Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.crewship.ai/llms.txt

Use this file to discover all available pages before exploring further.

This page is for users who used the earlier pipelines API or CLI before the Routines feature was finalized. TL;DR: nothing breaks. The user-facing label changed from “Pipeline” to “Routine”, but every backend identifier — table, package, route paths, journal types, WS event types, command-line aliases — stays the same. Existing scripts, agents, and saved pipelines keep working.

​
What changed

​
User-facing label: Pipeline → Routine

SurfaceBeforeAfter
Sidebar entry(absent)Routines under Work
Orchestration tab(4 tabs)5th tab Routines
Page route(absent)/routines full management page
Detail panel header”Pipeline""Routine”
Agent system prompt[AVAILABLE PIPELINES][AVAILABLE ROUTINES]
CLI primary namecrewship pipeline ...crewship routine ... (pipeline alias kept)
Bundle filenamepipeline-<slug>-bundle.jsonroutine-<slug>-bundle.json

​
What stays the same (backwards-compatible)

Backend surfaceIdentifier
Database tablepipelines (unchanged)
Go packageinternal/pipeline (unchanged)
HTTP routes/api/v1/.../pipelines/* (unchanged)
Sidecar routeslocalhost:9119/pipelines/* (unchanged)
Journal entry typespipeline.run.*, pipeline.step.* (unchanged)
WS event typespipeline.run.*, pipeline.step.* (unchanged)
CLI aliascrewship pipeline X still works (routine is preferred)
Database migrationsv78–v82 apply once on first boot; unchanged
The naming split mirrors GitHub’s pattern: their underlying table is issue_events but the UI says “Issues”. Backend stability + user-facing rename, no breaking change.

​
What’s new beyond the rename

Beyond the rename itself, the same release shipped the following net-new capabilities:

​
Triggers

  • Cron schedules — recurring auto-fire on cron expression. New tables: pipeline_schedules (v80). New CLI: crewship routine schedules ....
  • Webhooks — token-addressed event triggers with optional HMAC signing, per-token rate limiting. New tables: pipeline_webhooks (v82), pipeline_run_idempotency (v81). New CLI: crewship routine webhooks ....

​
Versioning

  • Immutable history + rollback — every save creates a new pipeline_versions row (v79). crewship routine versions <slug>, crewship routine rollback <slug> --to N.

​
Step types

  • 6 step types instead of MVP’s 2: agent_run, call_pipeline, http, code, wait, transform. See Step types reference.
  • DAG with needs[] — independent steps execute in parallel.
  • Conditional if — any step can carry a template-rendered boolean to skip itself.

​
HITL

  • Waitpoints — DB-backed approval primitive (v79). wait step of kind: approval parks the run goroutine on a token; approve via CLI (crewship routine waitpoints approve <token>), UI inbox, or API.

​
UX

  • /routines page — dedicated full-management surface as a clone of /orchestration. Three-column layout with filter sidebar, 4 main tabs, and a 7-tab detail panel.
  • + New routine dialog — JSON DSL editor with 3 starter templates and Test & Save flow.
  • Live waterfall — Runs sub-tab subscribes to step-level WebSocket events.

​
Tooling

  • crewship routine validate file.json — offline DSL parse + validate, exit 0/1 for CI gates.
  • crewship routine watch <slug> — live event stream, ANSI-coloured by default, --json for jq pipelines.
  • crewship routine logs <run_id> --slug X — one-shot post-mortem dump for a specific run.
  • crewship routine doctor <slug> — preflight ✓/⚠/✗ checklist that catches blind alleys (missing crew provisioning, agent slug typo, missing credential, contradictory validation gate, cost cap too tight) before run. CI-friendly via -f json + non-zero exit on any FAIL.
  • crewship routine bench <slug> — N-runs variance characterisation: pass-rate + cost mean/p95/max + duration p50/p95/max + fail-reason breakdown. Headline verdict (PRODUCTION_READY/FLAKY/UNRELIABLE/BROKEN).
  • crewship routine records <slug> — list runs from the indexed pipeline_runs projection (v83). Filterable by status, much faster than runs for large run history.
  • crewship eval scenarios / eval compare / eval baseline — cross-tier consistency suite for routines whose slug starts with eval-. Save baselines locally, diff in CI on every PR. See eval CLI.
  • --tier-override flag on routine run (and on bench, eval scenarios, eval compare, eval baseline) — force every agent_run step onto a specific tier without DSL changes. Step-level model_override still wins.
  • JSON Schema at schemas/routine.v1.json — IDE autocomplete + inline validation when authoring DSL files.

​
Security

  • HMAC save_token — /test_run returns a signed token bound to (workspace, definition_hash, user_id, ts). /save verifies the token instead of trusting last_test_run_at body claims. Closes the test-gate spoofing surface flagged in the original review.
  • CheckRedirect on HTTP step — egress allowlist re-validated on every 3xx hop, blocking SSRF via redirect chains.
  • JOIN on workspace_id in agent slug resolution — eliminates cross-workspace agent execution surface.

​
Migration paths

​
”I’m using the CLI”

Nothing to change. crewship pipeline list keeps working as an alias for crewship routine list. New subcommands (schedules / webhooks / waitpoints / validate / watch / logs / records / doctor / bench) are only available under the routine name; the pipeline alias doesn’t expose them. Recommended: switch your scripts to routine so both old and new commands are available.

​
”I’m using the API”

Nothing to change. All /api/v1/.../pipelines/* route paths are unchanged. The new endpoints (/api/v1/workspaces/{ws}/pipelines/save for UI authoring + /api/v1/workspaces/{ws}/pipeline-schedules, /api/v1/workspaces/{ws}/pipeline-webhooks etc.) are additive. The /test_run response now includes a save_token field (omitempty); legacy clients that ignore unknown fields see no change.

​
”My agents author pipelines via the sidecar”

Nothing to change. The sidecar localhost:9119/pipelines/* paths are unchanged. The agent system prompt now says [AVAILABLE ROUTINES] instead of [AVAILABLE PIPELINES], but agents read the bracket label as a section header, not a method name — they still POST to /pipelines/save and /pipelines/{slug}/run.
If you have skills or agent personas that explicitly mention “pipeline” in their prompts, consider updating them to “routine” for user-facing consistency. The agents themselves don’t care — but humans reading agent transcripts will see “I created a pipeline” vs “I created a routine”.

​
”I have a saved bundle export from before the rename”

Bundles are forward-compatible. The export format (crewship-pipeline-bundle/v1) is unchanged. Re-import works: crewship routine import bundle.json. The bundle filename format changed (pipeline-<slug>-bundle.json → routine-<slug>-bundle.json) but the filename has no semantic meaning — only the JSON shape inside matters.

​
Database migrations

Five migrations land in this release:
VersionPurposeIdempotent?
v78pipelines table + workspaces.execution_tiers_json✓
v79pipeline_versions + pipeline_waitpoints✓
v80pipeline_schedules✓
v81pipeline_run_idempotency✓
v82pipeline_webhooks✓
All migrations apply automatically on first boot of a server with the new binary. Backup integrity is preserved — crewship backup create includes the new tables. Rollback path: restore.go for v78 returns nil (data-bearing tables can’t be partial-rolled-back); for downstream rollback use the archive snapshot from before the upgrade.
Multi-instance dev (/opt/crewship_N ports 808N) shares a single SQLite file by default. Migrations run once on first-instance boot. If you have stale instances on older binaries, restart them in version order so they don’t try to apply already-applied migrations.

​
Tier override behavior change

Routines whose steps declare complexity: "fast" (or any other tier) now actually pin the agent’s CLI to the resolved model (claude --model claude-haiku-4-5-20251001). Previously the override was silently dropped and the agent’s default model ran every step. Before (silent drop): a routine with complexity: "fast" on Tomas (default model: Sonnet) ran on Sonnet at ~$0.06/run. After (override honored): same routine runs on Haiku at ~$0.003/run. ~20× cheaper, ~3× faster. Cost dashboards may show a step-down in pipeline-attributed LLM cost after this release. That’s the expected behaviour finally landing — not a metering glitch.

​
CodeRabbit-flagged fixes carried forward

The original PR review flagged 23 inline issues. The vast majority were addressed in the original 7715ee7 sweep; the remaining items shipped with this release. See the CHANGELOG for the full list. Highlights for security/stability-conscious users:
  • DAG completion bookkeeping — failed steps no longer mark completed=true (would have let downstream branches advance with empty inputs)
  • Waitpoint lost-wakeup race — listener pre-registration eliminates the timing window
  • Webhook rate-limiter — counter increment now atomic under mutex
  • Idempotency stale-row — sweep failure no longer leaks dead run_ids through subsequent reads
  • SSRF via redirect — already covered above
  • Cross-workspace agent execution — already covered above
  • Template validation breadth — bad templates in HTTP/Wait/Code/Transform/If fields now reject at save instead of crashing at runtime

​
Things still deferred (not in this release)

FeatureStatus
Persistent run state (pipeline_runs table with status enum + checkpoint)Follow-up PR. Currently runs are in-memory; restart loses active runs. Pending waitpoints survive (DB-backed); active step execution does not.
Distributed scheduler with leader electionFollow-up. Single-instance only — running multiple replicas would double-fire schedules.
Monaco DSL editor write mode in UIFollow-up. UI is read-only; authoring works through CLI / agent / API.
NL → cron expression converterFoundation PRD scope; not in MVP.
Errors fingerprinting + bulk replayFoundation PRD scope.
Auto-disable after N consecutive failuresFoundation PRD scope.
Inbound MCP server exposing routines as MCP toolsFuture.
Cross-adapter tier swap (Claude → Gemini)Same-provider model swap works (Haiku → Opus). Cross-provider adapter constants need shorthand→constant mapping.