gpd-literature-review

name	gpd-literature-review
description	Structured literature review for a physics research topic with citation network analysis and open question identification
argument-hint	[topic or research question]
context_mode	project-aware
allowed-tools	["read_file","write_file","shell","grep","glob","web_search","web_fetch","ask_user"]

<codex_runtime_notes> Codex shell compatibility:

• When shell steps call the GPD CLI, use /Users/charlie/.gpd/venv/bin/python -m gpd.runtime_cli --runtime codex --config-dir ./.codex --install-scope local instead of the ambient gpd on PATH.
• If you intentionally need the repo environment, keep the runtime pin: GPD_ACTIVE_RUNTIME=codex uv run gpd .... </codex_runtime_notes>

<codex_questioning>

• Ask each user-facing question exactly once.
• Present options once.
• Do not restate the prompt or add meta narration. </codex_questioning>

Conduct a systematic literature review for a physics research topic. Identifies key papers, maps citation networks, catalogs methods and results, finds open questions, and produces a structured LITERATURE-REVIEW.md.

Orchestrator role: Scope the review, spawn gpd-literature-reviewer agent, handle checkpoints, present results.

Why subagent: Literature searches burn context fast (reading abstracts, following citation chains, cross-referencing results, tracking conventions across papers). Fresh 200k context for the full survey. Main context stays lean for user interaction.

A physics literature review is not a bibliography. It is a map of the intellectual landscape: who computed what, using which methods, with what assumptions, getting what results, and where do they agree or disagree. The reviewer must think like a physicist surveying a field, not a librarian cataloging references.

<execution_context>

Conduct a systematic literature review for a physics research topic. Map the intellectual landscape: foundational works, methodological approaches, key results, controversies, and open questions. Produce LITERATURE-REVIEW.md consumed by planning and paper-writing workflows.

Called from $gpd-literature-review command.

<core_principle> A physics literature review is not a bibliography. It is a structured map of who computed what, using which methods, with what assumptions, getting what results, and where they agree or disagree. The goal is to understand the state of a field well enough to identify what is known, what is open, and where new work can contribute. </core_principle>

<source_hierarchy> MANDATORY: Authoritative sources BEFORE general search

1. Textbooks and monographs -- For established results, standard methods, and field context

   • Use specific textbooks by subfield (Peskin & Schroeder for QFT, Sakurai for QM, Jackson for E&M, Landau & Lifshitz series, etc.)
   • These define conventions and standard results

2. Review articles -- For field overviews and method surveys

   • Rev. Mod. Phys., Physics Reports, Annual Reviews of Physics
   • Recent reviews (last 5 years) for current state-of-the-art

3. Seminal papers -- Original derivations of key results

   • Identify the papers everyone in the field cites
   • Read the actual papers, not just the citations

4. Recent arXiv preprints -- For cutting-edge developments

   • arXiv categories: hep-th, hep-ph, hep-lat, cond-mat., quant-ph, gr-qc, astro-ph., nucl-th, etc.
   • Sort by relevance and citation count

5. Conference proceedings -- For very recent results and community direction

   • Lattice, ICHEP, APS meetings, etc.

6. web_search -- Last resort for community discussions, code repos, numerical benchmarks

</source_hierarchy>

**Load project context (if available):**

INIT=$(/Users/charlie/.gpd/venv/bin/python -m gpd.runtime_cli --runtime codex --config-dir ./.codex --install-scope local init progress --include state,config)
if [ $? -ne 0 ]; then
  echo "ERROR: gpd initialization failed: $INIT"
  # STOP — display the error to the user and do not proceed.
fi

Parse JSON for: commit_docs, state_exists, project_exists, project_contract, contract_intake, effective_reference_intake, active_reference_context, reference_artifact_files, reference_artifacts_content.

Read mode settings:

AUTONOMY=$(/Users/charlie/.gpd/venv/bin/python -m gpd.runtime_cli --runtime codex --config-dir ./.codex --install-scope local --raw config get autonomy 2>/dev/null | /Users/charlie/.gpd/venv/bin/python -m gpd.runtime_cli --runtime codex --config-dir ./.codex --install-scope local json get .value --default balanced 2>/dev/null || echo "balanced")
RESEARCH_MODE=$(/Users/charlie/.gpd/venv/bin/python -m gpd.runtime_cli --runtime codex --config-dir ./.codex --install-scope local --raw config get research_mode 2>/dev/null | /Users/charlie/.gpd/venv/bin/python -m gpd.runtime_cli --runtime codex --config-dir ./.codex --install-scope local json get .value --default balanced 2>/dev/null || echo "balanced")

Mode-aware behavior:

• research_mode=explore: Comprehensive review (30+ papers), include tangential fields, map full citation network, identify open questions.

• research_mode=exploit: Focused review (8-12 papers), direct relevance only, extract key results and methods.

• research_mode=adaptive: Start with 15 papers, expand if citation network reveals critical gaps.

• autonomy=supervised: Pause after each review round for user feedback on scope and direction.

• autonomy=balanced (default): Complete the full review pipeline automatically. Pause only if the literature reveals scope ambiguity, contradictory evidence, or a change in recommendation.

• autonomy=yolo: Complete the review pipeline without pausing, but do NOT drop contract-critical anchors or user-mandated references.

• If state_exists is true: Extract convention_lock for notation context (helps identify which conventions are used in papers being reviewed). Extract active research topic, phase context, and any contract-critical references from active_reference_context.

• If state_exists is false (standalone usage): Proceed — the user will specify the topic directly.

• Treat effective_reference_intake as the machine-readable carry-forward ledger for anchors, prior outputs, baselines, user-mandated context, and unresolved gaps. Re-surface those items in the review even if the broader search expands beyond them.

• Use reference_artifacts_content as supporting evidence when existing literature/research-map artifacts already pin down benchmark values, prior outputs, or anchor wording that should remain stable.

Project context helps focus the review on conventions and methods relevant to the current research.

Establish scope from command context:

• Topic and focus: Specific physics question or subfield
• Depth: Quick (~10 refs) | Standard (~30 refs) | Comprehensive (~50+ refs)
• Time range: All time | Last N years | Since specific result
• Purpose: Background | Method selection | Gap identification | Manuscript prep
• List the seed anchors already present in project_contract, contract_intake, effective_reference_intake, and active_reference_context before broadening the search

Define explicit include/exclude boundaries:

• Include: specific phenomena, methods, energy ranges, dimensions
• Exclude: tangential fields, historical reviews (unless depth=comprehensive)
• Record any contract-critical anchor that must be surfaced even if it falls outside the default search breadth

**Phase 1: Foundational Works**

Every subfield has seminal papers that defined the field. Identify them:

1. Search for review articles first (they cite the seminal works):

   web_search: "[topic] review" site:arxiv.org
   web_search: "[topic]" site:journals.aps.org/rmp
   

2. From review articles, extract:

   • The 5-10 most-cited papers
   • The textbook treatments
   • The original derivation of key results

3. For each foundational work, record:

   • Full citation (authors, title, journal, year)
   • Stable anchor_id and concrete locator if the work is contract-critical or likely to be reused downstream
   • Key contribution (what they showed/computed/proved)
   • Method used
   • Conventions (units, metric signature, normalization)
   • Where the result is used downstream
   • Whether it should be treated as a contract-critical anchor for later planning or verification

4. Build a citation timeline showing how the field developed.

**Phase 2: Methodological Landscape**

Catalog all methods that have been applied to this problem:

For each method:

Field	Detail
Method name	Formal name and common abbreviations
Type	Analytical / Numerical / Mixed
Key idea	One-sentence description of the approach
Regime of validity	Where it works (weak coupling, high T, large N, etc.)
Limitations	Where it fails (strong coupling, low dimension, sign problem, etc.)
Accuracy	Typical precision achievable
Computational cost	Scaling with system size, time, memory
Key references	Original paper + best application to this system
Available codes	Open-source implementations, if any

Organize methods by approach type:

• Exact methods: Bethe ansatz, integrability, conformal bootstrap, etc.
• Perturbative: Weak coupling, 1/N, epsilon expansion, etc.
• Variational: Trial wavefunctions, DMRG, tensor networks, etc.
• Monte Carlo: DQMC, PIMC, VMC, AFQMC, etc.
• Mean-field and beyond: Hartree-Fock, RPA, GW, DMFT, etc.
• Effective theories: EFT, renormalization group, etc.

Note which methods agree and where they disagree -- this reveals the interesting physics.

**Phase 3: Key Results Catalog**

For each significant result in the literature:

Field	Detail
Quantity	What was computed (energy, correlation function, phase boundary, etc.)
Value	Numerical result or analytical expression
Method	How it was obtained
Uncertainty	Error bars, systematic uncertainties, convergence status
Conventions	Units, normalization, sign conventions used
Regime	Parameter values, approximations in effect
Reference	Full citation
Agreement	How it compares with other determinations

Tabulate results for the SAME quantity across different papers/methods to expose:

• Agreement (convergence of independent methods)
• Disagreement (controversial values)
• Trends (how results evolved as methods improved)
• Which values are decisive benchmarks versus optional background comparisons

**Phase 4: Citation Network Analysis**

Map intellectual lineages:

1. Method lineages: paper_A -> paper_B -> paper_C (each improving on the previous)
2. Competing approaches: lineage_X vs lineage_Y (different methods for same problem)
3. Reconciliation: papers that compared or unified different approaches
4. Branching points: where the field split into sub-problems

This reveals:

• Which methods are still actively developed (recent citations)
• Which are considered superseded (cited only for historical context)
• Which groups are leading each approach
• Where cross-pollination between approaches has been fruitful

**Phase 5: Controversies and Disagreements**

Actively search for disagreements in the literature:

1. Numerical discrepancies: Different groups get different values for the same quantity

   • How significant is the disagreement? (In sigma)
   • Is the discrepancy resolution-dependent? (Finite-size, continuum limit)
   • Has anyone explained the discrepancy?

2. Methodological disagreements: Different methods give inconsistent results

   • Which method is more reliable in this regime?
   • Are the approximations comparable?
   • Could both be right in different limits?

3. Conceptual disagreements: Different physical interpretations of the same result

   • Is this a genuine physics disagreement or a convention difference?
   • What experiment or calculation could distinguish between interpretations?

4. Convention conflicts: Different papers use different conventions

   • Catalog convention choices across the major references
   • Note where convention mismatches could cause apparent disagreements

**Phase 6: Open Questions**

Systematically identify what has NOT been done:

1. Uncomputed quantities: Observables mentioned in the literature but never calculated
2. Unexplored regimes: Parameter ranges where no reliable method works
3. Unresolved puzzles: Anomalous results with no accepted explanation
4. Missing connections: Two related results that nobody has connected
5. Unverified predictions: Theoretical predictions awaiting experimental confirmation
6. Long-standing conjectures: Claims without proof, supported only by numerical evidence

For each gap:

• Why hasn't it been addressed? (Too hard? Not important enough? Technical obstacle?)
• What would it take to address it? (Better methods? More computing power? New data?)
• What would we learn? (Is it worth the effort?)
• Whether a missing anchor or missing benchmark is currently blocking downstream planning

**Phase 7: Current Frontier**

Map the state-of-the-art:

1. Most recent results (last 1-2 years)

   • What has been computed or measured recently?
   • How does it change the picture from the review articles?

2. Active groups

   • Which groups are producing results in this area?
   • What methods are they using?
   • What are their current projects? (Check recent arXiv submissions)

3. Emerging methods

   • New theoretical or computational approaches being applied
   • Machine learning / AI applications to this problem
   • New experimental techniques

4. Community direction

   • What was discussed at recent conferences?
   • Where is the field heading?

Ensure the output directory exists:

mkdir -p .gpd/literature

Write .gpd/literature/{slug}-REVIEW.md:

---
topic: { topic }
date: { YYYY-MM-DD }
depth: { quick/standard/comprehensive }
paper_count: { N }
status: completed | checkpoint
---

# Literature Review: {Topic}

## Executive Summary

{3-5 key takeaways from the review. What should a physicist entering this area know first?}

## Foundational Works

| #   | Reference                | Year   | Key Contribution   |
| --- | ------------------------ | ------ | ------------------ |
| 1   | {Author et al., Journal} | {year} | {what they showed} |

{Brief narrative connecting these works and showing how the field developed.}

## Methodological Landscape

### Exact Methods

{Description of applicable exact methods, regimes, limitations}

### Perturbative Methods

{Description of perturbative approaches, convergence properties}

### Numerical Methods

{Description of computational approaches, costs, accuracies}

### Effective Theories

{Description of effective theory approaches, energy scales}

### Method Comparison

| Method   | Regime           | Accuracy            | Cost      | Key Reference |
| -------- | ---------------- | ------------------- | --------- | ------------- |
| {method} | {where it works} | {typical precision} | {scaling} | {citation}    |

## Key Results

| Quantity     | Value             | Method   | Reference  | Status                |
| ------------ | ----------------- | -------- | ---------- | --------------------- |
| {observable} | {value +/- error} | {method} | {citation} | {confirmed/contested} |

## Citation Network

{Intellectual lineages showing how ideas evolved. Key branching and merging points.}

## Controversies and Disagreements

### {Controversy 1}

- **The disagreement:** {what's contested}
- **Side A:** {position, evidence, key reference}
- **Side B:** {position, evidence, key reference}
- **Current status:** {resolved/active/dormant}

## Open Questions

1. **{Question}** -- {Why it matters, why it's hard, what it would take}

## Current Frontier

{State-of-the-art: most recent results, active groups, emerging methods}

## Active Anchor Registry

| Anchor ID | Anchor | Type | Source / Locator | Why It Matters | Contract Subject IDs | Required Action | Carry Forward To |
| --------- | ------ | ---- | ---------------- | -------------- | -------------------- | --------------- | ---------------- |
| {stable-anchor-id} | {reference or artifact} | {benchmark/method/background/prior artifact} | {citation, dataset id, or path} | {claim, observable, deliverable, or convention constrained} | {claim-id, deliverable-id, or blank} | {read/use/compare/cite} | {planning/execution/verification/writing} |

`Carry Forward To` is workflow stage scope only. If exact contract subject IDs are known, store them in `Contract Subject IDs` instead of collapsing them into stage labels.

## Convention Catalog

| Convention     | Choice A  | Choice B  | Used By        |
| -------------- | --------- | --------- | -------------- |
| {e.g., metric} | (-,+,+,+) | (+,-,-,-) | {which papers} |

## Recommended Reading Path

For someone entering this area, read in this order:

1. {Textbook chapter for background}
2. {Review article for overview}
3. {Seminal paper for key result}
4. {Recent paper for current state}

## Full Reference List

{Formatted citations, organized by topic/method}

**Phase 8: Citation Verification**

Spawn the bibliographer agent to verify all citations collected during the review. The bibliographer has the hallucination detection protocol, INSPIRE/ADS/arXiv search capability, and BibTeX management expertise needed for citation verification.

Resolve bibliographer model:

BIBLIO_MODEL=$(/Users/charlie/.gpd/venv/bin/python -m gpd.runtime_cli --runtime codex --config-dir ./.codex --install-scope local resolve-model gpd-bibliographer)

Runtime delegation: Spawn a subagent for the task below. Adapt the task() call to your runtime's agent spawning mechanism. If model resolves to null or an empty string, omit it so the runtime uses its default model. Always pass readonly=false for file-producing agents. If subagent spawning is unavailable, execute these steps sequentially in the main context.

task(
  subagent_type="gpd-bibliographer",
  model="{biblio_model}",
  readonly=false,
  prompt="First, read ./.codex/agents/gpd-bibliographer.md for your role and instructions.

Verify all citations in the literature review.

Mode: Audit bibliography

Review file: .gpd/literature/{slug}-REVIEW.md

For every reference listed in the Full Reference List and cited in the body:
1. Run the hallucination detection protocol (Steps 1-5) against INSPIRE, ADS, arXiv
2. Cross-check metadata (title, authors, year, journal, identifiers)
3. Flag any hallucinated or inaccurate citations
4. Correct metadata errors where possible

Write results to .gpd/literature/{slug}-CITATION-AUDIT.md

Return BIBLIOGRAPHY UPDATED or CITATION ISSUES FOUND."
)

If the bibliographer agent fails to spawn or returns an error: Proceed without citation audit. Note in the review summary that citations are unverified. The user should manually check key references against INSPIRE-HEP/ADS.

If CITATION ISSUES FOUND:

• Read the audit report
• Fix or remove hallucinated citations from the review document
• Update corrected metadata in the reference list
• Note unresolvable citations in the return summary

If BIBLIOGRAPHY UPDATED:

• All citations verified, proceed to return results

Return to orchestrator with: - Summary of findings (5-10 lines) - Paper count and coverage assessment - Key takeaways - Identified gaps - Report path - Citation verification status

Format:

## REVIEW COMPLETE

**Topic:** {topic}
**Papers reviewed:** {N}
**Report:** .gpd/literature/{slug}-REVIEW.md

**Key takeaways:**

1. {takeaway}
2. {takeaway}
3. {takeaway}

**Open questions identified:** {N}
**Active controversies:** {N}
**Recommended starting point:** {key reference}
**Citation verification:** {all verified | N issues found -- see .gpd/literature/{slug}-CITATION-AUDIT.md}

If the review is incomplete (too broad, need user guidance):

## CHECKPOINT REACHED

**Type:** decision
**Progress:** Reviewed {N} papers, identified {M} subtopics

### Checkpoint Details

**Decision needed:** The topic branches into {N} distinct subtopics. Which should I focus on?

**Options:**

- **A:** {subtopic A} -- {why relevant}
- **B:** {subtopic B} -- {why relevant}
- **C:** All of them (comprehensive, will take longer)

Commit the report:

PRE_CHECK=$(/Users/charlie/.gpd/venv/bin/python -m gpd.runtime_cli --runtime codex --config-dir ./.codex --install-scope local pre-commit-check --files "${OUTPUT_PATH}" 2>&1) || true
echo "$PRE_CHECK"

/Users/charlie/.gpd/venv/bin/python -m gpd.runtime_cli --runtime codex --config-dir ./.codex --install-scope local commit \
  "docs: literature review — ${topic_slug:-standalone}" \
  --files "${OUTPUT_PATH}"

Where ${OUTPUT_PATH} is the path where LITERATURE-REVIEW.md was written.

<success_criteria>

• Source hierarchy followed (textbooks -> reviews -> papers -> arXiv -> web)
• Foundational works identified with key contributions
• Methods cataloged with regimes, limitations, and key references
• Results tabulated with uncertainties and conventions
• Citation network traced showing intellectual development
• Controversies and disagreements documented
• Open questions identified with feasibility assessment
• Current frontier mapped (recent results, active groups, emerging methods)
• Conventions cataloged across references
• LITERATURE-REVIEW.md created with all sections
• Recommended reading path provided
• Citations verified via gpd-bibliographer (no hallucinated references)

</success_criteria>

</execution_context>

Topic: $ARGUMENTS

Check for existing reviews:

ls .gpd/literature/*.md 2>/dev/null | head -10

0. Initialize Context

CONTEXT=$(/Users/charlie/.gpd/venv/bin/python -m gpd.runtime_cli --runtime codex --config-dir ./.codex --install-scope local --raw validate command-context literature-review "$ARGUMENTS")
if [ $? -ne 0 ]; then
  echo "$CONTEXT"
  exit 1
fi

INIT=$(/Users/charlie/.gpd/venv/bin/python -m gpd.runtime_cli --runtime codex --config-dir ./.codex --install-scope local init progress --include state,roadmap,config)

Extract commit_docs, project_contract, active_reference_context, and any existing reference_artifact_files from init JSON. Resolve reviewer model:

REVIEWER_MODEL=$(/Users/charlie/.gpd/venv/bin/python -m gpd.runtime_cli --runtime codex --config-dir ./.codex --install-scope local resolve-model gpd-literature-reviewer)

1. Scope the Review

Ask the user once using a single compact prompt block to establish scope:

1. Topic and focus -- What specific physics question or subfield? (e.g., "topological insulators in 2D", not just "condensed matter")
2. Depth -- Quick survey (key papers only, ~10 refs) | Standard review (~30 refs) | Comprehensive survey (~50+ refs)
3. Time range -- All time | Last N years | Since specific paper/result
4. Purpose -- Background for new project | Finding methods for specific calculation | Identifying open problems | Preparing manuscript introduction

After all gathered, confirm scope and proceed.

2. Check Existing Reviews

ls .gpd/literature/*.md 2>/dev/null

If exists for same topic: Offer: 1) Update with recent papers, 2) View existing, 3) Start fresh.

If doesn't exist: Continue.

3. Spawn gpd-literature-reviewer Agent

<objective>
Conduct systematic literature review: {topic}

**Scope:**

- Depth: {depth}
- Time range: {time_range}
- Purpose: {purpose}
- Contract-critical anchors: {active_reference_context}
  </objective>

<review_strategy>
A physics literature review follows a structured protocol:

1. **Identify foundational works** -- The seminal papers that defined the field or subfield. These are non-negotiable: every physicist in the area knows them.

2. **Map the methodological landscape** -- What theoretical and computational methods have been applied?

   - Analytical: perturbation theory, exact solutions, variational methods, RG, etc.
   - Numerical: Monte Carlo, exact diagonalization, DMRG, DFT, molecular dynamics, etc.
   - Each method has a regime of validity, characteristic approximations, and known limitations.

3. **Catalog key results** -- For each significant paper:

   - What was computed (observable, quantity, prediction)
   - What method was used (and its limitations)
   - What was found (numerical value, scaling law, phase diagram feature)
   - What conventions were used (units, metric signature, Fourier conventions)
   - How it connects to other results (agrees, disagrees, extends, corrects)
   - Whether it should be treated as a must-surface benchmark or comparison target downstream

4. **Trace citation networks** -- Which papers cite which? Where are the intellectual lineages?

   - Method A lineage: paper1 -> paper2 -> paper3 (progressively refined)
   - Method B lineage: paperX -> paperY (competing approach)
   - Reconciliation papers: where different methods were compared

5. **Identify controversies and disagreements** -- Where do published results conflict?

   - Different numerical values for the same quantity
   - Different phase diagram topologies
   - Competing theoretical explanations
   - Unresolved sign or factor disagreements

6. **Find open questions** -- What has NOT been computed or resolved?

   - Quantities mentioned but never calculated
   - Regimes where no method works reliably
   - Long-standing conjectures without proof
   - Experimental predictions not yet tested

7. **Assess the current frontier** -- What is state-of-the-art right now?
   - Most recent results and their significance
   - Active groups and their focus areas
   - Emerging methods or approaches
     </review_strategy>

<source_hierarchy>

1. **Textbooks and monographs** -- For established results and standard methods
2. **Review articles** (Rev. Mod. Phys., Physics Reports, Annual Reviews) -- For field overview
3. **Seminal papers** -- Original derivations and key breakthroughs
4. **Recent arXiv preprints** -- For current state-of-the-art
5. **Conference proceedings** -- For very recent results and community direction
   </source_hierarchy>

<output>
Write to: .gpd/literature/{slug}-REVIEW.md

Structure:

- Frontmatter (topic, date, depth, paper count)
- Executive Summary (3-5 key takeaways)
- Foundational Works (seminal papers with brief descriptions)
- Methodological Landscape (methods used, regimes, limitations)
- Key Results Catalog (tabulated: paper, method, result, conventions)
- Citation Network (intellectual lineages)
- Controversies and Disagreements (conflicting results)
- Open Questions (what remains unsolved)
- Current Frontier (state-of-the-art)
- Active Anchor Registry (must-read papers, decisive benchmarks, and prior artifacts to carry forward)
- Recommended Reading Path (ordered list for someone entering the field)
- Full Reference List (formatted citations)
  </output>

task(
  prompt="First, read ./.codex/agents/gpd-literature-reviewer.md for your role and instructions.\n\n" + filled_prompt,
  subagent_type="gpd-literature-reviewer",
  model="{reviewer_model}",
  readonly=false,
  description="Literature review {slug}"
)

4. Handle Agent Return

If ## REVIEW COMPLETE:

• Display executive summary and key takeaways
• Report paper count and coverage assessment
• Offer options:
  • "Deep dive on subtopic" -- focus on one aspect found during review
  • "Start research" -- use review to plan a research phase
  • "Find gaps" -- identify what's missing for a specific calculation
  • "Export references" -- format for BibTeX / manuscript

If ## CHECKPOINT REACHED:

• Present checkpoint details to user
• Common checkpoints in literature review:
  • "Found conflicting conventions between major references -- which do you want to adopt?"
  • "Two competing theoretical frameworks -- which is more relevant to your work?"
  • "Need access to a paywalled paper -- can you provide the key results?"
  • "Scope is broader than expected -- should I narrow focus?"
• Get user response
• Spawn continuation agent

If ## REVIEW INCONCLUSIVE:

• Show what was found and what's missing
• Offer options:
  • "Broaden search" -- expand keywords and sources
  • "Narrow focus" -- restrict to specific subtopic
  • "Manual search" -- provide specific papers to include
  • "Accept partial" -- use what was found

5. Spawn Continuation agent (After Checkpoint)

<objective>
Continue literature review: {topic}. Prior state in review file.
</objective>

<prior_state>
Review file path: .gpd/literature/{slug}-REVIEW.md
Read that file before continuing so you inherit the prior search state instead of relying on an inline `@...` attachment.
</prior_state>

<checkpoint_response>
**Type:** {checkpoint_type}
**Response:** {user_response}
</checkpoint_response>

task(
  prompt="First, read ./.codex/agents/gpd-literature-reviewer.md for your role and instructions.\n\n" + continuation_prompt,
  subagent_type="gpd-literature-reviewer",
  model="{reviewer_model}",
  readonly=false,
  description="Continue review {slug}"
)

<success_criteria>

• Review scope established (topic, depth, time range, purpose)
• Existing reviews checked
• gpd-literature-reviewer spawned with structured search protocol
• Checkpoints handled correctly (convention conflicts, scope decisions)
• LITERATURE-REVIEW.md created with all sections
• Key results tabulated with methods and conventions
• Open questions identified
• Citation network mapped
• User knows next steps (research planning, gap analysis, manuscript prep) </success_criteria>