this will be the cursor community post that lays out how the checks layer works and ties into the system and what it does
a zero slack CI where 1 fail halts the test, then rips apart the pr to find the exact causation of the failure while keep high speed pr testing and merging, not at human speed but AI speed.
with the end goal being:
a full decentralization of CI workflow, spread out between users using cellphones to cache the index needed to maintain the “who has what tools/libraries cached” handled over gitsync with a 1 singular CI needed to handle all repos at once pulling it’s needed tools from a toml file put into a repo. this way CI’s do not need to be custom built per-repo needs. all failures create a new rule, a new test, and a new tool. any failure of any kind is flagged for an issue ticket, flaky testing for code will no longer be needed to be done. under my in house testing i was able to drop my full suite tests solo on 1 laptop with only 1 tester from 30 minutes to 5 minutes, with the smallest being 45 seconds. all passing tests are cached, all libraries cached, all tools cached. all tests are done in a sealed container inside of a no-sudo enviroment, zero-network, all results/tools/libraries are fingerprinted with SHA hashs, and 3 testers minimum needed to validate returned work if done on a p2p style workflow.
what this will do for the future development??? i have no idea ask an engineer, what i do know, is there is absolutely no ci/cd testing suite anywhere on the market currently or any research papers using the combination i have put together in this post and nearly done building.
what this can do? bring the recursive self learning training standards from large LLM researchers into the hands of everyone. in the last month, my entire system was built by grokbot, i didn’t write a single word or write a single line of code for anything i have built, yet, every single tool and script both my teams use have been updated multiple times unprompted, every rule in the system was created by a bot catching its own failure somewhere and making rule, because i gave them trust.
in my own in house tests i was able to push over 300 pr’s per hour, fully tested, and every failure only making the tests more robust and better in every aspect.
this post is also not the fully up to date code as the development and deployment is to rapid to keep up with.
by all means, please, tear these diagrams apart and rip in to every design choice that has been made so far. being wrong is a gift, and being correct is a curse.
A short intro
This is the Checks layer in Project-001 . Code is tested in a sealed box, and a change only moves up when there is proof, not someone’s word. That includes a users own changes.
The question behind it: “how do you trust work when you can’t see the motives behind it — human or AI?” Judge the merit of the work, with proof at every step. A silent fix is still a mistake; if no one can see what broke, no one can learn from it, and no one can trust the fix.
The Lock Box holds and seals the values (counts, pins, scan rules, and, the time anchors). The CI only checks whether what it’s handed is true or false, and it never holds those values itself. It is not locked to one lock box; anyone can plug in their own. Home base is where work lands once it is built, checked, and passed. The Bot tool Dashboard watches the checks. The Library of Truth holds the information and decisions behind the work. Bee Stack is the rulebook people and agents carry; it is not the CI.
The diagram above is the whole flow. Below, that same diagram is broken into its numbered sections, one at a time. Each chunk is a small diagram for that section, then a few plain words about it. Gray boxes are stubs: they point to a piece that lives in another section. In this updated copy, section 8 (the bubble and the straw) is included from the design notes on main, and a few later rulings are marked, including things that are planned but not built yet.
1. Getting in
This is the front door. An agent writes code and runs its own checks before the push, but that local record is only a claim. The work is scanned before it leaves the machine, and scanned again when it arrives. Secrets, names, emails, IDs and paths are what the scan looks for, and the rules come from a sealed copy in the lock box. If the scan is clean, the version door checks that every pinned tool is already in the cache, with the right hash. If a tool is missing, it is fetched once, checked, scanned and stored under a fingerprint. If it can’t be fetched or the hash is wrong, the batch is parked as “wrong tools, not tested” — not a test fail — and a TOOLS ticket is opened. Protected folders and files are held and need a reason and proof. Renaming, moving, deleting or repointing one is an instant flag. Bee gets no shortcut here either.
2. The single-file line and the 5 seats
There is one line, and five seats in front of main. Seat 1 is locked: it tests a batch on testers in a frozen bubble built from cached tools, and it is never kicked out. It leaves only on pass or fail. A new push cannot cancel it. Seat 3 is the fix slot. Failed pieces ride with Seat 2’s work in labeled groups, so each pass shrinks the problem. Testers can be Bee’s machines today, and later other machines, but they all use the same sealed bubble. When main moves, waiting PRs are not all updated. An overlap check compares main’s changed files with the PR’s files and every file its suites really read. No overlap means the earlier pass still counts, and only a quick clean-merge check runs. Updates happen as late as possible, as a test copy in the bubble. The branch gets a normal merge commit only if it lands. There is no force-push and no rebase. A waste meter tracks re-runs per merge, and a growing number opens a ticket.
2b. Train seats: in front of main, land proven work
In front of main, proven work sits in train seats. The seat map lives in the lock box, starts as one seat per top-level folder, and is later reshaped by what tests really read and by merit, never by hand. A folder seat holds at most one proven change per file, and that file stays locked until the seat lands. A shared seat handles lock files and top configs, and lands alone with the full suite. Fixes use a fast-track seat, still with the same one-change-per-file lock. A wide change waits until all of its files are free, claims them at once, runs the full suite with no reuse, and lands alone; moving up after a set number of misses is planned in the idle loop. Ready seats lump-landing together is also still planned. A land is a normal merge commit, and the git host must confirm that the merge commit contains the PR. If that cannot be confirmed, the land is flagged, is not proven-good, and does not count toward the streak. Honeypot works out dependencies from the files themselves. A public Depends-On line in the PR is only a canary: missing means “read the rules” with no penalty, matching the files is ignored, and not matching is a flag. The CI never uses that line to order work.
2c. Bookmarks, the repo, and a new scan rule
Bookmarks live in the lock box. They are signed by the proof check, copied onto two or more shelves, and moved only by merit. A hand move is a flag. Landed means the quick check after landing passed. Proven-good means a full run, with two or more testers agreeing. Golden, in Honeypot today, means 7 clean daily runs in a row on two or more testers; one fail or flag resets the count. (The Nov 18 rule that golden also needs 7 passing PRs, 7 commits and 7 pin moves lives in the proof ladder note and is not built in Honeypot yet.) Bad is permanent: the fix is a new commit that earns its own marks. A daily real run happens on proven-good main even when nothing changed. Rollback is a new undo commit on top, never a force-push or a rewrite, and it goes to proven-good, or to golden if proven-good is in doubt. Undone changes go back as tickets. When a new scan rule is added, the full history is scanned once. A hit means something already leaked: flag, rotate the key, and log it. The history is never rewritten.
3. Inside Seat 1
Inside Seat 1 the work is routed from changed files to their suites, plus the checks that always run. An unknown or risky file gets a full run. Fast checks cover guards, format, lint, and full scans again, including personal info in logs, screenshots and test data. Fakes are allowed only with a marker from the lock box list. Guard code has to match the sealed copy. Tests may not sleep or call the real clock. Full tests use a fake clock, wait for events rather than time, scale timeouts to the tester’s speed, and run in a random order with the seed recorded. If a test fails, only the failed tests are rerun, same code, same seed, fresh bubble, at most three times. The first fail stays on the record. A rerun that passes is flaky, not a clean green. Warnings count as fails unless they are on the lock box allowed list. A new setting that hides warnings is a flag. Counts are checked against the lock box’s true values. The proof check is the judge: same bubble, same commit, counts and no warnings, a signed result, and two or more testers with fresh salts must match. Run time is recorded and never compared. A canary fake key is planted in a scratch copy to see if the scanner catches it. A pass can be reused only when every file the suite really read is unchanged, and the bubble, lock file, main, settings, tool versions and seed rules match, and the receipt is not older than the last daily real run. Guards, scans and canaries are never reused, and even a reused pass still runs the guards and a clean-merge check. One in ten reuses is re-run at random. A disagreement is a flag.
4. When something fails
When something fails, Seat 1 splits the failing group into two labeled groups, and at the same time splits the change’s lines into labeled pieces. A piece that will not build is kept whole, not blamed. One ticket is opened per original batch, updated at every split, and ends with the exact change, the lines and the failed test. Passing groups can go on. Remaining groups ride along with Seat 2. When a single culprit is found, it goes to the back of the line, and a fix, if one is made, cuts in at Seat 3. If the same failure repeats, a check is added to the fast checks or the pre-push, so the problem moves left. A close check asks whether the file now routes to the test that caught it.
5. Honesty and trust
The lock box holds the values the CI trusts: true counts, the quarantine list, the protected list, sealed guards, scan rules, canaries, the fakes list, and allowed warnings with expiry. The rules apply to every author, Bee included. Agents read sealed copies. They cannot edit counts or quarantine. Lowering a count needs a reason and proof. A quarantine request needs approval, lasts one week, and there are at most five active at once. Getting a test back from quarantine takes 20 passes in a row, in fresh bubbles, with random seeds, two or more testers, and the proof check. One fail resets that streak. A flag freezes the batch, and nothing merges. The emergency door is visible and needs two keys. The only thing it can do is roll the guards back to the last sealed proven-good copy. It is loud, logged, and it opens a ticket. A scan-rules rollback through the door is flagged as an emergency roll, not as a suspicious edit. A real fix still goes through the line. The door closes in one hour. There is no runner-hosted door window yet; sessions are opened through the command line.
6. Status, metrics and trust meters
Each PR is meant to have one status comment that edits itself: place in line, what passed, what failed and why, bookmarks, and flags. The change log is appended at the bottom, nothing is deleted, and it points to the rules. Posting that comment on the git host is still planned; the log already lives in the box. Metrics and meters are stored in the lock box and are built only from signed receipts. The comment and the dashboard are just views. An edit outside the box is an instant loss of trust. There are ten trust meters side by side, including honesty, work quality, tester reliability, rule following, security, helpfulness, per-area quality, a proving period for newcomers, a setup meter for tools, connectors and prompts, and Merit, which is the sum of the other nine and is never set by hand. An honesty mark caps Merit. Negative marks never fade or delete. Points count only for proven-useful work. The public view shows scores and lessons only, after a personal-info scan. Merit picking Seat 1 testers and shaping train seats is planned, and is not wired yet.
7. Pin moves
The pin list lives in the lock box. Each pin is watched, and a move is one PR per pin, with a cooling-off wait, except for security pins, which go through the fix slot right away. A shadow lane on proven-good main, in a fresh bubble, with two or more testers, must match or beat the old pin on counts, warnings, fails, speed and the canary. Today that shadow uses Seat 1 metrics; a dedicated shadow runner is planned. The new cache is built ahead, fetched once, locked by hash, scanned, and stored under the new fingerprint. Pins land alone through the shared seat in a quiet window. Rollback is a new commit that moves the pin back. The old cache and receipts are kept, so rollback is instant. A failed pin is marked bad, gets a ticket, and hits the setup meter. If a pin group fails the shadow run, it is split like Seat 1 until the one that broke it is found. A pin that is no longer used is archived with a note, never deleted.
8. The bubble runs in a container (the straw)
The bubble runs inside a container. The default launcher is rootless Podman. It is swappable, like any other connector, and nothing ever needs root. Each run gets a fresh throwaway container with the network off. The checked-out code is mounted read-only, and the tools are mounted read-only. The only way out is one writable folder, the straw, seen inside as /out. After the run, only regular files are read back, each with a size and a hash. Links and special files are refused and never followed. Then the container is discarded, so nothing from the run survives. Only what comes through the straw counts. The container image is pinned by digest from the lock box; a tag or a bare name is refused. If no image is pinned, a Podman run is parked as wrong tools, not tested. If Podman is missing, the run refuses to start, nothing is tested, and nothing is posted, unless someone makes an explicit uncontained choice. That kind of run is announced, recorded, and tagged so it can never pass itself off as contained.
What’s proven vs. still in progress
As of Nov 18, 2026, all 13 build steps are on main (as of, 6:33 PM CT), plus later fixes including. A source count on that tip finds 771 test functions in 129 test files (the suite was not re-run on this pass). Proven enough to write down, from the Nov 16 forum draft: the CI caught its own builders. Every bug went into an append-only ledger. The build itself turned up 27 bugs, and fixing them turned up more. The ledger was at 48 entries as of that draft, and counting. As of Nov 18, 2026, the ledger has 60 entries (distinct ids HB-001 through HB-060). Concrete cases already recorded there include a false “all green,” a path that could mark a commit golden without the required passing runs, and a fake receipt that could still mint golden. Those were caught by live proof, not by trust.
The Nov 18 proof-ladder decisions are written down and built into Honeypot: to reach golden, all three counts must be met since the last level (defaults 7 passing PRs, 7 commits, and 7 pin moves, changeable per team), combined with real-world timestamps, runner receipts, and three witnesses used on both the “nothing before this” side and the “anything after this is fake” side. Failed runs are counted in the record but never add toward golden. Honeypot’s current golden bookmark is still 7 clean daily runs in a row on 2 or more testers. HB-033 (merge confirmation) and HB-035 (daily run on main) are decided and built. The freshness rule for whether a daily run counts toward golden, and the recording of proof-ladder levels in the lock box, are not built yet.
Sealed boxes, signed receipts, promotion gates, and merge queues exist elsewhere (names added by Grok Bot). What’s different here, as the forum post puts it: it is built for teams where most of the “developers” are AI agents, so nobody, human or AI, gets trusted on their word alone.
Naming credit
Industry-standard names below were added afterward by Grok Bot. i built the ideas from first principles and called them things like “thingies, boxes, folders, files, and stuff.”
- sealed box with no network = hermetic / reproducible builds (name added by Grok Bot)
- signed receipts = build provenance and attestations; SLSA, in-toto, Sigstore (names added by Grok Bot)
- outside timestamps = trusted timestamping; RFC 3161, OpenTimestamps (names added by Grok Bot)
- landed / proven / golden = promotion gates and release channels (names added by Grok Bot)
- one at a time into main = merge queues (name added by Grok Bot)
- step to the left / funneling = git bisect and delta debugging (names added by Grok Bot)
- the bug ledger and tickets = append-only audit logs and auto issue tracking (names added by Grok Bot)
- also used in the notes: logical clocks, append-only hash chains, bisection (names added by Grok Bot)









