How it works

What actually happens between a request and a release.

No magic and no black box. Seven steps, in the order they run, with the files and checks each one leaves behind. Everything here is installed in your own accounts and stays there after we leave.

  1. Step 1 of 7

    You keep your accounts

    Salesforce sandboxes and production, GitHub, your AI subscriptions: all of it stays in your name. We never hold production credentials. The access we get is read-only and time-boxed, and it is revoked at handoff.

    During the setup we work from an invite you can withdraw at any moment. Production is never on the list.

    You keep your accountsA bracket labelled 'yours' holds three boxes: Salesforce with sandboxes and production, GitHub, and your AI subscriptions. Outside the bracket, a box labelled 'us' has dashed arrows marked read-only and time-boxed to the sandboxes and to GitHub. Production is marked with a cross: no access, ever. The AI subscriptions are marked: in your name.YoursSandboxesSalesforceProductionSalesforceGitHubyour repoAI subscriptionsin your name, paid by youUsread-only,time-boxedno access, ever
  2. Step 2 of 7

    A builder model writes the change in a branch

    It starts from a written brief that names the files it may touch and the tests it must pass. The builder works in a branch, never on main, never in production. A file the brief did not name is out of bounds, and the reviewers check for exactly that.

    The brief is a plain text file in your repo. A person writes it; it is the contract the review is judged against.

    A builder model writes the change in a branchA brief file lists the files the builder may touch and the tests it must pass. An arrow leads to a builder model, which commits to a branch drawn as three commits off the main line. The main line is marked: never on main. The branch ends at the commit that will be reviewed.Brieffiles it may touchtests it must passwritten by a personBuilderone modelmain: never written to by the builderthe commit to reviewbranch
  3. Step 3 of 7

    Two reviewer models review the exact commit

    Two reviewers from different model families each read the same commit, not a summary of it. Each writes a verdict file: GO, GO-WITH-CHANGES or NO-GO, a findings list, and the hash of the tree it reviewed. The builder never reviews its own build.

    A verdict without a tree hash is not a verdict. The hash is what ties the review to the bytes.

    Two reviewer models review the exact commitOne commit with hash a91f3c has two arrows out, to reviewer A from model family 1 and reviewer B from model family 2. Each reviewer writes a verdict file: reviewer A writes GO with hash a91f3c, reviewer B writes GO-WITH-CHANGES with hash a91f3c and two findings. A note reads: the builder is not one of them.Commita91f3cReviewer Amodel family 1Reviewer Bmodel family 2verdict-a.mdGO0 findingshash a91f3cverdict-b.mdGO-WITH-CHANGES2 findingshash a91f3cthe builder is not one of them
  4. Step 4 of 7

    The release gate refuses unless the hashes agree

    A script, not a person, reads the verdicts. It releases only when two different reviewers passed the same hash, and that hash equals the tree about to deploy, right now. Edit one line after review and it refuses. Reuse an old verdict and it refuses.

    The gate's seven refusals are listed on the playbook page, and the free Release Pack includes the script.

    The release gate refuses unless the hashes agreeThree inputs enter a box labelled release gate: the hash in verdict A, the hash in verdict B, and the hash of the tree about to deploy, computed now. If all three are equal and neither verdict is NO-GO, the gate releases. Otherwise it refuses. A side path shows one line edited after review: the tree hash becomes b07e21, no longer equal, and the gate refuses.a91f3cverdict Aa91f3cverdict Ba91f3ctree to deploy, hashed nowRelease gatesame hash?two reviewers?no NO-GO?ReleaseRefuseb07e21edit one line after review:not equal → refuse, every time
  5. Step 5 of 7

    Check-only validation, a human go, then the validated job

    A check-only validation runs in the target org and deploys nothing. A person reads the result and presses go, in writing, for this exact release. The validated job is then deployed as-is; no fresh package is built. A post-deploy check confirms the feature is actually on.

    Pressing go is a rule the team follows, not something the script can do. We say so because it matters.

    Check-only validation, a human go, then the validated jobFour boxes in a row. Check-only validation in the target org, which deploys nothing. A person presses go, in writing, for this exact release. The validated job is deployed as-is; no fresh package is built. A post-deploy check confirms the feature is actually on. Under the last box a note reads: a green deploy is not a live feature.Check-only validationin the target orgdeploys nothingA person says goin writingfor this releaseDeploythe validated jobno new packagePost-deploy checkis the feature actually on?a green deploy is not a live feature:flows active, switches where you set them
  6. Step 6 of 7

    Shared memory: one place, every agent

    Rules, lessons and org facts live as plain files in your repo. Every agent reads them before it works. A lesson written once, by anyone, applies to all agents from then on. There is no second copy to drift.

    Teach one agent, every agent learns. A correction becomes a line in a file, and the next run reads the file.

    Shared memory: plain files in your repo, read by every agentA folder in your repo holds three files: rules, lessons and org facts. Three agents, the builder, a reviewer and the post-deploy check, each have an arrow from the folder marked 'reads first'. One dashed arrow goes back from a person to the lessons file, marked 'written once'. A note says: applies to all agents from then on.your repo / memoryrules.mdlessons.mdorg-facts.mdreads firstBuilderReviewerPost-deploy checkA personwrites alesson onceevery agent reads it on its next run
  7. Step 7 of 7

    Budgets: a per-person daily limit, checked before each call

    Each person on the team has a daily spend limit on the AI plans your company holds. The limit is checked before each model call, not after the bill arrives. When it is reached, the call does not go out and the person is told why.

    Costs land on your own AI plans and vary with usage. The limit is the thing that keeps a bad day small.

    Budgets: a per-person daily limit, checked before each callA person's request goes to a box labelled budget check, which reads that person's daily limit and today's spend. If under the limit, the call goes out to the model on the company's AI plan. If over, the call does not go out and the person is told why. A note reads: checked before the call, not after the bill.A personone requestBudget checkthis person's daily limittoday's spend so farbefore the call goes outModel callon your planStoppedand told whyunder the limit: the call goes outat the limit: nothing goes out until tomorrow, or a lead raises itcosts land on your own AI plans and vary with usage

The gate, in full

Seven cases. The script refuses in every one.

The release gate's seven refusalsA table with seven rows. For each case the gate script refuses to release. 1. Fewer than two landed verdicts: Refuses. One review is not a review. 2. Both verdicts from the same reviewer: Refuses. Two different reviewers, or nothing. 3. Either verdict is NO-GO: Refuses. Fix, freeze again, review again. 4. A verdict has no tree hash: Refuses. A review that names no bytes reviewed nothing. 5. The hash moved during a review: Refuses. That review is void. 6. The frozen hash is not the tree about to deploy: Refuses. Edit one line after review and this fires. 7. A name, path or secret that must not ship: Refuses. The scrub runs on every changed file.WHENTHE GATE01Fewer than two landed verdictsRefuses. One review is not a review.02Both verdicts from the same reviewerRefuses. Two different reviewers, or nothing.03Either verdict is NO-GORefuses. Fix, freeze again, review again.04A verdict has no tree hashRefuses. A review that names no bytes reviewed nothing.05The hash moved during a reviewRefuses. That review is void.06The frozen hash is not the tree about to deployRefuses. Edit one line after review and this fires.07A name, path or secret that must not shipRefuses. The scrub runs on every changed file.
The same list is written out on the playbook page

What can still go wrong

Honest limits. We would rather you heard them from us.

  • Reviews check code against code, not against your orgA reviewer reads the commit. It cannot see your live picklist values, permission sets, data shapes or active automations. A person still checks those against the live org before the go.
  • A green deploy is not a live featureA deploy can succeed and leave a flow inactive or a switch in the wrong position. That is why the post-deploy check exists, and why it is a named step, not a hope.
  • Models have outagesWhen a model is down or refuses, the pipeline fails over to another model in the same role. Reviewer independence survives the failover: the builder still never reviews its own build.
  • Costs vary with usageThe models run on your own AI plans. A busy sprint costs more than a quiet one. The per-person daily limit caps the damage; it does not make usage free.

What we are not

Four things this is not.

Not a DevOps tool

We sit in front of DevOps Center, Gearset or Copado. They move and test metadata. The harness decides who may build, who must review, and when a person presses go.

No production access

We never hold production credentials. Nothing automated reaches production, during the setup or after it.

No hosting

Nothing runs on our servers. The scripts, skills and memory files live in your repo and run in your accounts.

No retainer

No subscription, no usage fee, no minimum. After handoff you owe us nothing unless you book an advice session.

What you own on day one

Everything. In your repo. In writing.

After handoff your team owns every script, prompt, skill, rule and memory file, in your repo, under your accounts. We keep no copy. You get a written deletion confirmation from us within seven days.

  • The release gate script and its tests
  • The two reviewer skill files and the verdict template
  • The brief template and every brief written during the setup
  • The rules, lessons and org-facts files
  • The budget check and each person's limit
  • The handoff document and the recording of the training
  • A written deletion confirmation within seven days

Start with the audit, or ask the question you still have.

The audit tells you which of these steps your team is missing, in writing, within 48 hours. A question gets a plain answer from a person.