@infinitedusky/indusk-mcp 1.24.1 → 1.24.3

package/dist/lib/eval/evaluator-runner.js

@@ -125,6 +125,7 @@ export function runEvaluatorBackground(opts) {
                 scorecardText = extracted;
             }
             const scorecard = JSON.parse(scorecardText.trim());
+            scorecard.timestamp = new Date().toISOString();
             if (usage)
                 scorecard.usage = usage;
             scorecard.telemetryPosted = false;
@@ -245,6 +246,7 @@ async function runEvaluatorSyncInner(opts, projectGroup) {
                     scorecardText = extracted;
                 }
                 const scorecard = JSON.parse(scorecardText.trim());
+                scorecard.timestamp = new Date().toISOString();
                 if (syncUsage)
                     scorecard.usage = syncUsage;
                 scorecard.telemetryPosted = false;

package/dist/lib/eval/persistent-evaluator.js

@@ -242,6 +242,11 @@ Output ONLY the JSON scorecard as before — no commentary.`;
                 return out;
             });
             const scorecard = JSON.parse(parsed.scorecardText.trim());
+            // Override the model-supplied timestamp with actual completion time.
+            // The model doesn't know the real current time and tends to round to
+            // 5-minute marks (e.g. 18:25:00). Use Date.now() so timestamps are
+            // accurate to the second.
+            scorecard.timestamp = new Date().toISOString();
             if (parsed.usage)
                 scorecard.usage = parsed.usage;
             scorecard.telemetryPosted = false;

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@infinitedusky/indusk-mcp",
-	"version": "1.24.1",
+	"version": "1.24.3",
 	"description": "InDusk development system — skills, MCP tools, and CLI for structured AI-assisted development",
 	"type": "module",
 	"files": [

package/skills/work.md

@@ -276,12 +276,17 @@ Skip silently if `mcp__indusk__highlight` is unavailable — highlights are best
 
 ## Commits (jj)
 
-Use the **describe-then-do** workflow from the jj skill:
+Use the **describe-then-do** workflow from the jj skill. The order is load-bearing — never reverse it.
 
-1. `jj new` before each logical unit of work
-2. `jj describe` to declare what you're about to do
-3. Do the work, check off the item
-4. Repeat
+**For every checklist item, in this exact order:**
+
+1. **`jj new`** — start a fresh empty commit. Do this BEFORE you touch any file.
+2. **`jj describe -m "..."`** — write the commit message describing what you're ABOUT to do. The description names the intent in present/future tense ("rename foo to bar") not past tense ("renamed foo to bar").
+3. **Do the work** — edit files, run tools, accumulate changes in `@` (the current commit).
+4. **Check the item off** in the impl.md.
+5. **Repeat** for the next item — back to step 1.
+
+**❌ Anti-pattern: describe-after-do.** Doing the work first and then describing what you did breaks two things: (a) the eval agent fires on `jj describe`, so the description is what it scores against — late descriptions mean the agent scores work that's already done without intent context; (b) the working copy IS the current commit in jj, so any uncommitted state automatically gets attributed to whatever the commit description currently says, even if that's the previous unit's description. Always: new → describe → work, in that order.
 
 **Default: one commit per checklist item.** Each impl checklist item is a logical unit of work — give it its own commit. This keeps history granular, makes blame and bisect useful, avoids the end-of-phase `jj split` chore, and lets the eval agent score each unit while context is fresh.