PR #81371: fix(cron): treat attempt dispatch as execution start

• Repository: openclaw/openclaw
• Author: hclsys
• State: open | merged: False
• Link: https://github.com/openclaw/openclaw/pull/81371

Description (problem / solution / changelog)

Fixes #81368

Summary

• Treat attempt_dispatch as an isolated cron execution-start milestone.
• Add a regression test proving the pre-execution watchdog is cleared after attempt_dispatch; the run then fails only at the configured job timeout.

Root cause

pi-embedded-runner can report attempt_dispatch, but isCronAgentExecutionStarted did not classify that phase as execution-started. For isolated cron runs that reach dispatch before the first model call, the 60s pre-execution watchdog could still abort the run as stalled.

Audits

• Existing helper: reused isCronAgentExecutionStarted; no new classifier.
• Shared caller: helper is private to src/cron/service/timer.ts and only broadens the set of positive execution milestones.
• Rival scan: no open direct PR for #81368. #80923 is closed/superseded and was a large dirty branch; #80888 is closed but the beta.4 repro in #81368 shows the local hotfix is still needed.

Tests

• pnpm format:check src/cron/service/timer.ts src/cron/service/timer.regression.test.ts
• node scripts/test-projects.mjs src/cron/service/timer.regression.test.ts
• pnpm check:test-types
• git diff --check

Real behavior proof

• Behavior or issue addressed: Isolated cron jobs that reach attempt_dispatch were still aborted by the 60s pre-execution watchdog as cron: isolated agent run stalled before execution start (last phase: attempt-dispatch), even though the configured job timeout was longer.
• Real environment tested: Reporter live setup from #81368: OpenClaw 2026.5.12-beta.4 (d124625), npm global install, Linux, isolated cron job, delivery disabled, model openai/gpt-5.5. Local patch in this PR applies the same classifier change on current main.
• Exact steps or command run after this patch: Reporter created the same no-delivery isolated cron job after applying the one-line case "attempt_dispatch": hotfix to the installed server-cron-*.js bundle and restarting the gateway:

openclaw cron add \
  --at "<now+15s ISO>" \
  --delete-after-run \
  --no-deliver \
  --session isolated \
  --agent cron \
  --model openai/gpt-5.5 \
  --timeout-seconds 240 \
  --message 'Run this exact shell command via exec: bash -lc "sleep 75". After it exits, reply exactly: OK_CRON_WATCHDOG_BETA4'

• Evidence after fix: Copied live runtime output from #81368 after the same one-line classifier hotfix and gateway restart:

{
  "status": "ok",
  "summary": "OK_CRON_WATCHDOG_HOTFIX",
  "durationMs": 128739,
  "model": "gpt-5.5",
  "provider": "openai",
  "deliveryStatus": "not-requested"
}

• Observed result after fix: The isolated cron run continued past the previous 60s false-stall cutoff and completed successfully at 128739ms with deliveryStatus: "not-requested".
• What was not tested: I did not rerun the live OpenAI cron smoke from this workspace; the after-fix live evidence is the reporter runtime log from #81368. I locally verified this source patch with the focused cron regression shard and test typecheck.

Not changed

• No watchdog duration changes.
• No runner phase emission changes.
• No cron scheduling or retry policy changes.

Changed files

• src/cron/service/timer.regression.test.ts (modified, +82/-0)
• src/cron/service/timer.ts (modified, +1/-0)