sws:revise-paper

/sws:revise-paper — Full revision pipeline (sequential orchestrator)

Runs the four pipeline passes in strict order: consistency check → reviser → humanizer → style-enforcer. Each pass reads the previous pass's output from disk. This is sequential by design (D4): passes are not independent, so the cycle-#7 parallel fan-out pattern does not apply here.

Flags

• --skip-humanize — skip the humanizer pass; style-enforcer reads -revised.md files directly.
• --skip-lint — skip the AI-tells linter that runs inside the humanizer agent. The linter is informational (R6): it does not gate the pipeline, but disabling it skips the cleanup nudge entirely.
• --skip-style — skip style-enforcer; pipeline ends after humanizer. No .docx is produced. Use when you only want polished markdown.

Steps

1. Resolve marker and overlay

Check $PAPER_ROOT/.sws-project.local.md. If absent, print "not an SWS project (no .sws-project.local.md found)" and exit.

Run the overlay resolver:

"${CLAUDE_PLUGIN_ROOT}/scripts/sws_python.sh" "$PAPER_ROOT" \
    "${CLAUDE_PLUGIN_ROOT}/scripts/resolve_overlay.py" \
    --paper "$PAPER_ROOT"

If profile_set: false, print "no profile set — run /sws:set-profile first" and exit. Read RESOLVED_WORD_TOTAL from the resolved overlay for the tier decision in step 4.

2. Locate draft files

Glob ${PAPER_ROOT}/_drafts/<section>.md (plain drafts, not -revised or -humanized suffixed). If none found, print:

no draft files found in _drafts/ — run /sws:draft-paper or /sws:draft-section first

and exit. List the found files so the user knows what will be revised.

3. Dispatch consistency-checker

Dispatch the consistency-checker agent via the Task tool. Pass PAPER_ROOT and CLAUDE_PLUGIN_ROOT in env. The agent:

• reads all _drafts/<section>.md files in profile section order
• runs scripts/sws_consistency_check.py
• writes ${PAPER_ROOT}/_review/consistency-report.md

If block-severity findings exist: print a summary of the findings (count by category, first three excerpts) and ask the user:

Consistency check found <N> block-severity issue(s). Continue with revision? [y/n]

Do not auto-proceed. Wait for explicit user confirmation before continuing. Warn-severity findings are logged to the report but do not pause the pipeline.

Note: if profile == funding-proposal, consistency-checker prints "profile not supported for consistency check in v0.1; manual review required" and exits 0 (D19). The pipeline continues.

4. Decide reviser tier and dispatch

Read RESOLVED_WORD_TOTAL from the overlay:

• RESOLVED_WORD_TOTAL < 1500 → dispatch reviser-fast (per-section passes; writes _drafts/<section>-revised.md + _review/revision-notes-<section>.md per section).
• RESOLVED_WORD_TOTAL >= 1500 → dispatch reviser-full (whole-paper pass; writes _drafts/<section>-revised.md for all sections + _review/revision-notes-full.md).

The tier decision is made here in the skill, not inside the agent prompts (D4 short_profile_carve_out). Profiles that naturally fall below 1500 words (editorial, communication, mini-review, commentary-reply) get reviser-fast automatically. Dispatch via the Task tool; wait for completion before proceeding.

5. Dispatch humanizer

Unless --skip-humanize is set: dispatch humanizer agent on the revised drafts (_drafts/<section>-revised.md). The humanizer:

• reads each -revised.md file
• runs the AI-tells linter (scripts/sws_lint_ai_tells.py) internally (unless --skip-lint was passed to this skill, in which case --skip-lint is forwarded to the humanizer)
• rewrites flagged tells via sentence-level polish
• writes _drafts/<section>-humanized.md

Important: humanizer skips a section if <section>-humanized.md already exists and is newer than <section>-revised.md. This preserves any hand-edits the user made to a previous humanized draft (R4).

The linter is informational only inside this pipeline (R6). It does not gate the humanizer or the skill — the humanizer resolves flagged tells via rewrite and moves on. To view raw linter output without running the full pipeline, use /sws:lint-ai-tells <file>.

6. Dispatch style-enforcer

Unless --skip-style is set: dispatch style-enforcer agent. It reads humanized drafts (or revised drafts if --skip-humanize) in profile section order and:

1. calls scripts/sws_write_docx.py to generate ${PAPER_ROOT}/Manuscript/<paper-name>.docx with SWS style canon
2. calls scripts/sws_apply_chemistry_format.py on the result (skipped automatically if format: latex)

Backup discipline: the cycle-#5 PreToolUse hook fires automatically before style-enforcer's docx write, producing <paper-name>.backup_pre_style_enforcer_<timestamp>.docx. No additional backup logic is needed here (D16).

7. Final summary

Print a structured summary:

/sws:revise-paper complete
  consistency:  <N> block, <M> warn — see _review/consistency-report.md
  reviser:      <tier> — <words before> → <words after> words
  humanizer:    <N> sections polished, <K> AI-tells resolved  [skipped if --skip-humanize]
  style:        SWS canon applied, chemistry pass done        [skipped if --skip-style]
  output:       Manuscript/<paper-name>.docx

If --skip-style, omit the docx line and note that no .docx was written.

When to invoke

• Explicit /sws:revise-paper.
• User triggers from the description above.
• Do NOT invoke without a marker, without a set profile, or without at least one _drafts/<section>.md file.