The propose loop
How work gets done on the access repo is one loop, and it was not invented here. It is what is left when you take three Fountain apps that were built separately, Mend and Rounds and dns-desk, and write down what they have in common. They differ on two axes and agree on everything else, which is the reason a course can teach one thing rather than three.
The two axes are whether a person is in the room, interactive or ambient, and what starts the work, an audit or a request. Four boxes, and every app this course reads is in one of them.
The parts
| Part | What it is | Mend | Rounds | dns-desk |
|---|---|---|---|---|
| target | The thing under management. The agent does not control it and treats its contents as untrusted. | A repo’s CI, Kubernetes, Docker and cloud config. | The same, enrolled. | The Cloudflare zones a token can see. |
| read | A deterministic read of the target. An audit yields findings sorted by fix confidence (deterministic, judgement, report-only). A request yields the current state, re-read before any apply. | An audit over ten catalogs. Silent catalogs are shown greyed. | The same, on a schedule. | dns-state, incremental per zone. |
| operator | An agent on a toolkit environment with the tool on its PATH and a credential that can read the target and nothing more. One per target. | One mender per repo. A per-repo read-only token in its own vault. | The same, holding a grant it trades for a one-hour read-only token. | One desk per token. The vault’s zone list is the blast radius. |
| plan | What the operator proposes per finding or per request. Applied means mechanical from the tool’s own diff. Proposed means a judgement call it was confident in. Skipped means it turns on intent it cannot see, with a note. Rendered as a diff. | mend-plan plus one mend-fix block per fix, independently selectable. | One cluster per file, with status. | dns-plan, a diff against the last read. |
| verify | Re-run the read after the change and before proposing. Findings gone, merge-worthy count not up, files still parse. Or re-read the target and re-plan if it moved. | Before the plan is shown. | A failed verify opens nothing. | Re-read the zone before apply. Re-plan instead of applying stale. |
| propose | The only path that writes, held by something that is not the operator. | The human’s browser with the human’s token. Every context line re-verified against the target right now. | A server minting a one-target write token for one proposal, checking policy and history, rendering the PR body from the findings. | The desk applies after an APPROVE plan-id message, with a token scoped before the conversation began. |
| rules | Constraints that hold when the prompt is ignored, enforced where the write happens. | The browser refuses a stale patch. | Never reopen what a human closed. Never a second PR for the same thing. Never outside the derived branch. At most N open. enabled: false means nothing happens. | The token’s scope. Today the approve is convention and the audit trail, until fountain#643. |
| record | Where the state lives. It is where the person who decides is already standing, with nothing to keep in sync. | The conversation. Report, plan and patch are derived from turns and blocks. | GitHub. Branch name and a marker in the PR body. The conversation holds the round report. | The conversation. Plan status is always derived, never stored. |
| refusal | An outcome, not a failure, rendered as one. | skipped, with the note. | already-open, declined, deferred, clean. A decline sticks until rounds:reconsider. | A request outside the token’s zones. |
Two things hold down every column, and they are worth reading the table twice for.
The claim and the report are one copy. The plan a person approves is the same artifact the machinery acts on, rendered once by something deterministic, so there is no summary that could be generous about what the diff does. An operator that wrote its own account of its own work would be the only witness to it.
The operator never holds a write. Whatever the mode, the thing that reads and proposes is not the thing that applies, and the credential that could apply lives somewhere the operator cannot reach. Every refusal in the last row of that table is cheap because of it. An operator that could apply would have to be trusted; one that cannot only has to be read.
The four forms
| audit-driven | request-driven | |
|---|---|---|
| interactive | Mend. Point it at a target, watch, ask for changes, take the patch or open the PR yourself. | dns-desk. Ask in words, read the plan as a diff, approve, done. |
| ambient | Rounds. Schedule, reconcile against its own past work, cap, server as propose, declines stick. | A ticket webhook starting a desk conversation. No app yet. |
The loop on IAM
IAM is the case where the target and the repo are the same object, and that collapses part of the loop in a useful way. The estate is declared in files, so a proposal is a diff rather than a description of one, and propose is a pull request in every form. Approval is the merge and never a message in a conversation, which is decision 18 and the reason there is no approve button anywhere in this course.
The applier is Terraform and never the agent. The AWS desk is the app the lessons build, and it proposes in repo mode holding nothing that can reach the account. It can apply, in direct mode, with a role bounded by the estate’s own permission boundary, and the course is careful about when it says that is acceptable.
| Form | IAM instance | Build or reuse | Lessons |
|---|---|---|---|
| interactive, request-driven | The concierge, in the desk’s form. The estate on screen. “tickets-api needs read on the receipts bucket.” One deterministic file edit. The plan is the rendered access delta. The PR is opened with a PR-only token. CODEOWNERS approve. The gated job applies. Refusals name the enrollment or escalation path. | Build. | I12 |
| ambient, audit-driven | The watcher, in Rounds’ form. Unused-access findings and expiring grants on a schedule. One PR per finding, capped, declines stick. The drift reconcile runs under the same rules. | Rounds as-is for what its catalogs cover. IAM projections added. | I7, I13 |
| interactive, audit-driven | Mend the access repo. Point Mend at the access repo. Its aws catalog reads the declared Terraform. | Mend as-is. | I3 |
| ambient, request-driven | A ticket that opens a desk conversation and tracks the PR as comments. | Later. | I12 |
The same eight parts, with this estate’s names in them. Reading this table beside the first one is the quickest way to see that nothing about IAM needed a new mechanism, only new instances of the old ones.
| Part | IAM on the water park estate | Lessons |
|---|---|---|
| target | The access repo and the live estate it owns, plus one satellite repo. | I1, I8 |
| read | The guardrails, the drift watch, the Access Analyzer proofs, the unused-access and expiring-grant projections, and the estate view. | I3, I6, I7, I11 |
| operator | A teammate on the reference Environment with a checkout, SKILL.md and a code-host token that can open PRs and nothing else. No cloud credential. | I12, F4 |
| plan | The access delta rendered by the deterministic renderer, never the model. A directed refusal for a boundary exception or an unmapped requester. | I12, I14 |
| verify | The guardrails and a terraform plan against Floci on every PR with no credential. check-no-new-access where an untrusted author cannot trigger it. The digest re-checked at apply. | I4, I6, I14 |
| propose | The PR. Merge then apply by a gated job on a protected branch. The watcher’s propose endpoint with cap and declined list. | I6, I13 |
| rules | Generated CODEOWNERS and branch protection. The apply role’s own boundary. The satellite’s iam:PermissionsBoundary condition. Rounds’ rules on reconcile and watcher (decision 28). | I5, I6, I7, I8 |
| record | Git for blame, the PR and provenance by PR and sha. The conversation for the request. Reconcile state in the code host. | I6, I7, I12 |
| refusal | The guardrail in the editor. The double refusal. An unmapped identity with the enrollment path. A reconcile PR closed unmerged. | I3, I8, I12, I7 |
What the loop asks of Fountain
The loop asks for less from a runtime than you would expect, and the Fountain course teaches each piece where it is first needed.
An environment with the tools on it and no secrets in it, which is lesson 1 and lesson 4. A vault per target, bound to the conversation when it starts, so the blast radius is decided before any work is, which is lesson 4. A teammate per target with a computer that persists, which is lesson 5. A schedule, for the ambient forms, which is lesson 6. Protocol blocks and the conversation as the record, for the interactive ones, which is lesson 7. And a sandbox provider that actually isolates, wherever the operator reads untrusted input while holding anything at all, which is lesson 10’s honesty about what a self-hosted runner does not give you.
One thing the loop does not ask for, and this is the load-bearing absence. There is no approval gate inside the runtime. Fountain has none, lesson 11 is the lesson about not having one, and propose lives outside the loop for exactly that reason. The gate is wherever the write lands, which on this estate is a protected branch and a job holding a bounded role. A loop that waited for a runtime to ask permission would be a loop with its safety in the one place nobody can audit after the fact.