The API no longer waits for the cycle, and nothing inside its release rolls a Container out.
Before is the first batch of 16 Nov 2026 (api-v0.3.167, 18m13s). After is the fourth (api-v0.3.170, 6m12s), run with engine 0.56.1 after every lane of the day had merged. Each figure comes from the batch's own summary line or the Actions run it names, and the two marked by design hold because of the merged declarations.
Six lines, one plan, and the API released beside the cuts.
The command reads every line against its newest release, decides which to cut, and writes one plan: tag lines, pin pull requests and the API release, each step waiting only on what it needs.
| Line | Repository and branch | Cut when | What the cut produces | Who pins it |
|---|---|---|---|---|
| framework | framework, main | main holds commits past its newest release | a TypeScript and a Go package at one version | collections (catalogue entry and Go requires) |
| engine | cloud, main | main holds commits past its newest release | cloud-cli-vX.Y.Z: the oc binary and the engine image that release.yml runs | salesforce and collections (cloud/oc.lock) |
| evidence | cloud, main | commits past its release that are not under ignore | the sealed evidence bundle | nobody; sealed and attested only |
| ops | salesforce, main | commits past its release that are not under ignore | the salesforce ops image for oc salesforce | collections (cloud/salesforce.lock) |
| connector | salesforce, main | commits past its release, or its pins move | connector-vX.Y.Z: the Container image, rolled out on its own line | the API binds the newest healthy one at release time |
| api | collections, develop | every batch | api-vX.Y.Z: three Workers and the migrate Container on staging | none; it is the target |
--version line=X.Y.Z names the version.A step starts once the steps it waits on have passed. Steps with no such wait run together, so the engine cut, the ops cut and the API release shared wave 1.
Never two steps that deploy the same Worker. The API release and the connector rebind lock the three staging Workers, so they never overlap.
A failure starts no further step. Running the command again resumes from the step that stopped, re-planned on the newest heads.
Lock-only pin moves release the green head first. When the pins change only lock files and develop's head is green on trusted CI, the head is released at once beside the cycle; the pin pull request lands after it. A code change in the pins, a head not yet green or --serial keeps the pins first.
| Wave | Step | Waits on |
|---|---|---|
| 1 | engine: tag cloud-cli-v0.56.2 once ci.yml passes on main | nothing |
| 1 | ops: tag ops-v0.25.3 once ci.yml's gate job passes | nothing |
| 1 | api: release develop's head as api-v0.3.170 before its pins land | nothing |
| 2 | salesforce: one pull request moving engine 0.56.1 → 0.56.2, merged on green | engine |
| 3 | connector: release connector-v0.26.3 from that merge | salesforce pin |
| 4 | api: roll the connector Worker out on the same head, binding connector-v0.26.3 | connector, api |
| 5 | collections: one pull request moving engine and ops, merged on green | engine, ops, both api steps |
The API target is met; the cycle is now a serial chain behind it.
Each batch below was run from this laptop with the engine the collections lock pinned at the time. The third ran with an engine that lacked the own-line-first rule, and shows what the rule removes.
| Stage | Measured | Inside it | On the API path |
|---|---|---|---|
| API live: bundle, then release.yml | 6m12s | bundle 2m12s · release 2m59s (gate 14s, rollout 55s, connector Container kept) | yes |
| Engine cut | 7m21s | ci.yml 3m22s (the push run, already running) · release-cli.yml 3m36s | no, beside it |
| Salesforce pin pull request | 1m13s | opened, merged four seconds later | no |
| Connector line | 8m12s | ci.yml 1m43s · bundle 3m01s · release 3m12s, its own Container rollout 1m30s | no |
| Connector rebind on the API head (api-v0.3.171) | 4m51s | bundle rebuilt 1m31s · release 2m51s, rollout 23s, no API Worker deploy | no |
| Collections pin pull request | 1m08s | engine 0.56.2 and ops 0.25.3, merged four seconds after opening | no |
| Cycle complete | 23m27s | engine → salesforce pin → connector → rebind → collections pin |
| Cost on the API path in batch 1 | Then | Batch 4 | Removed by |
|---|---|---|---|
| Plan, ops gate, tag, pin pull request merge before the API could start | 2m36s | 0 | cloud #950: lock-only pins no longer gate the API |
| Trusted CI queued behind the previous develop push | 2m02s | 0 | cloud #950: the pin commit's CI runs beside the release |
| Trusted CI on a pin-only commit, every tier | 4m01s | 0 | collections #2706: pins class runs static and bundle tiers only |
| Product bundle: stale early run, queue, then the release's own | 2m28s | 2m12s | cloud #949 cancels the stale run; collections #2707 predicts the connector the release will bind |
| Connector Container rollout inside the API release | 3m49s | 0 | cloud #951, #954 and salesforce #376: the connector line rolls its own Container out |
| Guard, plan, Workers, health, journal | 3m14s | 2m59s | health no longer waits for the maintenance tick once no Container restarts |
A release runs its target and everything the target requires.
Trusted CI publishes an attested bundle for every develop commit. The release seals it, hands it to cloud's release.yml, and the engine walks the deployable's graph from the target backwards. A node outside that closure never runs, which is what the guard in cloud #955 now refuses for anything but a health gate.
oc release probe.| Node | Kind | What it does | Batch 4 |
|---|---|---|---|
api.store, api.connector-store | store | verifies every Secrets Store reference the Worker binds; a release never creates one | seconds |
api.image | image | copies the attested migrate image to Cloudflare under the release's tag | seconds |
api.migrate-worker | worker | deploys the migrate Worker, whose Container rolls out with it | seconds |
api.migration | migration | asks the ledger what is applied; when nothing is pending no workflow starts, otherwise it waits for the Container on the new image and runs the files | nothing pending |
api.connector-worker | worker | deploys the connector Worker and pins the Container image the bundle names | kept |
api.worker | worker | uploads the API Worker without traffic, proves it at 0% through a version header, then moves all traffic to it | about 30s |
api.worker-health | health | the rollout's gate: /ready pinned to the new version, 90s budget, so a failing version switches back fast | seconds |
api.health | health | the target: both Workers ready, every box linked and maintained, 900s budget for a cold start | 9s |
One command for the batch, and three ways to narrow it.
Every variant runs the same plan; the flags decide how much of the cycle it takes with it. Nothing runs on the laptop but gh: the workflows do the work, and the command follows them.
| Command | What it does | When |
|---|---|---|
oc release staging --batch --no-tail --wait | the full cycle: cuts every line that moved, merges each pin pull request as it passes, answers the one question, waits for trusted CI, releases develop's head | the routine release; exits in seconds when nothing moved |
oc release staging --batch --dry-run | prints the cycle and the release with every blocked step marked and numbers what to fix; dispatches nothing | before every batch |
oc release staging --no-upstream --yes --wait --no-tail | skips the cycle and releases develop's green head alone, reusing the early bundle trusted CI already published | the hotfix lane; api-v0.3.163 ran release.yml in 2m59s |
oc release staging --batch --only engine --version engine=X.Y.Z | cuts one line in isolation at a named version; no pin pull request opens | a sealer floor raise, which needs the engine at exactly that version first |
oc release probe | runs the checkout's health gate against staging without releasing | after a rollback, or to read the connector's link state |
cloud/oc.lock pins (0.56.2 today). Batch 3 ran with 0.55.0 and waited for the whole cycle, 33m32s.Seven steps in five waves, then what is left to decide.
Pick a wave to see its steps as batch 4 ran them, each with what ran it, its measured time and the condition that closed it. The last picker holds the lanes not yet opened.
What can still go wrong.
Ordered by severity. Each risk states the number behind it, what it means for the business, and where the evidence lives. Pick a severity to narrow the list.
Every change, what, why, how and where.
Every pull request of the release-speed day, newest first, each linking to its pull request and naming the release that carries it. The tabs pick a part of the flow; each entry's own tabs switch between what changed, why, how and where.