omp-peak-gate0.3.0

omp-peak-gate

Park remote DeepSeek LLM calls while DeepSeek's peak-pricing window is open — with opt-in fallback to your configured next model.

DeepSeek bills full rate Mon–Fri 01:00–04:00 and 06:00–10:00 UTC. This plugin engages the same process-global agentPauseGate the TUI /pause command uses at those boundaries, so:

• in-flight provider streams finish (already billed),
• no new model call starts while the window is open,
• every agent loop parks instead of aborting — queued prompts and steering messages deliver normally once the window closes.

Scope: only remote DeepSeek models. The gate engages only when the session's current model has deepseek in its provider, model id or name and is served off this machine. Everything else keeps running during the window:

• other models (Claude, GPT, your default model, …) — never gated,
• DeepSeek checkpoints running locally — Ollama, llama.cpp, LM Studio, vLLM and friends (loopback baseUrl or a known local provider id) — never gated, they are not billed at DeepSeek peak rates.

/peak-gate pause still pauses every model regardless of scope.

Fallback (opt-in). If the gated model has an entry in oh-my-pi's own retry.fallbackChains settings, the session switches to the next usable model at the window's start instead of parking, and switches back when the window closes. No chain configured → the gate parks as before; there is nothing to set in this plugin. Candidates that are themselves remote DeepSeek models are skipped (they would not escape peak pricing); a local DeepSeek checkpoint in the chain is a valid, free target.

The gate decision is a pure function of the wall clock, the session model and a static window table, so suspend/resume, clock jumps and timer jitter cannot drift it. An exact edge timer engages/releases at the boundary; a 30-second heartbeat re-derives state (and notices model changes).

Windows are a static weekly UTC table. There is deliberately no holiday calendar — the subjectively-set, yearly Chinese holiday dates are not modeled. Every weekday listed in the table is subject to the gate.

Resuming after the pause

Parking only works while a loop is alive. A run can end while the gate holds it, and then nothing is left to wake when the gate opens. The plugin therefore covers the omp edge cases around that:

Situation	What happens
Loop parked at a model/tool boundary	agentPauseGate.resume() wakes it; no prompt is injected (a live loop makes the check a no-op).
Run aborted while parked (host timeout, deadline, provider abort)	One continuation prompt is sent when the gate opens and the session is idle.
wait tool timed out or returned during the pause	The tool result wakes the loop; if the loop already settled, the aborted-run path above sends the prompt.
Gate released manually while an interruptible wait is still in flight	After a hold of at least 5s, the plugin sends an aside message that ends the wait, so the turn returns to the model instead of sitting out the 30-minute cap.
Steer/follow-up queued while an idle session was paused	The queue never starts a turn by itself; the plugin starts one on release so the queued batch drains.
Process died mid-park, or a host deadline (--max-time) disposed the session before the gate opened	A per-session marker (~/.omp/agent/peak-gate-resume.json, 24h TTL) causes one continuation prompt when that session is next resumed. Cleared on an ordinary graceful shutdown, kept when a resume is pending.
User pressed Esc / deliberately interrupted	Never auto-resumed; a user-interrupt abort is distinguished from a host/deadline abort.
Fallback model took over	The gate only pauses briefly (or not at all); no resume prompt is generated.

Auto-resume is bounded to two injections per hold, never runs in subagent sessions, and can be disabled entirely with OMP_PEAK_GATE_RESUME=0. The injected prompt is attributed to the agent, not the user. /peak-gate status shows whether a run is live, whether it observed a pause, whether a wait is active, and whether a resume is armed.

Install

# local checkout / development (symlinked, edits are live)
omp plugin link /path/to/omp-peak-gate

# git
omp plugin install github:alosslessdev/omp-peak-gate-main

# marketplace (this repo ships .omp-plugin/marketplace.json)
/marketplace add alosslessdev/omp-peak-gate-main
/marketplace install peak-gate@omp-peak-gate

# npm
omp plugin install omp-peak-gate

Check it landed: omp plugin list, then /peak-gate status in a session.

Do not also keep a copy at ~/.omp/agent/extensions/peak-gate.ts — both paths load, and the gate would be registered twice.

Commands

Command	Effect
/peak-gate	Show state, windows, model scope, fallback, next boundary
/peak-gate pause	Hold LLM calls paused until the next window boundary
/peak-gate resume	Release now, skip the current window (restore a fallback)
/peak-gate reload	Re-read the config file and re-arm the timers

While the gate is held, the status line shows ⏸ peak-gate until <UTC>; while running on a fallback model it shows ⏸ peak-gate ↓<model> until <UTC>. pause pauses every model regardless of scope; resume releases the gate and switches back to the original DeepSeek model if a fallback was applied.

Configuration

~/.omp/agent/peak-gate.json (agent dir follows PI_CODING_AGENT_DIR):

{
  "windows": [
    { "startMinute": 60, "endMinute": 240, "weekdays": [1, 2, 3, 4, 5] },
    { "startMinute": 360, "endMinute": 600, "weekdays": [1, 2, 3, 4, 5] }
  ]
}

• startMinute / endMinute are UTC minutes of day, half-open [start, end); startMinute must be less than endMinute (a midnight-crossing window is two entries). endMinute: 1440 means midnight.
• weekdays uses Date#getUTCDay numbering (0 = Sunday). Omit it for daily.
• { "enabled": false } disables the plugin without uninstalling it.

Fallback (opt-in, configured in omp itself)

Fallback uses oh-my-pi's standard retry.fallbackChains — nothing plugin-side to configure. In ~/.omp/agent/config.yml (or project .omp/config.yml):

retry:
  fallbackChains:
    deepseek/deepseek-v4-flash:
      - xiaomi/mimo-v2.6-flash:high
      - anthropic/claude-sonnet-4
    default:
      - ollama/qwen3:8b

At the window's start the gated model's chain is walked (same resolution core uses on provider errors: exact model key, provider wildcard, model role, then default) and the first usable non-remote-DeepSeek model takes over. Without an entry the gate parks instead — fallback stays opt-in.

Environment overrides:

Variable	Effect
OMP_PEAK_GATE=off	Disable the gate for this process
OMP_PEAK_GATE_WINDOWS=JSON	Replace the window table ([{"startMinute":60,"endMinute":240}])
OMP_PEAK_GATE_QUIET=1	Suppress the "a call slipped through" warning
OMP_PEAK_GATE_KEEPALIVE=0	Let a parked headless run exit instead of waiting out the window
OMP_PEAK_GATE_RESUME=0	Disable auto-resume (prompt injection, wait interrupt, restart marker)
OMP_PEAK_GATE_DEBUG=/path	Append decision lines to a log file

What is and is not gated

Gated (the gate is polled before each model call and each tool call): the main agent, in-process subagents, advisors, tool loops — i.e. everything that spends through an agent loop — but only while the session model is a remote DeepSeek model (see scope above). The gate itself is process-global: a session parked on DeepSeek holds every loop in the process until the window closes or the fallback switch releases it.

Not gated: a few side channels stream outside any agent loop (/btw side turns, session title generation, remote compaction). before_provider_request cannot veto them — that hook may only replace the payload and a thrown error is swallowed — so they are surfaced as a throttled warning instead of being billed silently. Use OMP_PEAK_GATE_QUIET=1 to silence it. Cache warming is stopped for the duration of a hold (cache_warming_decision → stop), since it is a replay outside the loop; it resumes on the next real request.

Never gated: non-DeepSeek models and locally served models — they do not match the scope, so no window can pause them. To pause everything manually, use /peak-gate pause; to change when the window applies, edit windows.

Tests

bun test

test/peak-gate.test.ts loads the real extension against a mocked omp runtime (fake clock, timers, pause gate and session context) and covers the resume watchdog, wait interruption, gate ownership across sessions, and the restart marker.

Layout

package.json                     omp.extensions -> ./extensions/peak-gate.ts
extensions/peak-gate.ts          the extension module
test/peak-gate.test.ts           deterministic edge-case tests (bun test)
.omp-plugin/marketplace.json     catalog so the repo is its own marketplace