Full-Cycle Multi-Agent Workflow And Journal Fit Design

Goal

Make Qiongli support a strict end-to-end academic paper workflow that can move from topic selection to broad evidence search, data and methods, manuscript writing, review, strong judging, journal recommendation, and revision feedback without losing stage state or drifting away from locked decisions.

This becomes the next priority roadmap slice before additional subject expansion work. The reason is simple: subject expansion, feedback-aware routing, and marketplace readiness should be validated against a full-cycle harness rather than only against isolated stage tests.

The first implementation should be a thin vertical slice. It should prove the control surface, evidence handoff, judge gates, and reverse journal-fit recommendation path. It does not need to generate a complete publishable paper in one run.

Current Context

The repository already has most of the building blocks:

• The canonical workflow contract defines stages A, B, C, D, E, F, G, J, H, I, and K.
• /paper routes users to individual task IDs such as A1, B1, C1, F3, G3, J4, H3, and H4.
• Stage handoff rules require context/stage_handoff.md with completed artifacts, decisions, unresolved questions, evidence dependencies, and revisit triggers.
• context_package already distributes boundary review and writing harness data to multiple agents.
• qiongli_task_run is preview-first and can run local agents only when explicitly requested.
• Literature provider work now separates discovery, full-text candidates, Zotero attachment verification, and evidence limits.
• venue-analyzer supports target-venue-first analysis early in Stage A.
• Stage H already contains submission packaging, peer-review simulation, and fatal flaw detection.
• Venue profiles exist for core and subject-specific venues.

The gap is orchestration quality. Qiongli can perform many pieces, but it does not yet have a release-grade full-cycle harness that proves stage artifacts, agent roles, judge decisions, and journal recommendations remain aligned across multiple rounds.

Product Decision

Create a new full-cycle workflow line and merge it into the existing roadmap as the next priority before the remaining subject expansion stages.

The design has two linked capabilities:

1. Full-cycle multi-agent workflow harness
2. Reverse journal-fit recommendation from an existing manuscript

The full-cycle workflow owns the long-running state machine. Reverse journal fit is a required checkpoint near submission, but it can also be invoked directly for an already written manuscript.

Non-Goals

• Do not replace /paper as the single-task router.
• Do not make one command silently run a complete paper project by default.
• Do not launch real local agents unless run_agents=true and the existing runtime safety checks pass.
• Do not claim exhaustive literature coverage only because metadata search ran.
• Do not recommend a journal from title or abstract alone when manuscript, methods, evidence, or claim maps are missing.
• Do not recommend a top venue based only on impact, topical salience, or user preference.
• Do not bypass the existing workflow contract, stage handoff contract, or claim-evidence ledger.
• Do not fully implement new subject activation in this slice.

Product Model

Existing single-task path

Users can still run focused tasks:

/paper topic
/lit-review topic
/paper-write topic
/submission-prep topic

Those commands remain useful for local work.

New full-cycle path

The new path plans and checks a whole paper lifecycle:

/paper-lifecycle topic

The first version should support preview mode as the default. It produces a lifecycle plan, stage gate report, required artifact list, and next-action recommendations without launching agents.

Optional local-agent execution remains explicit:

{
  "run_agents": true,
  "execution_mode": "triad",
  "controller": "codex",
  "primary": "claude",
  "reviewer": "antigravity"
}

Direct reverse journal-fit path

Users with an existing manuscript can ask:

Recommend the best journal for this manuscript.

The system should route this to reverse journal fit, not to early-stage venue-analyzer, unless the user is still choosing a target before writing.

Full-Cycle State Machine

The full-cycle workflow is a state machine over existing stages. It should not invent new academic stages when an existing stage can hold the work.

Required lifecycle checkpoints:

Checkpoint	Main stages	Required proof
Idea lock	A1, A2, A4, A5	research question, contribution, boundary review, initial venue assumptions
Evidence base	B1, B2, B3, B6	search plan, search log, dedup log, retrieval manifest, literature map
Design and data	C1, C3, C4, optional I3-I8	study design, variable or construct spec, data plan, analysis or reproducibility status
Manuscript build	F1-F6	outline, draft, claim-evidence map, figures or tables plan
Compliance and proofread	G1-G4, J1-J4	reporting checklist, cross-section integrity, tone/proofread reports
Strong judge	H3, H4	peer review simulation and fatal flaw analysis
Journal fit	new H5	ranked journal fit report with evidence-based recommendation
Feedback loop	H2, H2_5	response matrix, revision plan, reviewer empathy check

The state machine should expose these statuses:

• not_started
• blocked_missing_artifact
• blocked_unresolved_boundary
• ready_for_agent
• ready_for_human_review
• passed
• reopened_by_revisit_trigger

Stage Handoff And Drift Control

Every stage transition must read and update:

• context/research_state.md
• context/decision_log.md
• context/boundary_review.md
• context/stage_handoff.md
• evidence/claim-evidence-ledger.csv

The full-cycle harness must block or warn when:

• a downstream task changes the research question without a decision log entry,
• a manuscript claim has no evidence ledger row,
• a methods or data claim exceeds the stage C/I evidence status,
• a journal recommendation ignores current manuscript limits,
• a strong judge reports a fatal flaw but the lifecycle status still says submission-ready,
• a revision promise is made without a source artifact or feasibility note,
• an agent output omits the locked non-goals or claim-strength boundary.

Multi-Agent Roles

The first full-cycle workflow should support these logical roles:

• controller: builds the lifecycle plan, assigns tasks, and resolves stage status.
• evidence_agent: checks literature, full-text status, Zotero verification, and evidence ledger coverage.
• methods_agent: checks design, data, analysis plan, and reproducibility commitments.
• writer_agent: drafts or revises manuscript text under the writing harness.
• review_agent: reviews cross-section integrity, compliance, and reviewer risks.
• strong_judge: cannot draft text; it can only block, pass, or request revision with evidence-backed reasons.
• journal_fit_agent: ranks journals from manuscript and venue-profile evidence.

These are roles, not mandatory separate model processes. In solo mode, Qiongli can execute them sequentially with role gates. In duo or triad mode, the context package must make each role's allowed actions explicit.

Strong Judge Contract

The strong judge is a gate, not a coauthor.

Allowed outputs:

• pass
• revise
• block_submission
• reopen_stage

Required fields:

• decision
• severity
• affected claim or artifact
• evidence basis
• missing artifact if any
• required revision
• stage to reopen if blocked
• whether journal recommendation remains valid

The strong judge must not:

• rewrite the manuscript directly,
• invent reviewer comments,
• recommend a venue without reading the journal fit report,
• override a locked decision without a revisit trigger.

Reverse Journal-Fit Recommendation

Problem

Current venue support is mostly target-first. A user can name a target venue, and Qiongli can analyze fit and formatting implications. The missing path is manuscript-first: read an existing draft and recommend the best venues.

New task

Add a canonical Stage H task:

H5: Reverse journal-fit recommendation

Primary output:

submission/journal_fit_recommendation.md

Optional machine-readable output:

submission/journal_fit_recommendation.json

Required inputs

The recommender should require enough evidence to avoid shallow ranking:

• manuscript draft or structured manuscript sections,
• title and abstract if available,
• research question and contribution statement,
• methods, data, or evidence design summary,
• claim-evidence map,
• reporting or compliance status if available,
• current limitations or fatal flaw report if available,
• candidate subject or discipline,
• venue profile catalog.

If the manuscript is incomplete, the recommender should return a blocked or partial report. It may suggest what to collect next, but it must not claim a best journal.

Ranking dimensions

Each candidate journal should be scored and explained on:

• scope fit,
• contribution fit,
• method and evidence fit,
• article type fit,
• audience fit,
• reporting and data-policy fit,
• manuscript maturity,
• reviewer-risk fit,
• desk-reject risk,
• required revision before submission,
• stretch or safe positioning.

Recommendation classes

The ranked report should include:

• primary: best realistic fit,
• stretch: attractive but high-risk venue,
• safe: good fit with lower positioning risk,
• fallback: useful if top options are blocked,
• do_not_submit: venues where current manuscript is a poor fit.

The recommender must explain why a higher-status journal is not primary when the manuscript evidence does not support it.

Harness Modes

Preview harness

Default mode. It reads fixtures and project artifacts, then reports stage and journal-fit gate status. It must not call external providers or local agents.

Provider-connected harness

Optional mode. It can call literature provider tools to verify search plan, coverage diagnostics, and full-text candidate status. It still must not launch writing agents unless explicitly requested.

Local-agent harness

Opt-in mode. It can run a bounded lifecycle slice with local agents after the same safety model as qiongli_task_run.

Release harness

Release checks should start with deterministic preview fixtures. Provider and local-agent variants should be maintainer opt-in until they are reliable on CI or a controlled runner.

Harness Report Schema

The full-cycle harness should produce a JSON report:

{
  "schema_version": "1.0",
  "mode": "preview",
  "topic": "demo-paper",
  "paper_type": "empirical",
  "lifecycle_status": "blocked_missing_artifact",
  "stage_gates": [
    {
      "stage": "B",
      "status": "passed",
      "required_artifacts": ["search_strategy.md", "search_log.md"],
      "missing_artifacts": [],
      "warnings": []
    }
  ],
  "drift_checks": {
    "locked_question_preserved": true,
    "claim_evidence_coverage": "partial",
    "unresolved_judge_blocks": 1
  },
  "journal_fit": {
    "status": "partial",
    "primary": null,
    "blocking_reasons": ["missing claim-evidence map"]
  },
  "recommended_next_tasks": ["F4", "H4", "H5"]
}

Markdown reports should be concise and point to the exact missing artifact or stage to reopen.

Fixture Strategy

The first implementation should add deterministic fixtures for:

• clean full-cycle preview path,
• missing literature coverage,
• manuscript drift from locked research question,
• unsupported claim in manuscript,
• fatal flaw unresolved before submission,
• journal overreach where a top venue is not justified,
• incomplete manuscript where journal fit must block,
• revision feedback reopening an earlier stage.

Fixtures should use small text files and static venue profiles. They should not depend on external provider calls or model output.

Roadmap Integration

The adaptive subject roadmap should be reconciled as follows:

1. Mark the local-agent smoke and subject-gate foundations as at least partially implemented on dev.
2. Insert the full-cycle workflow harness as the next priority stage.
3. Treat reverse journal fit as a required subfeature of that stage.
4. Resume subject expansion after the full-cycle harness can test stage handoff, drift prevention, and journal-fit gates.
5. Keep feedback-aware subject routing and marketplace/read-only fallback as downstream work, but make them report into the full-cycle harness once they are implemented.

Success Criteria

• A preview full-cycle harness can evaluate a sample project without launching agents.
• The harness blocks when required stage artifacts are missing.
• The harness detects manuscript drift from a locked research question.
• The harness reports unsupported manuscript claims from the claim-evidence ledger.
• The reverse journal-fit recommender ranks at least three candidate journals from venue profile evidence.
• The recommender blocks best-journal claims when manuscript evidence is too thin.
• Strong judge findings can reopen a specific earlier stage.
• /paper remains backward compatible as a single-task router.
• Generated docs and workflow packages expose the new path consistently.

Risks And Mitigations

• Scope creep: start with preview harness and one empirical fixture.
• False confidence: block best-journal recommendation when required manuscript evidence is absent.
• Agent drift: use context package role constraints and judge gates.
• Overly rigid workflow: allow stage reopen with explicit revisit triggers.
• Provider instability: keep provider-connected harness optional at first.
• Subject expansion coupling: keep subject-specific logic behind existing subject contracts and venue profiles.

Rollout

1. Documentation and workflow contract update.
2. Deterministic lifecycle harness preview.
3. Reverse journal-fit recommender over local venue profiles.
4. MCP/CLI exposure for preview-only calls.
5. Optional local-agent lifecycle slice.
6. Release-readiness integration after preview fixtures stabilize.