fix(agent): drop the hour from the system-prompt date line (day-stable prompt-cache prefix)

[philmerrell]

Why

get_current_date_pacific() rendered YYYY-MM-DD (Weekday) HH:00 TZ, and SystemPromptBuilder.build() appends it to the system prompt, which is the head of the Bedrock prompt-cache prefix. Every Pacific hour boundary flipped systemPromptHash and re-wrote the whole cached prefix for every active session, at the cache-write premium.

The 2026-09-15 prod cost audit measured 23 distinct systemPromptHash values in one 85-call session and attributed ~2.6% of September cache-write spend to this.

What changed

• backend/src/agents/main_agent/utils/timezone.py: the line now renders date, weekday and timezone only (2026-09-15 (Tuesday) PDT), in all three code paths (zoneinfo, pytz fallback, UTC fallback). Function name and signature are unchanged. The docstring records why the hour must stay out of the prefix and where to put one if a flow ever needs it (the user turn, never the system prompt).
• backend/src/agents/main_agent/core/system_prompt_builder.py: two-line comment at the render site so the hour isn't reintroduced.

Hour-dependency check

Grepped develop for the helper, the %H:00 format, (Weekday) HH:00 pins, and "current time" / "time of day" language across the agent code, shared APIs, skills, local tools, the BFF chat route, docs and the SPA. The only consumers are the builder and the two test files below. No prompt, skill or tool reads the hour, so nothing needed to move into a turn-scoped message.

Tests

• tests/agents/main_agent/utils/test_timezone.py: new format regex, plus frozen-clock tests that the string is identical across all 48 half-hours of a Pacific day, flips exactly at Pacific midnight, uses the Pacific (not UTC) date, and renders PST in standard time.
• tests/agents/main_agent/core/test_system_prompt_builder.py: the six mocked return values lose the hour.
• tests/agents/main_agent/core/test_bedrock_cache_points.py: unchanged, still passes.

Suite	Result
timezone + builder + cache-points	43 passed
all of tests/agents/main_agent	1,274 passed
ruff on changed files	clean

Compaction code is untouched (that work is on feature/compaction-model-relative-thresholds).

Second commit: pre-existing SPA spec order dependency (unrelated to the fix above)

The first CI run failed twice in Test frontend (coverage build) on file-preview-panel.component.spec.ts even though this PR touches no frontend file. Cause: DocxViewerComponent memoizes its dynamic import('docx-preview') on a static field, and the builder runs vitest with isolate: false, so when docx-viewer.component.spec.ts runs first in a worker it pins its mocked module for the panel spec, whose mockRejectedValue then programs a function the component never calls.

Develop's exact frontend tree had never been exercised by this job: CI runs only on PRs, the docx specs landed yesterday in 76e022a, and every PR that passed since adds or changes spec files, which reshuffles worker assignment. Any PR without frontend spec changes would hit this.

Fix is test-only (9 lines): both specs null the memo in beforeEach. Reproduced locally with a single-worker, alphabetically ordered runner config, then verified with the full coverage build (257 files / 3,212 tests pass).

Expected effect

Once deployed, systemPromptHash on a session's C# rows should change at most once per Pacific day (and at the PST/PDT switch) instead of hourly. Verify via GET /admin/costs/sessions/{id}/calls or the distinctSystemPromptHashes diagnosis.

🤖 Generated with Claude Code