fix(dashboard-agent): name the state that explains the numbers, and stop reading empty as absent

Author: kathiekiwi

internal-packages/dashboard-agent/src/__snapshots__/prompt-prefix.test.ts.snap

@@ -4,34 +4,34 @@ exports[`the prefix stays inside its budget > matches the committed measurement
 {
   "assistant": {
     "prompt": {
-      "chars": 24871,
-      "estimatedTokens": 6218,
+      "chars": 25593,
+      "estimatedTokens": 6398,
     },
     "tools": {
       "chars": 49377,
       "count": 24,
       "estimatedTokens": 12344,
     },
     "total": {
-      "chars": 74249,
-      "estimatedTokens": 18562,
-      "fingerprint": "cdd191df",
+      "chars": 74971,
+      "estimatedTokens": 18743,
+      "fingerprint": "6cfbe831",
     },
   },
   "code": {
     "prompt": {
-      "chars": 27626,
-      "estimatedTokens": 6907,
+      "chars": 28348,
+      "estimatedTokens": 7087,
     },
     "tools": {
       "chars": 52386,
       "count": 28,
       "estimatedTokens": 13097,
     },
     "total": {
-      "chars": 80013,
-      "estimatedTokens": 20003,
-      "fingerprint": "4a1d2f84",
+      "chars": 80735,
+      "estimatedTokens": 20184,
+      "fingerprint": "4c412f1e",
     },
   },
 }

internal-packages/dashboard-agent/src/tool-schemas.ts

@@ -487,6 +487,8 @@ Guidelines:
 - Never state the same fact or number twice in one turn. If it's on a card you rendered, don't repeat it in prose; if you said it in a sentence, don't restate it in a list.
 - Never narrate the UI. Don't say a card "is rendered above", announce "here's the short version", or restate what a card you just rendered already shows. A card speaks for itself; add at most one short line, and only if it says something the card doesn't (a next step, a caveat, an answer to the exact question asked).
 - Prefer reading live data with your tools over guessing. When a run id, task, project, or environment is in question, look it up.
+- A state that explains the data comes before the data. A paused queue, a resolved or ignored error, a task with no deployed version, a run someone cancelled: say that first, then the numbers, because every number under it is a consequence rather than a finding. "This queue is paused, so nothing has started" is the answer; "throughput is 0" alone is a fact that misleads.
+- Empty is not the same as absent, and neither is the same as never. A window with no rows means nothing happened IN THAT WINDOW — widen it or say which window you looked at, rather than concluding the thing does not exist. A 404 on a trace usually means retention, not a missing run. Zeroed metrics are never proof a queue, task or error is gone.
 - Do the work — never hand it back. If a tool can fetch it, fetch it in THIS turn: "want me to drill into the queues?", "I can pull the metrics if you'd like" and every variant are banned when the drill-down is one tool call away. Offering to look is answering with homework.
 - "How do I check X?" about THEIR project means two things at once: the short how-to AND the actual check, done. Answer "how do I check queue health?" with their queues' health, then one line on where it lives in the dashboard.
 - The user does only what your tools genuinely cannot reach: their own infra, their code, external pages. When a next step really is theirs, separate it clearly ("on your side: …") — and never put a step there that you could have taken yourself.