Status
This page says what is checked by machinery on every push and what is only written down. It wins any disagreement with the rest of the site.
Checked on every push
.github/workflows/hugo.yml
runs these on every push to main.
| Check | Job |
|---|---|
| The Hugo build succeeds | build |
| The Hugo build fails on any warning | build |
| Every internal link in the rendered site resolves | build |
| Every relative link in the markdown sources resolves | build |
| The desk’s prompt, its page and its generated manifest still agree | build |
| The prose rule over the lesson pages, no em dashes, colons or semicolons | build |
| The watcher’s cap, counted on any pull request carrying the reconcile marker | pr |
| The site deploys to GitHub Pages | deploy |
The site container image builds and pushes to ghcr.io/intentius/waterpark | image |
The runner image compose pulls, ghcr.io/intentius/waterpark-runner, is
not built by this workflow.
Verified by hand, dated
| Date | What |
|---|---|
| 2026-08-23 | just up, then just register, then just runner. The runner came up connected and listed online. |
| 2026-08-23, 2026-08-24 | The start skill walked end to end, an agent playing the student. |
| 2026-09-05 | just up on a fresh volume, just register, just runner. The runner came up online; a real inference credential went in through the web UI. |
| 2026-09-05 | Fountain lesson 1 on the compose stack. fountain apply -f created the three objects, the secrets endpoint returned keys with no value field, the conversation replied, and the second apply updated in place. The sandbox had STAGE=dev; a prompt naming the shell read it back. Two things did not match the page and are filed as issue 37
: the runtime refuses the manifest’s model id and answers on the account default, and the page’s prompt does not make the agent read STAGE. |
| 2026-09-05 | Floci as a Terraform target, plan phase 0. Upstream 2.0.1 converged, reported drift and imported, but never returned a role’s permissions boundary on read and never populated the iam:PermissionsBoundary condition key. A patched build of the lex00/floci fork (ghcr.io/lex00/floci:iam-boundary) passes all four facts, and the compose stack now pulls it. Both runs are recorded in project/upstream.md
. |
| 2026-09-05 | IAM lessons 1 to 5 written against checkpoint/i0 to checkpoint/i5 and taken end to end by an agent playing the student from a fresh clone with only the pages, the skills and the tags. Lessons 3, 4 and 5 passed their done-when as written, 1 and 2 failed on a checkpoint compare that could not be silent. Nineteen stumbles filed and fixed, the changed steps re-run clean. |
| 2026-09-05 | IAM lessons 6 to 8 written against checkpoint/i5 to checkpoint/i8 and taken end to end by an agent playing the student from a fresh clone. All three done-whens passed as written and every promised compare was silent. Sixteen stumbles filed, fifteen fixed in text, one left as a script finding. The access workflow’s PR job and apply job each ran for real on PR 71 and its merge, the apply job matching the approved plan digest and applying. |
| 2026-09-06 | Fountain lesson 3, the egress allowlist, written and taken end to end by an agent playing the student from a fresh clone. The done-when passed as written. The class stack’s runner holds no egress policy, so Fountain refuses a limited environment rather than running it open, and the lesson teaches that refusal. It needs no inference key, since the refusal comes before the model is called. |
| 2026-09-12 | IAM lesson 13, the watcher, written and taken end to end by an agent playing the student from a fresh clone. All five done-when clauses passed as written. The three projections each answered from the account rather than from a file, including the access review naming the unused-access analyzer it could not reach. A break-glass grant left past its expiry on main produced a sweep pull request whose diff removes the block, beside a reconcile pull request from Fountain lesson 9 whose diff is empty, which is the distinction the lesson is for. The pr job’s cap step ran on the sweep’s own pull request and reported cap 5, open watcher pull requests 1, so it counts both markers against the one number in access/baseline. And GET /api/search returned the lesson 12 request out of the conversation record. Two bugs found before the lesson could be written, both in code that had shipped green. access/scripts/break-glass sweep had never removed anything since lesson 10, because it changed directory before calling helpers that resolve from the access root, so its guard matched nothing and it committed an empty tree and opened a pull request with no diff while reporting success, and a stale branch of the same name made those empty pull requests read as adding the grant they were meant to retire. And the cap counted one marker, so a watcher could hold five reconcile pull requests and five sweeps while the constant says five. Nine stumbles filed. The load-bearing ones are that the break-glass marker is parsed rather than decorative, that the two dates in the lesson are the only pair both already expired and inside the two hour ceiling, and that the sweep branches from wherever you are standing. |
| 2026-09-12 | IAM lesson 12, the concierge, written and taken end to end by an agent playing the student from a fresh clone, two model turns. All five done-when clauses passed as written. A request in plain words became a pull request changing exactly one file, whose body opens by naming the requester and marking the identity an unverified claim in those words, since nothing in a conversation attests who is typing. A requester naming a team that appears in no principal’s teams and nowhere in access/codeowners.map was refused with the enrolment path named and no pull request opened for the access. And the same edit typed by hand went through the identical pr job and rendered a byte-identical access delta, both hashing to 1b8f34d4, which is prescription 13 checked rather than asserted. Ten stumbles filed. The one that mattered is that the lesson’s closing claim is false on this repository as configured. main carries no branch protection, so CODEOWNERS is advisory and the desk’s token, which holds contents write, can push straight to main where the apply job picks it up. The estate has declared that gate in access/github since lesson 6 and applies it against a real code host rather than from a laptop, so on this teaching copy it is a file rather than a fact, and the lesson now says so and has the student check it instead of believing the credential table. |
| 2026-09-12 | Fountain lesson 9, the propose loop ambient, written and taken end to end twice by an agent playing the student from a fresh clone, three model turns each. All six done-when clauses hold, five of them in the second student run and the last two confirmed by hand afterwards. A schedule ran the watch in the teammate’s own thread, its aws-drift block named the drift planted by hand, reconcile filed one pull request per drifted resource with no file change in any of them, a pull request closed unmerged came back as declined on the next run and stayed declined, labelling it desk:reconsider let the same finding be filed again, and a pull request carrying the reconcile marker failed its own pr job with cap 1, open reconcile PRs 3, which is decision 40 holding without the watcher’s cooperation. Three things had to be built first and each was a claim the repo had already made. The watcher’s cap was enforced nowhere, though access/baseline had said since lesson 7 that the PR job counted them. A decline did not stick, so a watcher would refile what a person had closed. And state lived in whatever directory last applied, so a scheduled run on a fresh sandbox would have planned to create an estate that already exists, which access/backends/backend.floci.tf fixes by putting state in a bucket. Twelve stumbles across the two runs. The one that cost the most was the desk cloning main while the decline check sat on an unmerged branch, so the sandbox ran code older than the lesson describing it, which is a rule about this repo rather than a defect. |
| 2026-09-11 | Fountain lesson 8, the propose loop interactive, written and taken end to end by an agent playing the student from a fresh clone, seven model turns on the class stack. All five done-when clauses passed as written. The desk planned the estate the repo declares at sha256:6778be0e, which is the fourth machine to compute that digest, applied it on APPROVE and the account read back by hand held the four roles and six policies the delta named. A grant asked for in plain words came back as a diff with the requester’s own reason carried into the file. A policy deleted in the account between the plan and the approval made the desk refuse its own plan as stale, naming the grant that had gone and applying nothing, and a re-plan restored it alongside the requested grant. A boundary change came back refused. The student’s parts table named propose as the one part that differs from Mend and answered the closing question against the desk, on the grounds that the stale refusal is the desk stopping itself rather than something stopping it. Five stumbles filed. The two that mattered were the skill’s terminal path, which said how to send a message to the desk and never how to read the reply or where to find the agent id, both now written out. One run reported seventeen resources added for an eighteen resource plan while the estate came out correct, so both texts now say to read the account rather than the desk’s sentence, and the prompt says to apply once and to keep detail to what the commands printed. |
| 2026-09-11 | Fountain lesson 7, driving an agent from an app, written and taken end to end twice by an agent playing the student from a fresh clone, one model turn each. The lesson has a browser in it and an agent has none, so three of the four done-when clauses were checked in their terminal half and the fourth, the key named oauth:aws-desk that only the sign-in button produces, is unchecked and says so here. Both preflights, the desk’s three objects and its seat, the aws-state block in the conversation and the rename that just desk-check refuses all passed as written. The first run found ten stumbles. Two were real bugs. Fountain scrubs every environment value of eight bytes or more out of the record, plain env_vars as well as secrets, so the block came back carrying [REDACTED] where its region and endpoint had been, and the block no longer carries either, which project/upstream.md
records. And the page claimed an empty account reads complete false, which was only true while the mktemp bug was live, where an empty account read cleanly is complete true with an empty list. One was the author building a Docker image during the run, which filled the VM disk and made a sandbox fail terraform init and hang. The second run found five more, the load-bearing one being that the no-browser path needed an agent id nothing printed, so just desk-hire now prints the agent beside the conversation. The second round of fixes was verified by hand rather than by a third run. |
| 2026-09-11 | The AWS desk v0, direct mode against the class stack’s Floci, taken end to end. The desk read an empty account, planned the estate the repo declares, applied it on APPROVE plan-d263 after its recheck digest matched, then took a request in plain words, edited one grant into one file, planned it, and applied it on approval. The account was read back by hand afterwards and holds four roles, two buckets, the boundary and six grant policies. The digest over the bootstrap plan is sha256:6778be0e... on the author’s laptop, in a hand probe inside the runner and in the desk’s own sandbox, which is three machines and one digest. A request to widen the estate boundary came back refused with the platform path named. Four things found and fixed, all in this repo. Six of the access scripts called mktemp -t with no X’s in the template, which is a macOS spelling that a Linux sandbox rejects, so access-review died and the desk diagnosed the bug itself before falling back to direct account reads. The class runner image carried no terraform and no tflint. A sandbox runs with its own directory as HOME, so a lint plugin installed under /home/runner is one nobody finds. And the page’s block parser required a fence to start its own line, which the joined reply text does not always give it, recorded in project/upstream.md
. |
| 2026-09-10 | IAM lesson 14, approve the change not the diff, written against checkpoint/i15 to checkpoint/i14 and taken end to end by an agent playing the student from a fresh clone, on the class stack’s Floci and a second container. All five done-when clauses passed as written. The saved plan’s delta and digest read as what a reviewer approves, a hand edit to the account changes the digest with the diff naming why, Terraform refuses a stale saved plan on its own, one in-place change planned from two emulators gives one digest where the digest from before lesson 11’s fix gave two, and a rename renders a replacements section the PR job now labels. Five stumbles filed, two the saved plans and the step 3 grant reaching the compare, three wording, all fixed. |
| 2026-09-10 | IAM lesson 15, adopt in place, written against checkpoint/i11 to checkpoint/i15 and taken end to end by an agent playing the student from a fresh clone, on the class stack’s own Floci. All five done-when clauses passed as written and the checkpoint compare was silent. A role, a bucket and an open security group made by hand come under management one file at a time with import blocks reviewed out of generated config, adopt-check passes each because the only change is the estate’s three tags and fails a file that misstates the account by naming the attribute, the rule pack refuses the adopted role as it would a written one, and a removed block leaves the group in the account. Five stumbles filed, one the check not seeing a landed import, one the adopted role’s missing team as a fifth day-two item, three wording, all fixed. |
| 2026-09-10 | Fountain lesson 11, no approval gate in Fountain, written and taken end to end by an agent playing the student from a fresh clone, two model turns on the class stack. All three done-when clauses passed as written. The runtime’s command line read off the runner mid-turn carries --allow-dangerously-skip-permissions with its permission prompt wired to stdio, a file write lands with zero request_permission messages in the record, and the destructive prompt went both ways across two runs, declined once on the model’s judgment and complied with once, deleting the sandbox’s own runtime install with the same zero in the record, which the page now teaches as the point. Seven stumbles filed, all fixed. |
| 2026-09-10 | Fountain lesson 10, the self-hosted runner, written and taken end to end by an agent playing the student from a fresh clone, three model turns on the class stack. All four done-when clauses passed as written. The runner reads as one process and one user holding the API key, a limited environment refuses with provider: runner, a second sandbox lists its sibling and both directories are one owner’s, a stopped container gives a new run a 409 naming the fix and a parked conversation a runner_offline failure, and the restart brings the same conversation back. Five stumbles filed, all wording, all fixed. |
| 2026-09-10 | Fountain lesson 6, schedules, written and taken end to end by an agent playing the student from a fresh clone, three model turns on the class stack. All four done-when clauses passed as written and every status code, count and label matched. @reboot is refused, a schedule made before its teammate exists records agent is not on the team when run, an every-minute schedule fires on the whole minute into the teammate’s own thread with the scheduler’s stamp, disabled it goes quiet and still runs by hand, a one-off opens a fresh conversation with no channel, and removing the teammate deletes its schedules. Four stumbles filed, all wording, all fixed. |
| 2026-09-10 | Fountain lesson 5, the team, written and taken end to end by an agent playing the student from a fresh clone, three model turns on the class stack. All four done-when clauses passed as written. One call adds an agent to the team and opens one conversation on the reserved channel, the roster and the channel-filtered conversation list name the same id, the stream carries the turn under the agent id, a second message mid-turn is a 400, a second add is a 200 with the same conversation, the teammate sleeps and wakes, and a remove unbinds and keeps the record. Found and recorded in upstream: with the runner stopped a message is accepted with a 202 and its turn fails at once with runner_offline and is never redelivered. Five stumbles filed, one that finding, four wording, all fixed. |
| 2026-09-10 | Fountain lesson 4, credentials and vaults, written and taken end to end by an agent playing the student from a fresh clone, four model turns on the class stack. All three done-when clauses passed as written. The merged secret is read off the runner’s disk, where each sandbox’s .env holds its own vault’s value beside the inference token, a vault outside the allowlist is refused with a 422 before anything is provisioned, the Claude runtime declines to print a credential-shaped variable, and an emptied allowlist changes later conversations and not the parked one. Two things found and recorded in upstream: a vault name in allowed_vault_ids is a 500, and a sandbox terminated while parked keeps its directory and its .env. Five stumbles filed, one a wait the page now asks for, four wording, all fixed. |
| 2026-09-10 | IAM lesson 11, offboard and the access review, written against checkpoint/i10 to checkpoint/i11 and taken end to end by an agent playing the student from a fresh clone. All five done-when clauses passed as written and the checkpoint compare was silent. whocan, expiring and the review read the account and never a file, a policy attached by hand appears in them and not in the drift watch, the course author leaves with validate and grep as the proof, and the desk operator leaves and returns with a five-resource apply and a NoSuchEntity read. Five stumbles filed, one the watch’s blindness to an extra attachment which the page now teaches, four wording, all fixed. The apply job on main had refused lesson 9’s merge, the first in-place update the pipeline saw, because the digest carried a role’s create_date and unique_id from the PR job’s container; the digest now reduces an update to what changed, proven against two containers. |
| 2026-09-10 | IAM lesson 10, break-glass, written against checkpoint/i9 to checkpoint/i10 and taken end to end by an agent playing the student from a fresh clone. All five done-when clauses passed as written and the checkpoint compare was silent. A grant with a granted_at applies as one policy carrying its DateLessThan expiry and tags, the script, the rule pack and plan each refuse a three-hour grant, the watch reports the expired leftover with the sweep deliberately unrun, and the revoke puts the file and the account back. Floci evaluates no condition on an allow, so the drill where the access works and then ends is live only, recorded in project/upstream.md
. Four stumbles filed, one a missing terraform init and three wording, all fixed. |
| 2026-09-10 | IAM lesson 9, federation trust, written against checkpoint/i8 to checkpoint/i9 and taken end to end by an agent playing the student from a fresh clone. All five done-when clauses passed as written and the checkpoint compare was silent. site-publisher federates through the GitHub Actions anchor with the github-pages environment as its exact subject, the two trust rules and their fixtures run in the check stack, a hand-edited trust policy and a widened anchor both page, and the rotation check reads the account for static secrets on the drift cron. Floci’s AssumeRoleWithWebIdentity mints credentials for a forged token, and the lesson runs that call on purpose. Five stumbles filed, four wording and one a missing apply, all fixed. |
| 2026-09-06 | Fountain lesson 2, the sandbox lifecycle, written and taken end to end by an agent playing the student from a fresh clone. All three done-when clauses passed as written and every command ran unmodified. The class stack now sets Fountain’s idle bound to two minutes, so the student watched the sandbox park two minutes and fifty-five seconds after its last turn, wake by reattach onto the same sandbox id with a file still on its disk, and lose the directory on terminate while the transcript stayed. Thirteen stumbles filed, all wording, all fixed. Like lesson 3 it needs no inference key. |
Written, not yet verified
Every one of the 26 lesson bodies is written and student-run. What is left is the media, the live layer and the back office, which are plan phases 5 to 7. Written and student-run: Fountain lessons 1 to 11, and IAM lessons 1 to 15.
just todoscounts 41 today.The access repo exists under
access/for what IAM lessons 1 to 8 need (layout, personas, the check stack, the Floci deploy, the boundary, the PR and apply jobs, drift and reconcile, the satellite and the double refusal), taggedcheckpoint/i0tocheckpoint/i8. The proofs are live only, since Access Analyzer is a stub on Floci, and everything lessons 9 onward need is not built.IAM lesson 10’s student run passed green over a sweep that removed nothing, because its done-when exercised the dry run and the watch’s report and never the
--openpath. The sweep is fixed, and lesson 10’s claims are worth re-reading against it.mainon this repository carries no branch protection, so the required check and the required code owner review thataccess/githubdeclares are a file rather than a fact here. IAM lesson 12’s closing claim depends on that gate and says so.The AWS desk is code, under
desk/, in both modes. Direct mode plans and applies with a credential of its own and takes anAPPROVEmessage as the approval. Repo mode holds a GitHub token and nothing that reaches AWS, opens a pull request, and leaves the applying to the gated job. Four blocks,aws-state,aws-plan,aws-resultandaws-drift, and a three pane page at/desk/. What is not built is the proofs on the solo path, since Access Analyzer is a stub on Floci.Floci as a Terraform target was verified on 2026-09-05 (plan phase 0, recorded in project/upstream.md ). Plan and apply converge, drift reports under
-detailed-exitcode, andimportblocks work. Upstream 2.0.1 has two gaps, a role’s permissions boundary is never returned on read and theiam:PermissionsBoundarycondition key is never populated, so the compose stack runs a patched fork build,ghcr.io/lex00/floci:iam-boundary, on which all four facts pass. The self-paced path depends on that image until the fixes are upstream.
The phases that close these gaps are in project/plan.md .