v0.25.0

Changed

• read_pad is now a fixed-window tool by contract. Caps at ~2,000 words per call. Docs at or below the cap return in full as before. Above the cap, the response includes the doc opening (title + intro + first sections — the most context-rich slice) plus lastNodeId and a continuation hint pointing at peek_doc({ around: lastNodeId, after: N }), outline_doc({ docId }), or search_docs({ query, docId }). No force flag — the cap is the architecture, not a soft suggestion. First truncation in a session prefixes an FYI so the agent learns the new behavior; subsequent truncations skip the explanation.

  Removes the failure mode where an agent could token-blow a 50k-word doc by calling read_pad blindly. Side-effect: surfaces monolith docs (8k+ words in one file) as the wrong unit for AI-assisted writing and nudges users toward chapter/section/topic-sized docs that read complete in one shot.

• Skill v0.15.0. Read ladder rule rewritten to reflect the truncation contract — step 6 now reads "read_pad(docId) — first ~2,000 words of the body, ALWAYS truncated above the cap" instead of the prior "escape hatch for everything." Single-document and Multi-document workflows updated to orient with outline_doc + peek_doc for long docs and use read_pad only when the doc is known to fit under the cap. New paragraph on doc-structure implications: the cap makes monoliths visibly expensive and rewards smaller docs.

Added

• truncateRead(doc, maxWords) + countWords(nodes) in peek-outline.ts. Truncates at top-level node boundaries — never splits a list, blockquote, or code block mid-way — so the returned markdown stays structurally valid. Always includes at least one top-level node so callers never receive an empty body. Returns { doc, truncated, totalWords, returnedWords, lastNodeId, remaining } for callers that need the cost-and-continuation metadata.