feat(billing): drop trace table, make 'analyse my spend' clickable

[pauldambra]

Problem

Two issues with the PostHog Code token spend banner:

1. The Top traces table is noise. Trace IDs are opaque strings (UUIDs or JSON-shaped device blobs like {"device":"...","session":"..."}), the list is long, and there's no actionable next step from seeing it. Users get more information than they can use.
2. The "drop the exploring-llm-costs skill into your agent" line is a dead link. It links to the SKILL.md on GitHub, but that doesn't actually move the user toward the analysis they want. They have to manually start a new task, paste the cost data in, ask for advice, and remember to load the skill — friction nobody overcomes for a "maybe save some money later" task.

Changes

1. Drop the TraceTable component and the trace-related branch in generateSuggestions. Backend PR #59796 deprecates the top_traces API field (returns empty); removing the rendering here means we don't show an empty section.
2. Replace the static exploring-llm-costs GitHub link with a Sparkle-iconed button: "Ask an agent to analyse this and suggest reductions". Clicking it calls navigateToTaskInput({ initialPrompt }) with a markdown report of the user's spend (summary, by_product, by_tool top 10, by_model) plus a prompt instructing the new agent to load the skill from the PostHog skill store via mcp__posthog__exec → llma-skill-get and rank reduction advice by impact.

The prefill saves the new task an API round-trip and gives the agent the full breakdown in its initial context, so it can answer "what should I do to spend less" immediately without fetching data first.

How did you test this?

I'm an agent — Claude Code via PostHog Code.

• pnpm exec biome check apps/code/src/renderer/features/billing/components/TokenSpendAnalysisBanner.tsx — clean.
• pnpm -F @posthog/code exec tsc --noEmit — only the existing unrelated auth/service.test.ts errors, none in the modified file.
• No manual UI run — I haven't booted the app. Worth a human eyeball on (a) the button styling/placement, (b) the prefilled prompt arriving correctly in the new task input, (c) the new task actually loading the skill on its first turn.

Publish to changelog?

no

---

Created with PostHog Code

[greptile-apps]

Comments Outside Diff (1)

1. apps/code/src/renderer/features/billing/components/TokenSpendAnalysisBanner.tsx, line 422-425 (link)

   The initial callout (shown before data loads) still advertises "by trace" — but the trace table was removed in this PR, so the description will be stale. Consider dropping "by trace" from the copy to match the current feature set.

   Prompt To Fix With AI

   This is a comment left during a code review.
   Path: apps/code/src/renderer/features/billing/components/TokenSpendAnalysisBanner.tsx
   Line: 422-425
   
   Comment:
   The initial callout (shown before data loads) still advertises "by trace" — but the trace table was removed in this PR, so the description will be stale. Consider dropping "by trace" from the copy to match the current feature set.
   
   How can I resolve this? If you propose a fix, please make it concise.

Prompt To Fix All With AI

Fix the following 2 code review issues. Work through them one at a time, proposing concise fixes.

---

### Issue 1 of 2
apps/code/src/renderer/features/billing/components/TokenSpendAnalysisBanner.tsx:422-425
The initial callout (shown before data loads) still advertises "by trace" — but the trace table was removed in this PR, so the description will be stale. Consider dropping "by trace" from the copy to match the current feature set.

### Issue 2 of 2
apps/code/src/renderer/features/billing/components/TokenSpendAnalysisBanner.tsx:242-261
Markdown table injection from unescaped data values — if a `product`, `tool`, or `model` name contains a pipe character `|`, the generated markdown table will have misaligned cells, causing the agent to misread row boundaries. For example a tool name like `bash | grep` would silently split into extra columns. Consider replacing `|` with `\|` (or a similar escape) in all cell values before interpolation.

_{Reviews (1): Last reviewed commit: "feat(billing): drop trace table, make 'a..." | Re-trigger Greptile}

[pauldambra]

Note

🤖 Automated comment by QA Swarm — not written by a human

Single-reviewer pass: paul-reviewer (focused review chosen for this small frontend-only change).

Verdict: 💬 APPROVE WITH NITS → 1 fix pushed, rest deferred

The diff is doing the right things — drop the unactionable trace table, swap the static GitHub link for a "send the spend data into a fresh agent task" affordance that mirrors the inbox `ReportDetailPane` pattern. No blockers.

Fixed in `a457d50e`

• 🟢 (NIT) Button copy — "Ask an agent to analyse this and suggest reductions" read like it kicked off something inline. Changed to "Open a task to analyse this with an agent" to make the destination explicit.

Deferred (non-blocking)

• 🟡 (MEDIUM) MCP soft dependency — the prompt asks the new agent to load `exploring-llm-costs` via `mcp__posthog__exec` → `llma-skill-get`. If the receiving agent doesn't have the posthog MCP wired up (or uses a different server name), the instruction silently no-ops. Mitigation: the prompt already includes the full data in the initial context, so even without the skill the agent has the numbers and can give generic advice. The skill is the playbook; the data is the substance.
• 🟢 (LOW) Missing analytics event — no `track(ANALYTICS_EVENTS.X)` call in the button handler. This repo's analytics layer would need a new event added to `ANALYTICS_EVENTS` + `EventPropertyMap`, which is a bigger change than the finding warrants. Worth a follow-up so we can measure: did anyone click the button → did it open a task → did the task produce a useful answer?
• 🟢 (LOW) Format-string duplication — `buildAnalysisPrompt` re-derives the same `formatUsd` / `formatTokens` / "(no tool)" / "(unknown)" fallback strings the React tables use. If someone tweaks the tables, the prompt drifts silently. Acceptable for now; consider extracting shared format helpers if a second prompt-builder appears.
• ⚪ (NIT) Skill name string coupling — `SKILL_NAME = "exploring-llm-costs"` is a magic-string contract with the PostHog skill store. If the skill gets renamed, this silently asks for one that doesn't exist. Future fragility note, not a fix today.

Reviewer summary

Reviewer	Assessment
👤 paul	Ship as you see fit. Diff is the right shape; the main concerns are about resilience of the MCP-skill-load handshake and a missing analytics event so we can measure adoption.

---

Automated by QA Swarm — not a human review