Likely duplicate of #10448 — same per-turn live time injection in run_agent.py. Also related to #15866 (prompt cache invalidation from timestamp).

Likely duplicate of #10448

Thanks for reviewing @alt-glitch! This is not a duplicate of #10448 — they fix the same symptom but with fundamentally different architectures that have real production implications.

Key difference: prompt cache preservation

The Hermes codebase explicitly documents this principle (in run_conversation, around line 9449):

"Context is ALWAYS injected into the user message, never the system prompt. This preserves the prompt cache prefix — the system prompt stays identical across turns so cached tokens are reused."

This isn't theoretical — Anthropic's 2026 best practice also recommends per-turn context in the user message specifically for cache stability. When the system prompt changes each turn, every API call pays full token cost for the system prompt instead of leveraging the cached prefix.

What #10448 does better

• ✅ Covers the codex responses path (we don't yet)
• ✅ Clean _build_live_time_context() helper function
• ✅ More test paths

What this PR does better

• ✅ Preserves prompt cache — critical for production cost and latency
• ✅ Uses the existing plugin injection mechanism (_plugin_user_context) — composes naturally with plugins
• ✅ Changes "Conversation started" → "Session started" — eliminates the ambiguity that caused the bug
• ✅ Adds timezone name to both timestamps

Proposed resolution

The ideal fix would combine:

1. User-message injection (from this PR) for cache preservation
2. Codex responses path coverage (from fix(agent): inject turn-level live time context #10448)
3. _build_live_time_context() helper (from fix(agent): inject turn-level live time context #10448, but output injected into user message)

I'm happy to add codex responses coverage to this PR if the maintainers prefer the user-message approach, or to close this PR in favor of #10448 if they accept the cache tradeoff. But they are architecturally different and shouldn't be merged as duplicates — the cache question needs a deliberate decision.

Thanks for the thorough review @Bartok9! Appreciate the ✅ on all four changes.

Re: your suggestion about dropping the minute portion from Session started: — that's a great point. Since the per-turn injection now provides precise time, the session start only needs to establish when the session began for context. Date + hour (e.g. Session started: Sunday, April 26, 2026 05:xx AM HKT) would be sufficient and would further stabilize the cache prefix across sessions that start in the same hour.

I'll hold off on making that change for now to keep this PR focused on the core fix. Happy to follow up in a separate PR if the maintainers want that optimization.

Also — I've rebased #12630 and #12987 onto the latest main (both were dirty/unstable from the upstream merge wave today). #12630 had one conflict in cli.py where upstream's ModelInfo refactor already removed a fallback branch that the original PR was patching — resolved by accepting upstream's version. #12987 rebased cleanly with zero conflicts.

This PR looks like the preferred salvage path for #17459/#17476 because it separates stable session-start context from live per-turn time context rather than mutating the cached system prompt.

Please align the final version with the fan-out criteria:

• keep volatile Current time out of the cached system prompt/cacheable prefix;
• include user timezone in the ephemeral runtime context;
• align timezone resolution with fix(timezone): propagate configured timezone to agent prompt and terminals #10061;
• add/keep tests proving cache stability while the per-turn time context updates;
• avoid adding a separate gateway-only timestamp-prefix mechanism.

Related: #10421, #17459, #17476, #17474, #17475.

Thanks for the clear alignment criteria @markojak. I've addressed all five points in the latest update (single squashed commit 046f2deb):

Alignment checklist

1. ✅ volatile Current time out of cached system prompt — injected into _plugin_user_context (user message), not system prompt. Tests confirm "Current time:" never appears in _cached_system_prompt.

2. ✅ user timezone in ephemeral context — format now aligned with fix(timezone): propagate configured timezone to agent prompt and terminals #10061: Timezone: Asia/Hong_Kong (UTC+08:00) (IANA name + UTC offset). New get_timezone_display() in hermes_time.py computes the offset from hermes_time.now().utcoffset(), so DST-aware zones show the correct offset at runtime. Agents that don't understand IANA names can still determine the offset from the UTC±HH:MM portion.

3. ✅ timezone resolution aligned with fix(timezone): propagate configured timezone to agent prompt and terminals #10061 — get_timezone_display() uses the same resolution path as fix(timezone): propagate configured timezone to agent prompt and terminals #10061's _resolve_configured_tz_name(): reads both HERMES_TIMEZONE env var and ~/.hermes/config.yaml timezone key via hermes_time.get_timezone(). The display format IANA (UTC±HH:MM) matches fix(timezone): propagate configured timezone to agent prompt and terminals #10061's Timezone: Asia/Shanghai (UTC+08:00) exactly.

4. ✅ tests proving cache stability — 5 tests cover:

   • test_excludes_current_time_from_cached_prompt: system prompt must not contain "Current time:"
   • test_timezone_display_format_aligns_with_10061: system prompt contains "Timezone: Asia/Hong_Kong (UTC+08:00)" on a separate line
   • test_turn_level_time_injected_into_user_message: Current time in user messages, not system
   • test_summary_injects_current_time_into_user_message: max_iterations recovery path also injects time
   • test_run_conversation_codex_injects_current_time_into_user_message: codex_responses mode covered

5. ✅ no gateway-only timestamp-prefix mechanism — all time injection at AIAgent level via _plugin_user_context, same as plugins. No gateway timestamp prefixing.

What changed from previous version

• get_timezone_name() → get_timezone_display() (IANA + UTC offset format)
• System prompt timezone: inline → separate Timezone: line matching fix(timezone): propagate configured timezone to agent prompt and terminals #10061
• Per-turn + max_iterations: same aligned format

Not changed (by design)

• Terminal TZ propagation is fix(timezone): propagate configured timezone to agent prompt and terminals #10061's scope
• Session started: still has minute precision (follow-up optimization per @Bartok9)

Rebased onto latest main (May 2)

Resolved the merge conflict in tests/run_agent/test_run_agent.py — upstream added two codex-related tests (test_codex_summary_sanitizes_orphan_tool_results, test_api_sanitizer_matches_responses_call_id_when_id_differs) in the same test class as our test_summary_injects_current_time_into_user_message. Both sets are independent; the resolution keeps all tests.

Current status vs. related PRs

What I would adopt from #18135

The prepend_current_time_context() helper in hermes_time.py is well-designed — especially the idempotent check and multimodal list-content support. If maintainers prefer, I can refactor this PR to use that helper (keeping _plugin_user_context as the injection channel, timezone aligned with #10061, and max_iterations + codex_responses coverage).

The transcript-cleanliness concern (_persist_user_message_override) is also valid — worth addressing in a follow-up or as part of the helper refactor.

Summary

This PR is now conflict-free, approved, and aligned with all five criteria from @markojak. Ready for merge or further iteration per maintainer preference.

Adopted format_current_time_context() helper from #18135

After reviewing #18135 (WXBR), I've adopted its best idea — a centralized formatting helper in hermes_time.py — while keeping our architectural advantages (cache-safe injection, #10061 timezone alignment, max_iterations + codex coverage).

What changed (commit 31ed790)

New: hermes_time.format_current_time_context() — single source of truth for per-turn time formatting, replacing inline formatting in run_conversation and _handle_max_iterations.

# Before (duplicated in 2 places):
from hermes_time import now as _ht_now, get_timezone_name as _get_tz
_current = _ht_now()
_time_ctx_parts.append(f"Current time: {_current.strftime(...)}")
_tz = _get_tz()
if _tz:
    _time_ctx_parts.append(f"Timezone: {_tz}")

# After (one import, one call):
from hermes_time import format_current_time_context
_time_block = format_current_time_context()

New: 5 unit tests in TestFormatCurrentTimeContext covering the helper directly:

• Timezone line present when configured, absent when not
• Explicit datetime passthrough for deterministic testing
• DST-aware offset verification (EDT vs EST)

What we kept from #18135

• ✅ Centralized helper — no more duplicated formatting logic
• ✅ Explicit datetime parameter — testability without monkeypatching now()
• ✅ Single format contract — format_current_time_context() guarantees consistent output

What we kept from our approach (not from #18135)

• ✅ IANA + UTC offset format — aligned with fix(timezone): propagate configured timezone to agent prompt and terminals #10061, not just %Z
• ✅ User-message injection via _plugin_user_context — composes with plugins
• ✅ _handle_max_iterations coverage — recovery path also gets current time
• ✅ codex_responses path coverage — tested separately

What we deliberately did NOT adopt from #18135

• ❌ [System note: Current time is ... .] format — less clear than Current time: ...\nTimezone: ...
• ❌ Multimodal prepend_current_time_context() — Hermes user-message injection goes through _plugin_user_context (string), not direct content mutation
• ❌ _persist_user_message_override transcript cleaning — valid concern but orthogonal to the core fix; worth a separate PR

Test results

30 passed in 10.13s  (all timezone + run_agent tests)

This PR now has: approved review ✅, no conflicts ✅, #10061 alignment ✅, centralized helper ✅, 11 tests (5 new + 6 existing) ✅.

Nice work adopting the format_current_time_context() helper from #18135 — that's exactly the right call for long-term maintainability. The centralized helper in hermes_time.py keeps the volatile/stable split clean without duplicating formatting logic. The rebase onto latest main looks solid too. ✅

Approve — user-message injection is the right call

This is the correct fix. The core insight is right: injecting Current time: into the system prompt (as #10448 does) changes the cache prefix every turn and burns cached tokens on every message. User-message injection composes with the existing plugin context mechanism and costs nothing.

The two-layer split is clean:

• Session started: in the system prompt — frozen, cache-stable, unambiguous label
• Current time: in the user message — dynamic, injected per-turn via _plugin_user_context

The label rename from Conversation started → Session started matters more than it looks. "Conversation started" is genuinely ambiguous — an agent seeing that at 9 AM after a 5 AM session start has a reasonable basis to interpret it as current time. "Session started" makes the frozen nature of the timestamp explicit.

The max-iterations path injection (_handle_max_iterations) is a good catch — recovery prompts need current time too, easy to miss.

4 tests covering cache safety, user-message injection, and the max-iterations path. All passing.

Relation to #10448: Same bug, wrong injection layer. This PR should supersede it.

Thanks for this PR — this is the correct approach.

I have been manually injecting current time into the user message (same location, same reasoning: preserve prompt cache) across two versions of the codebase now, and it works exactly as expected. It solves the stale-timestamp problem without the cache penalty that #10448 would introduce.

This is a ~5-line change with passing tests, and the core issue has been reported across multiple threads (#10421, #9628, #27742, #17459, #17476). It feels like the only thing missing is someone with a merge button making a call.

Just wanted to add a user voice to the discussion — is there anything blocking this that the community can help with?

Rebased onto latest upstream/main (v2026.5.29, commit 8738cb9).

Changes adapted to the 3-module refactor:

• run_agent.py -> agent/system_prompt.py (Conversation->Session rename + timezone)
• run_agent.py -> agent/conversation_loop.py (per-turn time injection into _plugin_user_context)
• run_agent.py -> agent/chat_completion_helpers.py (max-iterations time injection)
• hermes_time.py: get_timezone_name() helper (unchanged)
• tests/run_agent/test_run_agent.py: 5 tests (updated 3 + added 2 new)

All 5 tests passing, clean single commit (+95/-6).

Maintainer alignment criteria (per @markojak):

1. Volatile Current time out of cached system prompt
2. User timezone in ephemeral context (IANA name)
3. Timezone resolution aligned with fix(timezone): propagate configured timezone to agent prompt and terminals #10061
4. Tests prove cache stability while per-turn time updates
5. No gateway-only timestamp-prefix mechanism

Ready for merge.

Rebased onto v2026.6.5 (commit c94e93a)

Adapted to the turn_context.py extraction refactor:

• Time injection moved from conversation_loop.py → agent/turn_context.py (new file)
• format_current_time_context() helper in hermes_time.py (single source of truth)
• get_timezone_display() for IANA + UTC offset format (fix(timezone): propagate configured timezone to agent prompt and terminals #10061 alignment)
• Max-iterations injection in agent/chat_completion_helpers.py
• 11 tests, all passing

Diff: +220/-1 across 5 files. Ready for merge.