Skip to main content

Gitmoot v0.9.0

An email can now start a pipeline and hand it the message — end to end, declared in one YAML file, with the Activepieces glue generated and owned by gitmoot so you never open a flow-builder UI. Alongside it: a one-command Activepieces bootstrap, a production wave for pipelines (skipped decisions, intra-stage progress, PR-bound reviews, opt-in self-merge, data-writing produce stages), a kernel-enforced Landlock write sandbox, and a self-balancing memory hierarchy. Everything additive; existing setups behave identically unless you opt in.

What's new since v0.8.8​

Email-triggered pipelines, end to end (#851, #861, #863)

name: triage-email
repo: owner/repo
trigger:
  kind: email
  map: {subject: subject, sender: from_address, body: text}
stages:
  - id: route
    cmd: 'echo "got: $GITMOOT_TRIGGER_SUBJECT"'
  - id: triage
    agent: lead
    prompt: Read the trigger payload and decide what to do.
    needs: [route]
  • One-command Activepieces bootstrap (#851). gitmoot activepieces setup --yes starts Activepieces + Postgres + Redis in Docker, spawns the token-authed gitmoot bridge, creates the admin account, installs the published @gitmoot/piece-gitmoot from npm, wires the bridge connection, and imports two starter flows. connect gmail creates a live-validated IMAP connection from the CLI (wrong app password fails immediately with the provider's error); --with-smtp, non-Gmail hosts, and test servers are flags away.
  • Declarative triggers — the AP UI becomes invisible (#861). A trigger: {kind: email} block on a pipeline makes pipeline add --enable generate, import, metadata-stamp, and publish an owned gitmoot: <name> flow via the Activepieces API. Ownership is durable (stored flow id + metadata, never name adoption); pipeline bind-trigger repairs — including recreating a flow deleted in the AP UI; disable is fail-closed (the bridge rejects runs of disabled pipelines even when AP is unreachable); removing the trigger block cleans up the stale flow.
  • Run inputs — the email reaches your stages (#863). The bridge accepts a bounded string payload (64 KiB body, ≤32 entries, ≤48 KiB decoded; duplicate JSON keys rejected at any nesting — unescape-aware), snapshots it immutably on the run, and exposes it two ways: shell stages get GITMOOT_TRIGGER_<KEY> environment variables delivered via execve (payload can never reach shell source text), and every agent stage's prompt is prefixed with a fenced UNTRUSTED trigger block (fence sized from post-truncation values so payload backticks can't escape). trigger.map compiles a closed selector set (subject, from_address, text, message_id, date) into the generated flow — raw expressions are never accepted. A new allow_triggered_writes double-key mirrors allow_scheduled_writes: email content cannot silently drive write-capable stages. Companion piece release: @gitmoot/piece-gitmoot 0.1.4 (optional payload object; 0.1.3 flows and older bridges interoperate unchanged).
  • Local flow execution actually works now. The embedded stack sets AP_INTERNAL_URL, fixing a bug where every flow run on a setup stack died with INTERNAL_ERROR (engine self-calls hit the host-only published port); trigger poll cadence became configurable. Setup/down polish (#858): down warns instead of stranding the bridge, --stop-bridge stops only the verified process it spawned, failed template imports are retryable, and prompts never hang in scripts.

Pipelines: the production wave (#819, #821, #823, #839, #826, #829)

  • First-class skipped decision (#819) — a no-op stage advances the line instead of parking the run.
  • Intra-stage progress (#821) — elapsed + last-activity for long-running stages, bridged into the dashboard funnel (#841).
  • PR-bound reviews (#823) — source: binds a review stage to the PR its implement stage opened: [implement] → [review the PR] → [gate] is first-class, with a report-only guard.
  • merge: auto (#839) — an opt-in, double-keyed gate mode that lets a pipeline merge its own PR when the gate is green.
  • produce stages (#826) — agents that write data, never a PR: declared writable paths, a check: QA loop with bounded retries, token rollup; codex-first, now with Landlock-backed claude/kimi support (#840).
  • Wedge fix (#829) — implement stages no longer stall on non-implemented success decisions.

Landlock launch sandbox (#840)

Kernel-enforced write confinement for produce-stage agents: the launched runtime can only write its declared paths, enforced by the kernel rather than prompt discipline. Off unless the stage declares writes.

Memory: self-balancing knowledge (#838, #837, #827)

  • Auto-balanced cluster hierarchy (#838) — recursive agglomerative clustering with bounded fan-out keeps the Knowledge view legible as facts grow.
  • Brick-splitting groomer (#837) — oversized memories split losslessly at seams, automatically.
  • Fact-to-fact wiki-link edges (#827) — [[links]] between facts become first-class graph edges (the fact-galaxy server half).

Coordination & dashboard (#852, #866, #862, #836)

  • External-coordinator job grouping (#852) — jobs driven by an outside coordinator (like a Claude Code session) group under one workflow with a journal and a memory hook.
  • Workflow visibility (#866, #862) — galaxy hubs, a run strip, and a /workflows forest + journal view, with the server halves to feed them.
  • Checkout self-healing (#836) — registration guard plus self-heal for dangling checkouts.

Docs for all of the above are live on the site (Gmail + pipelines workflows, bridge payload reference, llms.txt/llms-full.txt regenerated). Upgrade is a drop-in binary replacement: gitmoot update.