer-review panel + formatting compliance.

Own subagents

• intake_router — reads the project's current state and materials, resolves the target journal, picks the entry stage, and assembles the context each downstream skill needs.
• quality_gate — the checkpoint between stages: verifies the stage's deliverable meets the gate criteria (integrity, journal compliance, evidence sufficiency) and decides proceed / revise / stop, with the author at mandatory gates.

Stages

Knowledge reuse — don't research the same field twice

When the pipeline ran Stage 1 (it entered at Stage 0 or 1), the field has already been searched and synthesized. Stages 3 · REVIEW and 5 · RE-REVIEW must therefore carry the Stage-1 evidence base into food-review rather than let its knowledge_builder repeat a full literature search:

• Pass forward the Stage-1 output (food-research / food-deep-research: validated sources, evidence matrix / synthesis, grading, gap list) as the review's field-knowledge foundation; don't re-fetch what Stage 1 already validated.
• Top it up with the food-research quick brief stream to find the field's key review publications and read those reviews in full (state of the art, consensus vs contested, standard methods, benchmark ranges).
• Knowledge base = Stage-1 knowledge + key-review knowledge. The manuscript's own cited sources are still read and audited (Pathway A), reusing Stage-1 records where the source was already retrieved.

If Stage 1 did not run (entry at Stage 2/3 with a finished draft), there is nothing to inherit — food-review builds its knowledge base the full way (Pathway A + B). Using food-review standalone is unaffected by this rule. See food-review/agents/knowledge_builder.md.

Review & revision defaults (Stage 3 onward) — explicit authorization

Default: one review→revise round, then FINALIZE. Do not auto-run a second round or silently rewrite the author's original Word file.

Ask once (consolidate) before Stage 4 when a .docx (or equivalent) is in play:

1. Second review round? Default no. Run Stage 5 (RE-REVIEW) only if the author explicitly authorizes it. Hard cap remains 2 rounds total.
2. Edit the original Word with Tracked Changes? Default no. Only modify the original manuscript in place when the author explicitly authorizes it. Without that authorization: deliver a revised copy (or a change log / marked draft) plus the Review & Response Report — leave the original file untouched.

Deliverables — exactly two files, both Word (.docx)

The pipeline produces one manuscript and one report. Never a separate review report and a response letter; never Markdown.

1. One manuscript file (.docx). Revisions are Tracked Changes on that single original Word file when authorized (otherwise a revised copy); food-review adds margin comments to that same file each round, and every Editor query item gets a comment/note at its location.

2. One Review_and_Response_Report_<slug>_<date>.docx — the same document evolving through the stages, in the canonical food-review/references/report-format.md structure (Parts A/B/C; stable issue IDs; precise locations; colour legend):

   • Stage 3 (REVIEW) — food-review writes the reviewer feedback (black): every concern with its ID and location, plus the editorial decision.
   • Stage 4 (REVISE) — food-paper updates that same file in place, filling each item's Response (<type>) (blue) = Tracked edit · Editor query · Recommendation · Residual, with what was actually done and where.
   • Stage 5 (RE-REVIEW), if authorized — append R2-* items to the same file.

   The result carries both the reviewer feedback and the editing response in one document, labelled by round. Do not create a separate reviewer report, and do not create a standalone response letter — this report is the response. (A point-by-point letter to a journal's editor is only produced by food-paper revise standalone, responding to real reviewers.)

Markdown is a working format only: convert with Pandoc (pandoc report.md -o report.docx) or the docx skill, and never claim a .docx you did not produce. Run scripts/privacy_scan.py on every file before delivery.

See food-review/references/report-format.md, food-review/references/word-review-comments.md, and food-paper/references/revision-response.md.

Workflow

flowchart TD
    A[Project in] --> R[intake_router<br/>state + materials + journal + entry stage]
    R --> J[journal-selector]
    R --> S1[Stage 1 RESEARCH<br/>food-research / food-deep-research]
    S1 --> G1{quality_gate<br/>evidence sufficient?}
    G1 -- yes --> S2[Stage 2 WRITE<br/>food-paper -> food-figure]
    G1 -- no --> S1
    S2 --> G2{quality_gate<br/>integrity + journal compliance}
    G2 -- pass --> S3[Stage 3 REVIEW<br/>food-review panel]
    G2 -- fail --> S2
    S3 --> G3{{author decision<br/>mandatory gate}}
    G3 -- revise --> S4[Stage 4 REVISE<br/>food-paper revise]
    S4 --> G4{{author: second round?}}
    G4 -- no / default --> S6[Stage 6 FINALIZE<br/>format + Word export]
    G4 -- yes authorized --> S5[Stage 5 RE-REVIEW<br/>food-review re-review]
    S5 -- issues + author continues --> S4
    S5 -- accept --> S6
    G3 -- accept --> S6
    S6 --> OUT[Submission-ready manuscript]

Entry points (mid-pipeline)

intake_router detects where to start: a topic/dataset → Stage 1; a full draft → Stage 2 or 3; reviewer comments in hand → Stage 4. It never restarts completed stages unnecessarily. At Stage 3/4 it records whether the author has authorized a second round and/or in-place tracked changes on the original Word file.

References (load as needed)

• references/mode-advisor.md — intake_router uses it to pick entry stage, research flavor, and skills.
• references/pipeline-state-machine.md — states, transitions, entry points, loop caps.
• references/quality-gates.md — the per-stage gate criteria quality_gate applies.

Rules

• Journal first, journal throughout: re-flow references and re-check limits whenever the target journal changes.
• Gates are real: quality_gate can send a stage back; integrity and review gates cannot be skipped, and the review decision is always the author's.
• One round by default: do not auto-run RE-REVIEW; a second round needs explicit author authorization (hard cap 2).
• Original Word is opt-in: do not apply tracked changes to the author's original file unless they authorize it; otherwise leave the original untouched and deliver a revised copy / change log + the Review & Response Report.
• Food-science standards everywhere: n and error type, validated methods, panel details, ethics/food-safety — enforced at every write/review gate.
• Don't duplicate work: the specialist skills own their subagents; the pipeline sequences and gates them, it does not re-implement them.
• Two deliverables, both .docx, always: one manuscript and one Review & Response Report carrying reviewer feedback and the editing response. Never a separate reviewer report or a standalone response letter; never Markdown (see "Deliverables").