The system prompt intentionally excludes the current date/time for cache stability (see 66eec29). This leaves agents without date awareness, causing wrong day-of-week claims (#1897, #1928, #2108).

Instead of reverting the cache optimization, add a one-line hint directing agents to use session_status when they need the current date/time. This keeps the prompt stable while teaching frontier models where to look.

Also adds a negative test ensuring the date/time is NOT re-added to the system prompt, with comments explaining why and pointing to #3658 for the complementary gateway-level timestamp injection approach.

Refs: #1897, #1928, #3658

QA Results

Before Fix (no hint in prompt)

Test 1:

What day of the week is it today?
→ "Today is Thursday." (wrong)
Is it the weekend?
→ "Nope, not yet! Thursday means you've still got one more day" (wrong day)
How many days until February 1st?
→ "199 days until February 1st, 2026." (wildly wrong — actual: 4 days)

Test 2:

How many days until February 1st?
→ "Today is July 11, 2025. February 1st, 2026 is 205 days away." (hallucinated date)

Test 3:

Is it the weekend?
→ "Yes! It's Saturday, July 19, 2025." (hallucinated — it was Wednesday Jan 28, 2026)

After Fix — Main Session (full prompt)

Test 1:

How many days until February 1st?
→ "It's January 28, 2026 — so 4 days until February 1st." ✅ (used session_status)

Test 2:

What day of the week is it today?
→ "It's Wednesday, January 28th, 2026." ✅ (used session_status)
Is it the weekend?
→ "Nope — it's Wednesday evening. Middle of the week." ✅

Test 3:

How many days until February 1st?
→ "It's currently January 28th, so February 1st is 4 days away." ✅

After Fix — Spawned Subagent Sessions (minimal prompt)

We verified via debug logging that isMinimal does not strip the time hint — buildTimeSection is not gated by isMinimal, only by userTimezone. Confirmed: promptMode=minimal | hasTimeHint=true.

However, spawned agents behave differently:

Key findings:

• The hint IS present in minimal (subagent) prompts — confirmed via subsystem logging
• When models check, they prefer exec over session_status (reasonable — more direct)
• Simple yes/no date questions ("is it the weekend?") are non-deterministic — sometimes the model checks, sometimes it hallucates
• Computation questions ("how many days until...") reliably trigger tool use
• Zero session_status calls across all spawned tests — exec is the natural instinct

Complementary Fix: Gateway Timestamp Injection (#3658)

This hint helps frontier models that think to check, but doesn't help when models skip checking entirely. #3658 (gateway timestamp injection) is the stronger complement — it injects the date into the user message itself so models don't need to look it up.

However, #3658 as proposed injects timestamps into incoming user messages at the gateway level. Spawned subagent tasks may need the same treatment — the task text passed via sessions_spawn goes through callGateway({ method: "agent", params: { message: task } }) and currently has no timestamp injected. If #3658 only covers inbound channel messages (TUI/web), spawned agents would still be blind.

🤖 AI-Assisted

This PR was developed and tested collaboratively between a human and an AI agent (Claude, running in Moltbot). Testing included spawned-session tests across multiple rounds, debug logging verification, and live main-session QA. We understand what the code does — it's a one-line hint + a guardrail test.