Large skill catalog: list one line per skill (name, shortest location, 100-char line) and compact from 16 KiB of descriptions #4449

Summary

A large skill catalog (more than 100 skills or more than 64 KiB of descriptions) is already listed in compact mode, but each compact entry is still a five-line XML block with a full rN/<name>/SKILL.md location and a 160-character description cut mid-word. On a real 101-skill catalog that block is about 8,300 tokens (o200k estimate) on every request. The compact trigger also has a cliff: hiding a few skills (for example retired ones) drops a 98-skill catalog back to full multi-sentence descriptions, which is larger, not smaller.

Reproduction

1. Load ~100 skills whose descriptions average ~300 characters (a typical personal skill library).
2. Render the system prompt (formatSkillsForPrompt).
3. Each entry is <skill>\n<name>..</name>\n<description>..</description>\n<location>rN/name/SKILL.md</location>\n</skill>; 100 of 101 locations are exactly rN/<name>/SKILL.md and carry no information beyond the root alias.
4. Remove three skills (98 left, ~29 KiB of descriptions): the catalog falls below both thresholds and every full description is listed.

Measured (real 101-skill catalog, o200k estimate)

Expected (ideal state)

• A large catalog lists one line per skill: - name [location]: one-line description, inside the same <available_skills> wrapper and after the same <skill_roots> table.
• location is the root alias alone (r0) when the file is <root>/<name>/SKILL.md, else the path under the root (r1/bundle/SKILL-x.md), so every skill stays resolvable; one header line states the rule.
• The one-line description is at most 100 characters, cut at a word boundary.
• The compact trigger uses a 16 KiB description budget (count threshold unchanged), so a ~100-skill library stays compact when a few skills are hidden, while a default install with ~30 bundled skills keeps full descriptions.
• The names-only fallback (over the 48 KiB cap) uses the same one-line form without descriptions.
• Small catalogs are unchanged.

Actual

packages/coding-agent/src/core/skills-prompt.ts: renderSkills emits the XML block for compact entries; ONE_LINE_DESCRIPTION_MAX_CHARS = 160 cut without a word boundary; COMPACT_SKILL_CATALOG_DESCRIPTION_BYTES = 64 KiB.

Acceptance criteria

• Regression test: a 98-skill catalog with ~300-character descriptions is compact; every line is - name [location]: …, at most 140 characters; every location resolves to the skill's file through the roots table.
• Existing small-catalog and names-only bounds tests still hold (updated for the line form).
• Real first-request measurement before/after (input total and cacheRead) in the PR.

Related

• No per-request context budget or spend guard: 1M-window models (e.g. OpenGateway DeepSeek) grow to ~500K-800K tokens per call on unattended runs #3460 / fix(coding-agent): bound context files and large skill catalogs added to every request #3463 (compact catalog and prompt budgets), GPT-6 Astra over-prepares and over-verifies: skill catalog stance, forced first-turn verification phase, and preset gates codex does not carry #2505 (skill catalog stance)
• The Claude-subscription prompt extractor keys on the catalog marker and </available_skills>, both kept.