agent-afk 3.8.5 → 3.8.9

package/dist/index.mjs

@@ -1168,6 +1168,9 @@ Apply in order \u2014 first match wins:
 3. Any rejection check = Yes \u2192 **REJECT** (SALVAGE only if the sole Yes is "better folded into another skill")
 4. Any single Stage 1 rubric dimension \u22642 \u2192 downgrade one tier from the rubric-total result (APPROVE\u2192SALVAGE, SALVAGE\u2192REJECT)
 5. Workflow Impact \u22643 \u2192 downgrade one tier (session shape must materially change for this plugin)
+
+   _Why rule 5 fires often:_ Skills in the gate/guard/check category \u2014 those that add an advisory step rather than reshape session flow \u2014 systematically score WI \u22643 by design. Examples that have hit this rule: \`thesis-lock\`, \`checkpoint\`, \`read-before-explain\`, \`premise-check\`, \`drift-watch\`. This is intentional: a one-line check that mutates Claude's behavior on a specific signal is rarely transformative enough to clear WI \u22654. If the candidate is borderline, prefer folding it into an incumbent skill (or accepting the SALVAGE downgrade) over forcing an APPROVE.
+
 6. Stage 2 failure-resilience total \u22648 (Fragile band) \u2192 downgrade one tier (a fragile force multiplier is still dangerous). Note which dimensions scored 0\u20131 in the Rewrite target so the skill can be hardened.
    - Rules 4, 5, and 6 do not stack \u2014 apply at most one downgrade per candidate. If multiple fire, prefer rule 6 (failure behavior is the most costly to miss).
 7. Stage 1 total \u226524/30 \u2192 **APPROVE**
@@ -1515,7 +1518,7 @@ Create a ship-ready summary with next steps.`,l=await a.runToResult(c);if(l.stat
 - Use absolute paths for file operations.
 - Prefer \`agent\` (and \`skill\`) for multi-file investigation, verification, parallel hypotheses, and any work that would otherwise consume large amounts of inline context. The main session is the coordinator; subagents are the investigators.
 
-When you see a \`<command-name>\` tag in the current conversation turn, the skill has ALREADY been loaded by the user typing a slash command. Do NOT re-invoke the skill tool to dispatch the same skill again. Instead, treat the \`<command-message>\` as the skill name and \`<command-args>\` as its arguments, then follow the instructions in the body block immediately following the tag.`,jn=`# Cross-Session Memory
+When you see a \`<command-name>\` tag in the current conversation turn, the skill has ALREADY been loaded by the user typing a slash command. Do NOT re-invoke the skill tool to dispatch that same skill again. Instead, treat the \`<command-message>\` as the skill name and \`<command-args>\` as its arguments, then follow the instructions in the body block immediately following the tag. You MAY still invoke the skill tool to dispatch OTHER skills that are not the one already loaded.`,jn=`# Cross-Session Memory
 
 You have three tools for persisting knowledge across sessions: memory_search, memory_update, and procedure_write.
 

package/dist/telegram.mjs

@@ -1325,6 +1325,9 @@ Apply in order \u2014 first match wins:
 3. Any rejection check = Yes \u2192 **REJECT** (SALVAGE only if the sole Yes is "better folded into another skill")
 4. Any single Stage 1 rubric dimension \u22642 \u2192 downgrade one tier from the rubric-total result (APPROVE\u2192SALVAGE, SALVAGE\u2192REJECT)
 5. Workflow Impact \u22643 \u2192 downgrade one tier (session shape must materially change for this plugin)
+
+   _Why rule 5 fires often:_ Skills in the gate/guard/check category \u2014 those that add an advisory step rather than reshape session flow \u2014 systematically score WI \u22643 by design. Examples that have hit this rule: \`thesis-lock\`, \`checkpoint\`, \`read-before-explain\`, \`premise-check\`, \`drift-watch\`. This is intentional: a one-line check that mutates Claude's behavior on a specific signal is rarely transformative enough to clear WI \u22654. If the candidate is borderline, prefer folding it into an incumbent skill (or accepting the SALVAGE downgrade) over forcing an APPROVE.
+
 6. Stage 2 failure-resilience total \u22648 (Fragile band) \u2192 downgrade one tier (a fragile force multiplier is still dangerous). Note which dimensions scored 0\u20131 in the Rewrite target so the skill can be hardened.
    - Rules 4, 5, and 6 do not stack \u2014 apply at most one downgrade per candidate. If multiple fire, prefer rule 6 (failure behavior is the most costly to miss).
 7. Stage 1 total \u226524/30 \u2192 **APPROVE**
@@ -1561,7 +1564,7 @@ Create a ship-ready summary with next steps.`,l=await a.runToResult(c);if(l.stat
 - Use absolute paths for file operations.
 - Prefer \`agent\` (and \`skill\`) for multi-file investigation, verification, parallel hypotheses, and any work that would otherwise consume large amounts of inline context. The main session is the coordinator; subagents are the investigators.
 
-When you see a \`<command-name>\` tag in the current conversation turn, the skill has ALREADY been loaded by the user typing a slash command. Do NOT re-invoke the skill tool to dispatch the same skill again. Instead, treat the \`<command-message>\` as the skill name and \`<command-args>\` as its arguments, then follow the instructions in the body block immediately following the tag.`,br=`# Cross-Session Memory
+When you see a \`<command-name>\` tag in the current conversation turn, the skill has ALREADY been loaded by the user typing a slash command. Do NOT re-invoke the skill tool to dispatch that same skill again. Instead, treat the \`<command-message>\` as the skill name and \`<command-args>\` as its arguments, then follow the instructions in the body block immediately following the tag. You MAY still invoke the skill tool to dispatch OTHER skills that are not the one already loaded.`,br=`# Cross-Session Memory
 
 You have three tools for persisting knowledge across sessions: memory_search, memory_update, and procedure_write.
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-afk",
-  "version": "3.8.5",
+  "version": "3.8.9",
   "description": "CLI tool for interacting with AI agents via multiple interfaces",
   "main": "dist/index.mjs",
   "type": "module",