fix(system_prompt): anchor 'Conversation started' to the real session start

[bobaba76]

What

Conversation started: in the system prompt was stamped with hermes_time.now() at system-prompt build time, not the conversation's real start.

The system prompt is rebuilt on compression, fresh-agent gateway turns, and resume-without-stored-prompt. So the date silently advanced to whatever day the prompt was last rebuilt — a chat that started on a Wednesday read as Conversation started: Thursday after a Thursday-morning resume, contradicting the fresh per-turn time hint (Current time:), and leaving the model's sense of "today vs last night" wrong in resumed threads.

Fix

New _session_start_like(agent, now) resolves the true start, in priority order:

1. The timestamp embedded in the session id (YYYYMMDD_HHMMSS_...) — immutable for the session's life, so the line is byte-stable across every rebuild boundary (preserving prefix-cache KV);
2. agent.session_start (the session-creation stamp);
3. now() only as a last resort (initial build with neither available).

Box-local stamps are attached to the box's local zone, then converted to the rendered zone (now's tzinfo), so the displayed date stays consistent with the per-turn clock even when the host TZ differs from the configured IANA zone (e.g. a remote gateway in UTC).

The existing zone-suffix and _bot_chat_timeless_prompt behaviour are untouched.

Tests

• TestSessionStartLike: unit coverage for the id-embedded timestamp, the session_start fallback, the now fallback, and a non-matching (non-YYYYMMDD_HHMMSS) session id.
• A build-level regression: a session started Jan 1 must still render Conversation started: Thursday, January 01 when the prompt is rebuilt on Jan 2.
• Full tests/agent/test_system_prompt.py + related suites pass (87 passed locally on Windows and Linux-compatible).

Separate note (drop if you prefer)

The second commit also fixes a pre-existing Windows-only portability bug in test_coding_prompt_preserves_legacy_workspace_order: it hardcoded /hermes while production renders str(Path('/hermes')) (backslash on Windows), so the suite fails on Windows but passes Linux CI. Happy to drop it from this PR if you'd rather keep the scope strictly to the timestamp fix — it's in its own commit for exactly that reason.

[kshitijk4poor]

Thanks for this — the analysis of the drift problem is solid, and the test coverage is good.

However, the approach here is the opposite of what we want. The goal isn't to freeze the "Conversation started" date to the original session start — it's to make the timestamp truthful about when the prompt was actually rebuilt. When the system prompt is rebuilt on compaction (or a fresh gateway turn, or resume), the date should reflect that rebuild, not a stale birth date from days or weeks ago.

This is especially important for bot mode and messenger platforms where a single session lives forever — pinning to the original start date becomes confidently-wrong misinformation within days. The existing _bot_chat_timeless_prompt path already handles eternal sessions by dropping the date entirely, which is the right call for those. But for normal sessions, the current now-based timestamp is already the truthful answer — the "drift" across midnight is actually correct behavior (the prompt was rebuilt on the new day).

The better direction is PR #80721 (still open), which announces date changes via the per-turn api_content sidecar without touching the prompt cache. That approach keeps the prompt byte-stable while giving the model accurate, fresh time context.

A few other notes from review:

• The session_id regex only matches 3 of 6 session_id formats (CLI, agent_init, gateway). Cron (cron_...), API server (api_...), and ACP (uuid.uuid4()) session IDs don't match — the fallback to session_start covers them, but the regex is dead code for half the session types.
• session_start is naive datetime.now() at all three sites — the timezone conversion is accidentally correct but fragile (uses server-local tz, not the configured IANA tz from hermes_time).
• The zone suffix comes from now but the date comes from _start — across DST boundaries this produces a mismatch (e.g. "July 01 (UTC-05:00)" when the session started in EDT).
• The Windows portability test fix (second commit) is legitimate but belongs in its own PR.

Closing this — appreciate the investigation, and credit for the thorough test coverage.