🧪 fix: Count System Instructions in Mock Model Usage

[danny-avila]

Summary

Every PR that runs e2e against current dev fails the same three specs:

• usage.spec.ts:111 and :167: the context-usage popover never shows its Messages row
• scenarios/series-eight-theme.spec.ts:84: the Messages swatch (span.bg-series-1) is missing from the same popover

What broke it: #16301 (Surface Persistent Memory Before First Save). Memory is enabled in the e2e config, and an empty memory store now still adds the memory-instructions paragraph, so every mock chat now has a system prompt of about 57 tokens. #16301 skipped e2e (backend-only path filter), so the first PRs to run e2e afterwards inherited the failure.

Why a system prompt breaks the gauge in the mock only: the SDK passes a test override model the pruned messages without the systemRunnable pipe (prepareProviderRequest({ model: this.overrideModel ?? model })). The fake model therefore saw only [human "hello"] and reported input_tokens: 2. The context snapshot calibrates against that usage, so the popover showed "used" as 7 (2 input + 5 output) against a 57-token system prompt. That made messageBudget = max(0, 7 − 57) = 0, and the zero-value Messages row disappeared. A real provider bills the system prompt, so production is unaffected.

Fix: UsageEmittingFakeChatModel now counts input over the complete prompt, using the harness's existing getStreamAgentView. That helper already applies the agent's systemRunnable for prompt assertions, for the same reason. The only other change is replacing a nested ternary lower in the file, because the pre-commit hook lints the whole file and blocked the commit on it.

before:  fake input = [human "hello"]                 → input_tokens 2  → used 7 < system 57 → Messages hidden
after:   fake input = [system <instructions>, human]  → input_tokens ≈ 60 → Messages row renders

Testing

Run locally against edddf92 (current dev), in-memory stream store:

• On dev: usage.spec.ts:111 fails, reproducing CI.
• With 🪴 fix: Surface Persistent Memory Before First Save #16301 reverted: the same test passes, which confirms the cause.
• With this fix: usage.spec.ts and series-eight-theme.spec.ts pass, 8/8.
• Regression sweep over the other mock specs that assert usage, cost, compaction or context: results to follow in a comment.

[danny-avila]

Local regression check with this fix: 91/91 passed on first try, 0 flaky (chromium, memory store, retries=1, none used). Specs covered: activity-phases, agent-file-ownership, content-filters.persisted, message-tree, model-replay, quotes, compaction-rerun-controls, failed-summary-history, persisted-context-usage, subagent-context-total, tool-context-split, thread-fold, steering, tool-approvals, completion. These are the mock specs that assert on usage, cost, compaction or context, plus the long-running suites.