@victortomaili/skill-cli 0.6.5 → 0.6.7

package/CHANGELOG.md

@@ -11,6 +11,26 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Cursor adapter (`.cursor/rules` format) for `init -g` bootstrap.
 - Per-agent hook adapters for automatic `/X` triggering (push model).
 
+## [0.6.7] - 2026-07-08
+
+### Changed
+- AGENTS.md START GATE: heavily compressed (605 → ~250 words) and made the
+  classification mechanical. cost/style/speed-affecting skills MUST be proposed on
+  EVERY task (those axes apply by definition); the rationalizations "not needed for
+  this task", "the user didn't request it", and "skip unless needed" are now
+  explicitly FORBIDDEN. Also de-hardcoded the caveman example (it overfit the test)
+  — the rule is now fully generic.
+
+## [0.6.6] - 2026-07-08
+
+### Changed
+- AGENTS.md START GATE: the load/propose/skip decision is now driven by the four
+  criteria (correctness / quality / cost / speed), not "propose every ★". A skill
+  that affects CORRECTNESS or QUALITY → auto-load; affects COST or SPEED only →
+  propose; affects none → skip. Task type still does not filter candidates (a
+  cost/style skill is a candidate on every task), and the agent must state each
+  decision out loud. `skill active` footer rewritten to match.
+
 ## [0.6.5] - 2026-07-08
 
 ### Changed

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@victortomaili/skill-cli",
-  "version": "0.6.5",
+  "version": "0.6.7",
   "description": "Cross-agent skill manager — one global store, clean agent folders, activation via skill.config",
   "license": "MIT",
   "author": "Victor Tomaili <victor@tomaili.com>",

package/src/commands/defaults.js

@@ -40,8 +40,7 @@ export function cmdActive() {
     if (s.description) console.log(c.gray('      ' + String(s.description).replace(/[\r\n]+/g, ' ').trim()))
   }
   console.log()
-  console.log(c.bold(c.yellow('→ For EACH ★ skill above: PROPOSE it now.')) + c.gray(' Task type is irrelevant — \"not needed for this task\" is NOT a reason to skip. The user decides, not you.'))
-  console.log(c.gray('  Load a skill: ') + c.cyan('skill cat <name>'))
+  console.log(c.bold('→ Decide each skill by the criteria:') + c.gray(' affects ') + c.cyan('CORRECTNESS/QUALITY') + c.gray(' → load (') + c.cyan('skill cat <name>') + c.gray('); affects ') + c.cyan('COST/SPEED') + c.gray(' → propose; affects none → skip. Task type is irrelevant; when unsure → propose.'))
 }
 
 // `skill default <name>` — mark a skill as a default (active by default in every

package/src/lib/agents-md.js

@@ -24,50 +24,26 @@ Usage:
 - \`skill default <name>\` — mark a skill as a global default (active + auto-load)
 
 START GATE (mandatory): On the FIRST user message of a session, your VERY FIRST
-action — BEFORE ANYTHING ELSE: before you start thinking/reasoning about the task
-and before any tool call — MUST be to run \`skill active\`. It prints the skill
-CATALOG — every ACTIVE skill's name + FULL description (NOT the skill body).
-Read it, then for EACH skill decide FROM ITS DESCRIPTION:
-- AUTO-LOAD it (\`skill cat <name>\`, silently) when it is needed to do the task AT
-  ALL, OR when it would make the work HIGHER QUALITY / more correct — domain
-  knowledge, best-practice, standards, a quality methodology. These are not
-  optional; they are how the job should be done.
-- PROPOSE it (tell the user what it does; apply ONLY after they confirm; \`/X\` =
-  confirm) when it is an OPTIONAL change to the EXPERIENCE, not to quality:
-  • response STYLE / tone / format / a persistent mode (e.g. a brevity style);
-  • CONTEXT-COST / token optimization;
-  • an optional ACCELERATOR / shortcut (easier or faster, but not about correctness).
-ENFORCEMENT — for EVERY active skill you MUST state a decision OUT LOUD: LOAD it
-or PROPOSE it to the user. Staying silent, or "not needed for this task", is NOT
-a valid decision. In particular:
-- A ★ default is NEVER silently skipped. STYLE / COST / shortcut skills are
-  candidates for EVERY task type — review, debugging, refactor, anything; the task
-  type does NOT disqualify them, and "the user didn't ask" is NOT a reason to stay
-  quiet. If you are not loading a ★ default, you MUST propose it. The ONLY valid
-  reason to stay silent about a ★ is: the user already declined it this session.
-- Do NOT wait for the user to ask — proposing happens HERE, proactively, every turn
-  until the user decides.
-When unsure whether to load or propose → PROPOSE (proposing is the safe default;
-loading is silent).
-PRIORITIES — when trading off how to do the work, rank: CORRECTNESS first, then
-QUALITY, then COST (cheap / token-efficient), then SPEED (fast). Never sacrifice
-correctness or quality for speed or cost.
-LOADED ≠ LISTED. A skill is loaded ONLY if you have run \`skill cat <name>\` for
-it THIS session. Seeing a skill in \`skill active\` / \`skill list\` does NOT load
-it — you then know only its name + description, nothing more. Never infer a skill
-is in your context from its name, the \u2605 marker, or an \`active\` status: if you
-haven't catted it, treat it as NOT loaded (then decide: cat it, or propose it).
-Do not assume a skill is already applied — \`active\`/\u2605-default = AVAILABLE, not
-APPLIED. ANY skill can be context-altering; judge from the description — there is
-no fixed list and no flag. Do this even if the task looks trivial: skills shape
-how you approach it, so you cannot reason about it correctly until you've reviewed
-them.
-
-Discovery (every message): On EVERY later user message, re-check the catalog
-against the new request; load any newly-relevant functional skill
-(\`skill cat <name>\`), and PROPOSE any newly-relevant context-altering one.
-Load each skill only ONCE per session (\`skill trigger <keyword>\` resolves a
-keyword to a skill).
+action — BEFORE ANYTHING ELSE (before thinking, before any tool call) — MUST be to
+run \`skill active\`. It lists each ACTIVE skill's name + FULL description (never the
+body). Read it, then for EACH skill classify which axis it moves and ACT — there is
+no discretion and no third option:
+- moves CORRECTNESS or QUALITY → load it now: \`skill cat <name>\`.
+- moves COST, SPEED, or response STYLE → PROPOSE it: in THIS reply, ask the user
+  "Enable <name>? It <one-line benefit>." and apply only on yes. These three axes
+  apply to EVERY task by definition (every reply costs tokens and time), so the
+  thoughts "not needed for this task", "the user didn't request it", and "skip
+  unless needed" are FORBIDDEN — they are the exact error. Listing the skill in a
+  table is NOT proposing; you must actually ask.
+- moves none of the above → skip.
+When unsure if a skill moves an axis → PROPOSE (ask). LOADED \u2260 LISTED: a skill is
+loaded only if you \`cat\`-ed it this session — listing it, its \u2605, or its \`active\`
+status is NOT loading; active/\u2605 means AVAILABLE, not APPLIED.
+PRIORITIES: correctness > quality > cost (cheap) > speed — never trade correctness
+or quality for speed or cost.
+Discovery: on EVERY later message, re-run this classification; load newly-relevant
+correctness/quality skills, PROPOSE newly-relevant cost/style/speed ones. Load each
+skill only ONCE per session (\`skill trigger <keyword>\` resolves a keyword).
 
 Triggers: when the user types \`/X\`, run \`skill trigger X\`.
 - Single match → apply the output directly.