OpenClaw Temporal Halo

Temporal Halo is an OpenClaw plugin that maintains a living HALO.md, writes a compact HALO.delta.md sidecar for incremental refreshes, and injects current temporal context into each agent turn. It is tool/skill agnostic: the prompt tells the agent what to gather, and the agent uses whatever calendar/email/messages tools you already have connected.

1) Installation

openclaw plugins install @shayne/openclaw-temporal-halo
openclaw plugins enable openclaw-temporal-halo

Optional plugin config (plugins.entries.openclaw-temporal-halo.config):

{
  plugins: {
    entries: {
      "openclaw-temporal-halo": {
        enabled: true,
        config: {
          haloPath: "~/.openclaw/temporal-halo/HALO.md",
          dreamMarker: "[temporal-halo:dream]",
          fullRefreshMarker: "[temporal-halo:full-refresh]",
          maxChars: 25000,
          compactTargetChars: 20000,
        },
      },
    },
  },
}

Recommended immediately after install: run the one-off full refresh in section 3 to bootstrap HALO.md.

2) Set Up Cron Job (30m Delta Refresh)

Create a recurring main-session system event job:

openclaw cron add \
  --name "Temporal Halo: Dream (Delta)" \
  --every "30m" \
  --session main \
  --wake now \
  --system-event "[temporal-halo:dream] Refresh HALO.md with recent changes."

Delta mode behavior:

• Scheduled runs are incremental.
• Prompt steers the agent to use the HALO.delta.md sidecar as the first-pass change ledger, then fall back to the full HALO.md snapshot only when needed.
• If prior refresh timing is unknown, prompt falls back to about the last 30 minutes with a small overlap to avoid misses.
• Each medium/important publish can write both a refreshed HALO.md snapshot and a compact HALO.delta.md summary with added, updated, retired, and still-open items.
• Prompt steers the agent to retire stale bullets when they are resolved, expired, superseded, duplicated by fresher facts, or no longer useful for likely user action/disambiguation.
• The run remains map/reduce and subagent-first (sessions_spawn fan-out/fan-in).

3) Run Full Refresh (On-Demand)

Use this for initial bootstrap and occasional deep refreshes. This is intentionally on-demand (not recurring).

openclaw cron add \
  --name "Temporal Halo: Full Refresh (One-off)" \
  --at "+5s" \
  --delete-after-run \
  --session main \
  --wake now \
  --system-event "[temporal-halo:full-refresh] Rebuild HALO.md from a wider baseline."

Full refresh behavior:

• Uses the same map/reduce subagent orchestration.
• Scans a broader baseline (at least last 14 days) plus upcoming horizons to rebuild HALO.md.

Markers

• Delta marker (scheduled): [temporal-halo:dream]
• Full marker (on-demand): [temporal-halo:full-refresh]

Prompt Architecture

Delta and full refresh share one base dream prompt; mode-specific scope is added by mode lines in code:

• Dream mode detection: dream.ts
• Shared dream prompt builder: buildDreamInstructions in dream.ts
• Sidecar delta context block: buildHaloDeltaBlock in dream.ts
• Mode scope lines (delta vs full): buildDreamModeScopeLines in dream.ts

Notes

• One-off/cron runs can outlive short CLI timeouts; check run history if needed.
• HALO.md and HALO.delta.md can contain sensitive context. The full HALO.md snapshot is injected into each turn; HALO.delta.md is injected only for delta dream runs.

License

MIT. See LICENSE.