Repository navigation
Event-driven automation triggers (GitHub/CI/Linear/Sentry/webhook) + NL authoring + per-automation memory #242
Description
Inspired by Cursor's Automations, the richest part of their cloud-agent offering.
Today Solenta automations are schedule-based (#160 covers a failure circuit breaker). Add event triggers: GitHub PR opened/pushed/merged, label changed, CI completed, review submitted; Linear issue created/status changed; Sentry error spikes; generic inbound webhooks. Also:
- NL authoring: describe the workflow in a chat session, Solenta generates the automation config (Cursor's /automate skill).
- Per-automation persistent memory: each automation gets a scoped memory namespace (memory-server) so recurring runs remember and improve.
- Permission scopes per automation (private vs project-shared).
Activity
currentbits commentedon Aug 17, 2026
Research note (Amp agent-set schedules): beyond static triggers, let the agent itself create and wake schedules with its full context ('check this backfill every 10 min, ping me when done'). An agent-authored automation is a first-class creation path alongside NL authoring. Also from Tembo: agents-as-workflow-executors with integration tools (post Slack summaries, update Linear, query Postgres) make automations useful beyond code changes.
currentbits commentedon Aug 17, 2026
Research note (loop-engineering, 10k stars): package automations as named loop patterns (daily-triage, PR-babysitter, CI-sweeper) each with a readiness score and a phased autonomy ladder (L1 report, L2 assisted, L3 unattended) plus a token-cost estimate per pattern, instead of binary on/off.
currentbits commentedon Aug 17, 2026
Deep-dive detail (Paseo): heartbeat vs schedule — a cron that injects a prompt into an EXISTING thread (PR babysitting, 'check the build') vs one that spawns a fresh agent. One-line model addition, outsized value. Also (Paseo hub): unattended dispatch must REFUSE if the provider cannot honor exact tool preapproval — never degrade to broader access.
currentbits commentedon Aug 18, 2026
Round 19 intel from Cursor Automations (changelog 2026-06-18), all directly in this issue's scope:
currentbits commentedon Aug 18, 2026
Round 23 intel: Factory Droid shipped incident response as a trigger pattern (v0.140.0): the agent automatically investigates production alerts posted to a configured Slack channel and works toward a root-cause analysis. Noteworthy framing: the deliverable is an RCA (a diagnosis with evidence), not a fix PR. That is a safer default for unattended triggers: investigate-and-report first, fix-on-approval second. Maps well onto the Sentry/webhook triggers already scoped here; consider making "investigate only, report back" vs "fix and open PR" an explicit per-trigger mode, with the blast-radius guard from #510 applying to the fix mode.
currentbits commentedon Aug 18, 2026
Round 27 intel: Orca shipped issue-intake from two more sources: Jira (paste an issue URL or search by text when creating a workspace; the card shows key and summary with a View-on-Jira link) and Linear (browse and create issues alongside GitHub, with a team selector and persisted last-used source). The intake pattern to note: the issue URL is the thread-creation entry point, not just a trigger: paste a ticket, get a workspace with the ticket as context. That composes with the planboard thesis: GitHub issues are Solenta's native source, but Jira/Linear intake meets enterprise users where their tickets live.
currentbits commentedon Aug 18, 2026
Round 28 intel: GitHub Next's gh-aw compiles agentic workflows written in Markdown into GitHub Actions YAML, and its security model is directly relevant to automation triggers: read-only by default, with writes only through sanitized "safe-outputs". The agent never gets write tokens; it emits structured output that a separate, non-agent step validates and applies. That is a stronger isolation pattern than prompt-level instructions, and it composes with #510 (workflow files as protected surface): Solenta automations that touch GitHub should emit intended writes as data, with the orchestrator applying them after validation.
currentbits commentedon Oct 9, 2026
Ranked #8 in the 2026-10-09 gap pass. Today automations are schedule-only (hourly/daily/weekly presets,
electron/automations.js:3, minute ticker:274).notify.jswebhooks are outbound only. Smallest useful step: an inbound webhook URL per automation, with a shared secret, on the existing--serve-webserver. Add a delivery log (last 50 deliveries with status) and an optional hold-while-offline queue. Pair it with #160 (circuit breaker).currentbits commentedon Oct 11, 2026
Shipped on branch coder/event-driven-automation-triggers-github--c4ee0e (51b22a3):
<automation-memory>…</automation-memory>. Editable in the Edit form. Not in memory-server: runs bind coder-memory to the project scope (Isolate Kimi runs: no cross-project threads, memory, or workspaces #671), so the server rejects a second namespace.Split out: #1588 (Linear/Sentry/generic webhooks), #1589 (NL authoring), #1590 (permission scopes).
GitHub event triggers and per-automation memory (#242) (#1591)
GitHub event triggers and per-automation memory (#242)