Skip to main content

Scheduled Jobs

EstaCoda cron runs scheduled automation from the active profile. A cron job can run an agent prompt, run a script before the agent prompt, run a script-only no-agent watchdog, load skill instructions, include output from upstream cron jobs, and constrain the model, tools, and workspace used for the run.

Cron is also an evidence surface for Agent Evolution. Runtime-backed jobs create isolated runtime sessions and trajectories that can be inspected later. Cron does not automatically create or promote skills yet.

Storage​

Cron job storage depends on how the runtime constructs its CronStore.

Profile-wired runtimes use profile-local paths:

~/.estacoda/profiles/<id>/cron/jobs.json
~/.estacoda/profiles/<id>/cron/output/

Some default/manual paths can use:

~/.estacoda/cron/jobs.json
~/.estacoda/cron/output/

Execution history is stored in ~/.estacoda/sessions.sqlite in the cron_executions table. It records status, timestamps, output summary, failure metadata, delivery results, and runtime session/trajectory IDs where a runtime exists.

Creating Jobs​

Use CLI commands, slash commands, or the cronjob tool.

estacoda cron add --schedule "every 1h" --command "Summarize recent project changes"
/cron add --schedule "every 1h" --command "Summarize recent project changes"

Useful add/edit flags:

--skill <name>
--script <path>
--script-arg <arg>
--script-timeout-ms <ms>
--no-agent
--agent
--context-from <job-id>
--clear-context-from
--model <model>
--provider <provider>
--clear-model
--toolset <name>
--clear-toolsets
--workdir <absolute-path>
--clear-workdir

Supported schedules include relative delays such as 10m, intervals such as every 2h, five-field cron expressions, and date strings accepted by JavaScript date parsing.

Runtime Isolation​

Gateway cron, /cron tick, and estacoda cron tick all use isolated cron runtimes. Scheduled work does not reuse the current conversation runtime.

Cron runtimes force-disable the cron, messaging, and clarify toolsets, pass a generated cron session ID, and are disposed after execution. If a job has an enabledToolsets allow-list, tools outside that allow-list are removed from the scheduled runtime. cron, messaging, and clarify cannot be re-enabled.

The scheduler is sequential today. It runs due jobs one by one under a global tick lock and per-job locks. Non-blocking bounded dispatch is not implemented.

Scripts And No-Agent Checks​

Agent-backed script jobs run the script first, redact the result, and inject it into the scheduled prompt.

No-agent jobs run only the script:

  • --no-agent requires --script.
  • Non-empty stdout is delivered.
  • Empty stdout is a silent success.
  • A final non-empty JSON line { "wakeAgent": false } is a silent success.
  • Script failures and timeouts produce classified, redacted alerts.

No-agent jobs do not create runtime sessions or trajectories.

Skills And Context​

Skills attached to cron jobs load actual skill instruction bodies. EstaCoda uses providerInstructions.content when present, otherwise instructions. Missing skills produce a warning in the scheduled prompt instead of crashing the job. Per-skill injected text is capped and scanned with the assembled prompt.

contextFrom injects the latest output from upstream cron jobs. Outputs are loaded from the cron output root, kept in requested order, truncated, redacted, and labeled as data rather than instructions. Malformed persisted job IDs and path escapes are skipped.

Model, Tools, And Workdir​

--model and --provider set a per-job model route. If --model is supplied without --provider, EstaCoda pins the current provider before storing the job. Invalid provider/model routes are rejected before persistence.

--toolset creates an allow-list validated against actual registered runtime tools. Unknown toolsets and forbidden toolsets are rejected.

--workdir selects an effective workspace for the job. It must be an absolute existing path inside an allowed workspace root. EstaCoda canonicalizes paths with realpath, rejects symlink escapes, and derives trust from the workspace trust store. Cron does not grant trust to arbitrary absolute directories.

Safety​

Raw prompts are scanned on create/update and again at runtime for legacy persisted jobs. Script output, upstream output, and skill text are redacted before prompt injection. The final assembled prompt is scanned and invisible Unicode is sanitized before runtime.handle().

Inspect And Repair​

estacoda cron list
estacoda cron show <job-id>
estacoda cron history [job-id]
estacoda cron run <job-id>
estacoda cron pause <job-id>
estacoda cron resume <job-id>
estacoda cron remove <job-id>
estacoda cron tick

Clear advanced controls with:

estacoda cron edit <job-id> --clear-model
estacoda cron edit <job-id> --clear-toolsets
estacoda cron edit <job-id> --clear-workdir
estacoda cron edit <job-id> --agent

Limits​

  • Scheduler dispatch is sequential.
  • There is no cron web dashboard in this stack.
  • Cron does not automatically evolve or promote skills.
  • No-agent jobs do not create runtime trajectories.
  • Workdir support does not trust arbitrary absolute paths.