Proposal

Read from the proposal directory (quote paths — they may contain spaces):

PROP="$proposal_dir"

Load:

• "$PROP/config.yaml" — agency, mechanism, language, proposal metadata
• "$PROP/foa_requirements.json" — section requirements and word limits
• "$PROP/sections/objectives.md" — research objectives from /gw:aims
• "$PROP/landscape/competitive_brief.md" — competitive positioning
• "$PROP/landscape/literature.md" — literature context and gap analysis
• "$PROP/sections/preliminary_data.md" — PI's evidence
• "$PROP/sections/bibliography.md" — citation list

Load agency-specific section order:

uv run gw-agency sections "$agency" "$mechanism"

Read citation style from the agency manifest:

• "numbered": Use [1], [2] inline citations
• "author_year": Use (Author et al., Year) inline citations

2. Check multi-model availability

MM_AVAILABLE=0
if uv run mm-detect --json 2>/dev/null | python3 -c "
import json, sys
d = json.load(sys.stdin)
sys.exit(0 if any(v.get('installed') for v in d.get('providers', {}).values()) else 1)
"; then
  MM_AVAILABLE=1
fi

When --no-multi-model is passed, force MM_AVAILABLE=0.

3. Write Project Summary FIRST

The project summary (abstract) is written first as a standalone document with its own word limits. It provides the framing for all subsequent sections.

Before drafting, check whether this section is already complete:

uv run gw-state checkpoint-status "$PROP" proposal_writing project_summary

If next_step is "complete", skip the draft. Otherwise proceed.

Draft sections/project_summary.md:

• Concise statement of the problem and its significance
• Proposed approach and objectives (high-level)
• Expected outcomes and impact
• Respect word limit from agency manifest (typically 250-500 words)

Save the draft checkpoint:

cat sections/project_summary.md | uv run gw-state save-checkpoint "$PROP" proposal_writing project_summary draft

Human checkpoint: Present the project summary for PI review before continuing.

4. Per-section loop

Mark the phase in_progress exactly once before entering the loop:

uv run gw-state update "$PROP" --phase proposal_writing --status in_progress

For each section in the agency's section order:

a. Resume check

STATUS=$(uv run gw-state checkpoint-status "$PROP" proposal_writing "$section")
NEXT=$(echo "$STATUS" | python3 -c "import json,sys; print(json.load(sys.stdin)['next_step'])")
if [ "$NEXT" = "complete" ]; then
  echo "Section $section already done — skipping"
  continue
fi

b. Load template

uv run gw-agency info "$agency"

Read the corresponding section template from the printed template directory.

c. Draft section (Claude)

Write the section respecting:

• Word limit from agency manifest
• Template structure and required sub-sections
• Citation style from agency manifest
• Language setting (language: ro → Romanian headers with English scientific content)

Grant writing is coherence-sensitive — Claude does the writing directly. Do NOT delegate prose generation to a multi-model ensemble (that produces inconsistent voice). Use mm-ask ONLY for verification, not generation.

Reference figures using Markdown:

![Description](figures/filename.png)
*Figure N: Caption text explaining the figure content.*

Ensure every referenced figure exists in "$PROP/sections/figures/".

Save the draft checkpoint atomically:

cat "$PROP/sections/$section.md" | uv run gw-state save-checkpoint "$PROP" proposal_writing "$section" draft

d. Multi-draft variants (optional)

If --multi-draft is passed OR multi_draft.enabled: true in config:

uv run gw-multi-draft \
  --command "claude --print 'Write section $section variant {variant_id} at {output}'" \
  --variants v1 v2 v3 \
  --output-dir "$PROP/state/checkpoints/proposal_writing/$section/drafts" \
  --log-dir "$PROP/logs" \
  --target-words "$WORD_LIMIT" \
  --select-best \
  --output-file "$PROP/logs/${section}_multi_draft.json"

Then copy the best variant into sections/$section.md AND re-save the draft checkpoint so a subsequent resume picks up the winning variant, not the pre-multi-draft placeholder:

BEST=$(python3 -c "import json; print(json.load(open('$PROP/logs/${section}_multi_draft.json'))['best']['output_path'])")
cp "$BEST" "$PROP/sections/$section.md"
cat "$PROP/sections/$section.md" | uv run gw-state save-checkpoint "$PROP" proposal_writing "$section" draft

e. Fact-check the section (mm-ask preferred)

If MM_AVAILABLE=1 and multi_model.claim_verification: true:

Launch mm-ask to verify claims in parallel across 3 external models:

cat > /tmp/gw_section_verify.txt <<EOF
Verify the factual claims and citations in this grant section. For each
problem you find, output ONE JSON object per line (JSONL):

  {"claim": "...", "issue": "CITATION_HALLUCINATED|CLAIM_UNSUPPORTED|FACT_OUTDATED|...", "severity": "critical|warning|info", "fix": "..."}

Check: (1) every citation is real, (2) citations actually support the claim
next to them, (3) quantitative claims are accurate, (4) no internal
contradictions with other sections.

Return nothing if the section is clean. Do NOT wrap in markdown — just JSONL.

--- SECTION: $section ---
$(cat "$PROP/sections/$section.md")
EOF

uv run mm-ask \
  --models gpt-5.4-high,gemini-3.1-pro,grok-4-20-thinking \
  --prompt-file /tmp/gw_section_verify.txt \
  --output "/tmp/gw_section_${section}_verify.json" \
  --timeout 300

Merge the 3 reviewers' JSONL outputs: an issue flagged by ≥2 reviewers is high-confidence; single-reviewer issues are still worth checking but flagged as lower-confidence. Save the merged result as the section's verification