Context / regression

Quiet hours should not be a control-plane policy that silently withholds or delays agent actions. Per the tool-use principles we want the agent to see the relevant runtime facts and make/justify the decision, not have a background subsystem enforce user-contact policy invisibly.

The current direction is fragmented:

• Closed, not merged / should not be revived as-is
  • feat: Add profile-scoped gateway autonomy runtime #4711 feat: Add profile-scoped gateway autonomy runtime introduced gateway autonomy with quiet-hours controls in the control plane. It is closed and unmerged; do not resurrect the quiet-hours enforcement shape.
  • feat: drain Ockham outbox notifications via Hermes scheduler #7402 feat: drain Ockham outbox notifications via Hermes scheduler was closed and unmerged; it explicitly held approvals/digests during quiet hours. That is the wrong semantic for Hermes core unless the agent is told and chooses deferral.
• Open time-awareness work to salvage / consolidate
  • [Feature] Turn-level live time context for current date/time awareness #10421 captures the core need: turn-level live current time awareness.
  • fix: prevent stale timestamp perception by injecting current time per-turn #15872 is the preferred implementation shape: inject Current time + timezone into the user-message/ephemeral context path, preserving prompt-cache stability.
  • fix(agent): inject turn-level live time context #10448 and feat: inject current time into system prompt on every API call #5241 address similar bugs but put live time into the system-prompt path; that should be avoided because it undermines prompt caching.
  • fix(timezone): propagate configured timezone to agent prompt and terminals #10061 is adjacent and useful: propagate configured timezone to agent prompt/terminal environments.
  • feat(gateway): add optional inbound message timestamp prefix #9784/feat(gateway): inject current timestamp into user messages #5487 are gateway-only timestamp-prefix variants; we should avoid multiple parallel timestamp mechanisms and consolidate on one core ephemeral context path.
• Open caching regressions related to this work
  • [Bug]: get_tool_definitions() quiet-mode cache pollution causes duplicate LCM tool schemas in Gateway #17335: get_tool_definitions() quiet-mode cache pollution causes duplicate LCM schemas in Gateway. This should be treated as an immediate bug fix.
  • Question: does the minute-precision timestamp in _build_system_prompt invalidate prompt caching for upstream inference servers? #15866/perf: system prompt timestamp changes after compression, invalidating prefix cache on local LLM backends #8687/perf: stabilize system prompt timestamp across compression cycles #8689/feat(caching): multi-block system prompt with tiered TTLs (v2) #5713: system-prompt timestamp / prompt-cache invalidation concerns. The durable direction is stable cached system prompt + volatile current-time context in user-message/tool-runtime context.
  • fix: codex_responses prompt caching — session routing headers + cache_write_tokens field #10006/feat(prompt-caching): wire cache_control breakpoints into codex_responses adapter #16235: Codex/Responses caching still needs proper cache telemetry/routing/breakpoints.

Goal

Move from control-plane quiet-hours enforcement to agent-visible, timezone-aware tool context:

1. Fix the caching regression(s).
2. Give the agent reliable current time + user timezone context on every turn and on relevant tool calls.
3. For time-sensitive tools (email, calendars, agent mailbox, send/wait/scheduler-style tools, future autonomy/watch tools), surface a clear advisory when the action appears outside the user's normal contact window — but do not block or delay it in the control plane.

Proposed design

1. Stable prompt cache, live time outside the cached system prompt

• Keep the cached system prompt byte-stable across turns.
• Replace the ambiguous Conversation started: mental model with a stable session-start anchor only.
• Inject volatile time context through the existing ephemeral user-message context path, e.g.

Runtime time context:
- Current time: 2026-04-29 08:34:12 America/Los_Angeles (UTC-07:00)
- User timezone: America/Los_Angeles
- Session started: 2026-04-29 08:00:02 America/Los_Angeles

• Prefer/salvage fix: prevent stale timestamp perception by injecting current time per-turn #15872's approach over per-turn system-prompt mutation.

2. Tool-runtime time advisory, not enforcement

Add a small shared helper/context object available to tool handlers, for example:

ToolRuntimeTimeContext(
    now=<aware datetime>,
    user_timezone="America/Los_Angeles",
    contact_window=("08:00", "23:00"),
    within_contact_window=True | False,
    advisory="It is 02:14 in the user's timezone; this may be outside their normal contact window. Ask/consider deferring unless urgent."
)

Tools may include this in their returned JSON/result text when relevant. The control plane must not silently suppress, defer, or rewrite the action.

Examples:

• wait / process.wait / future agent-mailbox wait:
  • Return current local time and deadline/elapsed-time interpretation.
  • If waiting would cross the user's normal contact window, warn the agent.
• Email/calendar/mailbox tools:
  • Include current_time, user_timezone, and contact_window_advisory in reads/searches and before sends/creates where relevant.
• send_message / proactive delivery paths:
  • Warn if sending outside the window; do not block. The agent can explicitly say why it is sending now or choose to schedule/defer.

3. Remove/avoid control-plane quiet-hours semantics

• No core gateway/scheduler/autonomy subsystem should silently hold messages merely because of quiet hours.
• If a subsystem keeps a user-configured contact window, it should expose that as agent context/advisory only.
• Any future hard block must be a separate, explicit user-configured safety setting with visible tool feedback, not the default quiet-hours behavior.

Implementation slices

1. Immediate cache correctness

   • Fix [Bug]: get_tool_definitions() quiet-mode cache pollution causes duplicate LCM tool schemas in Gateway #17335: defensive-copy cached tool definition lists and dedupe context-engine tool injection.
   • Add regression tests for Gateway/quiet-mode repeated agent creation with LCM/context-engine tools.

2. Core time context consolidation

   • Salvage fix: prevent stale timestamp perception by injecting current time per-turn #15872: inject current time/timezone into ephemeral user-message context per turn.
   • Keep system prompt cache-stable; update tests to prove Current time: is not in _build_system_prompt().
   • Incorporate timezone helper pieces from fix(timezone): propagate configured timezone to agent prompt and terminals #10061 where still applicable.
   • Close/supersede fix(agent): inject turn-level live time context #10448/feat: inject current time into system prompt on every API call #5241/feat(gateway): add optional inbound message timestamp prefix #9784/feat(gateway): inject current timestamp into user messages #5487-style duplicate mechanisms once consolidated.

3. Tool runtime advisory helper

   • Add a shared helper for aware now, resolved user timezone, configurable advisory window, and advisory text.
   • Wire it into time-sensitive tools first: wait/process-wait, email/calendar/agent-mailbox if present, send/proactive-delivery surfaces.
   • Ensure tool output is advisory and visible to the model.

4. Docs / principles alignment

   • Document: "time policy is surfaced to the agent; control plane does not silently enforce quiet hours."
   • Add tests that verify no quiet-hours path silently suppresses delivery/action without returning an advisory to the agent.

Acceptance criteria

• get_tool_definitions(quiet_mode=True) cache pollution is fixed; repeated Gateway agents do not accumulate duplicate tool schemas.
• Current time + user timezone are available on every agent turn without mutating the cached system prompt.
• Prompt-cache stability tests prove volatile time context is excluded from the cached system prompt.
• Time-sensitive tools return explicit time/timezone/advisory metadata when appropriate.
• Quiet-hours/contact-window behavior is advisory by default and visible to the agent; no default control-plane suppression/delay.
• Duplicate timestamp-prefix/current-time PRs are either closed as superseded or rebased into the chosen core path.
• Docs explain the principle and the intended extension point for future tools.

Non-goals

• Building a full policy engine.
• Adding a broad autonomy control plane.
• Hard-blocking sends/actions outside contact windows by default.
• Moving volatile time data into the system prompt.

Suggested priority

P1. This blocks reliable gateway use because caching/schema pollution is already user-visible (#17335), and the quiet-hours direction is an architectural regression if it lands as hidden control-plane enforcement.