Phase 1: Token cost data correctness (FEA-1431/1432/1433)

apps/desktop/CLAUDE.md

@@ -136,6 +136,54 @@ Typical log locations:
 
 The Diagnostics tab shows the current in-memory gateway log plus a bounded previous-session tail read from `main.log` at startup. First-run or unreadable log files must not block boot; return an empty previous-session tail and continue.
 
+## Pricing Data
+
+Model pricing rules baked into the generated agent-monitor `db.js` come from
+two sources, merged at build time by `loadHostDefaultPricing()` in
+`scripts/build-agent-monitor.mjs`:
+
+1. **`scripts/litellm-pricing.json`** — auto-generated, sorted 6-tuples
+   (`pattern`, `display_name`, `input/Mtok`, `output/Mtok`,
+   `cache_read/Mtok`, `cache_write/Mtok`) derived from LiteLLM's
+   `model_prices_and_context_window.json`. Filtered to vendor prefixes
+   `claude-`, `gpt-`, `o1-`, `o3-`, `gemini-`. Committed.
+2. **`HOST_ONLY_OVERRIDES`** in `scripts/build-agent-monitor.mjs` — rows
+   LiteLLM does not carry (fallback patterns like `cursor-default`,
+   OpenCode-hosted free models), plus any *temporary* upstream-deviation
+   overrides documented inline with a TODO + ticket reference.
+
+The build runs two invariants. Both are hard build-time assertions —
+violation throws with an actionable message and aborts the build:
+
+- **OpenAI zero-write + sanity floor** — `gpt-*` / `o1-*` / `o3-*` rows
+  must have `cache_write = 0` AND `cache_write_1h = 0` (OpenAI publishes
+  no cache-write surcharge), AND `cache_read ≤ input` (cached input must
+  not cost more than uncached — a true sanity check, not a guess at the
+  discount ratio). LiteLLM is trusted for the actual cache_read rate
+  (currently 10% for the GPT-5 family, 50% for older GPT-4o-style models).
+- **Anthropic 1h cache floor** — every `claude-*` row with positive input
+  must have `cache_write_1h ≥ input × 1.5` (sanity floor for the 2×
+  documented tier).
+
+(A third invariant — an Opus 4.x input floor of $10/Mtok — was removed
+during the FEA-1431 bugfix pass: Anthropic re-priced Opus starting at 4.5
+down to $5/Mtok input, so the floor was rejecting correct data.)
+
+### Refreshing pricing
+
+```bash
+just desktop-refresh-pricing   # or: pnpm -C apps/desktop refresh:pricing
+```
+
+The wrapper fetches upstream, writes `litellm-pricing.json` +
+`litellm-pricing.meta.json` (with SHA pin + ISO timestamp), and then runs the
+build-time invariants. A regression in upstream rates (Opus floor) blocks the
+refresh — review the diff, then either add a temporary override row to
+`HOST_ONLY_OVERRIDES` or back out the refresh. Refresh roughly weekly.
+
+The JSON sits inside the `currentStamp()` hash inputs so a fresh refresh
+forces a rebuild of `agent-monitor/.generated/server/db.js`.
+
 ## Agent Monitor Sidecar
 
 The desktop app bundles the MIT-licensed `Claude-Code-Agent-Monitor`

apps/desktop/package.json

@@ -1,6 +1,6 @@
 {
   "name": "desktop",
-  "version": "0.15.93",
+  "version": "0.15.103",
   "description": "ClosedLoop Desktop",
   "author": "ClosedLoop AI <support@closedloop.ai>",
   "private": true,
@@ -14,6 +14,7 @@
     "prebuild": "node -e \"const{execSync:e}=require('child_process'),{writeFileSync:w}=require('fs');const h=e('git rev-parse HEAD').toString().trim();w('src/shared/build-info.ts','// AUTO-GENERATED — do not edit\\nexport const BUILD_COMMIT_HASH = \\\"'+h+'\\\";\\n');\"",
     "build": "pnpm clean:dist && pnpm prebuild && tsc -p tsconfig.json && pnpm build:agent-monitor",
     "build:agent-monitor": "node scripts/build-agent-monitor.mjs",
+    "refresh:pricing": "node scripts/refresh-litellm-pricing.mjs",
     "dashboard:reset": "node scripts/reset-dashboard-db.mjs",
     "dashboard:reset-packs": "node scripts/reset-dashboard-db.mjs --packs-only",
     "stage:package": "node scripts/stage-packaging-app.mjs",

apps/desktop/scripts/agent-monitor-client/Sessions.tsx

@@ -373,8 +373,64 @@ export function Sessions() {
                     <td className="px-5 py-4 text-sm text-gray-400">
                       {session.agent_count ?? "-"}
                     </td>
+                    {/*
+                      FEA-1433: render three cost states (matching the
+                      sidecar's calculateSessionCostFea1433 contract).
+                        cost == null → EVERY model in the session is
+                          unpriced (no model_pricing rule matched any
+                          row). Render an em-dash with a tooltip pointing
+                          at Settings → Pricing. This is the strong
+                          diagnostic signal — never silently show $0.
+                        cost > 0 && unpriced_models non-empty → MIXED:
+                          some models priced, at least one not. Render
+                          the partial $ value with an asterisk + tooltip
+                          so the user sees the partial total AND knows
+                          it understates true cost (FEA-1433 review fix).
+                        cost > 0 && unpriced_models empty → fully priced.
+                          Render the dollar value.
+                        cost == 0 → no usage on a priced model. Render
+                          "-" (distinct from the unpriced em-dash).
+                      Sort column treats null as end-sort in the sidecar
+                      (see calculateSessionCostFea1433 patch in
+                      build-agent-monitor.mjs).
+                    */}
                     <td className="px-5 py-4 text-sm text-gray-400 font-mono">
-                      {session.cost != null && session.cost > 0 ? fmtCost(session.cost) : "-"}
+                      {session.cost == null ? (
+                        <span
+                          className="cursor-help text-gray-500"
+                          title={(() => {
+                            const first = session.unpriced_models?.[0];
+                            return first
+                              ? `Model "${first}" not in pricing table — open Settings → Pricing to add a manual rate.`
+                              : "Model not in pricing table — open Settings → Pricing to add a manual rate.";
+                          })()}
+                        >
+                          —
+                        </span>
+                      ) : session.cost > 0 ? (
+                        (() => {
+                          const unpricedCount =
+                            session.unpriced_models?.length ?? 0;
+                          if (unpricedCount === 0) return fmtCost(session.cost);
+                          // Mixed: partial cost. Asterisk + tooltip.
+                          const sample = session.unpriced_models?.[0];
+                          const label =
+                            unpricedCount === 1
+                              ? `Partial — model "${sample}" is not in the pricing table; this total excludes its tokens. Open Settings → Pricing to add a rate.`
+                              : `Partial — ${unpricedCount} models (including "${sample}") are not in the pricing table; this total excludes their tokens. Open Settings → Pricing to add rates.`;
+                          return (
+                            <span
+                              className="cursor-help"
+                              title={label}
+                            >
+                              {fmtCost(session.cost)}
+                              <span className="text-amber-400 ml-0.5">*</span>
+                            </span>
+                          );
+                        })()
+                      ) : (
+                        "-"
+                      )}
                     </td>
                     <td
                       className="px-5 py-4 text-[11px] text-gray-500 font-mono"