recording-decisions now permits compaction and names four kinds that may never be
dropped. Applying it removed six entries and left fifty, and the ratio is worth
more than the saving.
Every entry dropped became droppable by being published. what this tool is, the
lock format, unlocked is a state, --help as the documentation surface, and the
visibility field all now live in .loom/published/guarantees.md or README.md, and
conversion-destroys lives in cart/v2 and cart.usages.md. Nothing was summarised
away — the content moved into documents somebody else can lock, which is a stronger
place for it than a log only we read. The entry names each one and where it went,
so the compaction is auditable rather than trusted.
What did not compress is the point. Almost everything left is a decline, a
measurement, a supersession, or a belief shown wrong. Declines are not recoverable
from code because the code is what they are absent from; measurements are facts
about other people's systems that no artifact here holds; supersessions are changes
of mind, which is most of what a reader wants. The log is long because the work was
exploratory, not because it was padded.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018UTxuSizozEA8yDitPuris
Publishes what the tool guarantees, on a condition that fired rather than a
decision reversed. Records the narrower promise that replaced promising the command
surface — a change to it lands on the page consumers already lock, so a rename
reports 200 the day it happens.
And records a limit found by accident that follows from a guarantee rather than
from anybody's code: an ETag is opaque, we promise never to interpret one, so two
of them can be compared and never ordered. A consumer can know two versions
disagree and can never know which is behind. Anybody building freshness on this
inherits it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018UTxuSizozEA8yDitPuris
Extracted at conversion rather than staged as it went — there was no wal this
round, and the answers lived in the daily that the rm destroys, which is precisely
the failure v2 added the rule to prevent. Worth noticing that the round which
answered a question about durable records is the one that nearly lost its own.
Records two declines with their reasons: a second orientation file, since the
publishing half varies not at all and is already adopted so it appears in the index
by itself; and .loom/externals/README.md as the path, measured rather than argued,
because check walks that directory and would report a generated file as unlocked in
every repository using the feature, forever.
Records the visibility field with the word it refuses — public or not-public and
never private, because an anonymous request tells those apart and nothing finer,
and private would claim precision the signal does not have.
And records a one-line bug whose trap is permanent: visibility was computed by one
function used both after a successful fetch and as a probe, where "no credential
configured" means opposite things. In the audit it would have silently cleared real
alarms, which is the worst failure available to a check.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018UTxuSizozEA8yDitPuris
All three prompts answered and extracted, the polad tree empty, and check clean
before the rm — which under v2 is the audit that has to happen, because the cart is
the only copy of everything in it.
The round's substance: three gaps closed at reconciliation, and in every case what
survived was a local workaround becoming specified behaviour — wal.md, which v2
named and cited our own sentence as the reason for; "a fetch you did not perform",
which had been a declined command and an insight without an implementation; and the
compaction rule, which came back with an addition we had not thought of.
Declines fixing annotating, because two of its four gaps are not gaps in the
agreement but a rule living in a local editor protocol instead of the document, and
deciding where a rule lives is not an amendment and not ours.
Records that a casting file goes stale on amendments that add no role — twice in
one day, v1 and v2 — and that the mechanism fired correctly both times and can only
point. And declines opening a round on the third-file question, on the grounds that
ours arrived as a conversion-timing failure the agreement already forbids, so it may
be a symptom rather than a category.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018UTxuSizozEA8yDitPuris
Closes the last unexercised assumption with a throwaway token: partial clone over
HTTPS with a token against a private repository, list through the tool, the
short-form raw URL redirecting to a resolved branch under auth, and a conditional
request returning 304 across that redirect. A private repository can adopt from
another private one, by name or by URL, and check works off the lock afterwards.
Fixes what would have shipped as noise. The warning fired whenever a fetch needed a
credential, which in a private repository adopting from a private repository is
every time and legitimate. add now resolves origin and makes one anonymous request
to learn whether this repository is public, so the warning fires when the source is
private and the destination is not. The limit is stated in the message rather than
implied away: the signal tells public from not-public and nothing finer, so two
repositories private to different people is the case that genuinely widens access
and the one this cannot see. No origin means cannot tell, which warns — cannot tell
must never read as not public.
Notes when an adopted document did not come from .loom/published/, without
refusing. What is not exported is not hidden, but a lock against it records a
dependency on something that was never a contract.
And records the framing that settles all of this: the tool is a mast, not a lock.
It grants no access, everything it does is possible with copy and paste, and the
locks mean nothing outside the tool and the discipline of the agreement — so it
makes the wrong thing deliberate rather than impossible. What it adds over a paste
is not restriction but provenance.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018UTxuSizozEA8yDitPuris
The config becomes kubectl-shaped — named contexts with one current, each holding a
host, its flavor and a read-only token — and a bare owner/repo resolves against it.
The point is where the details live: a host's raw-file route belongs to the host so
it sits in the context, the published directory belongs to the convention so it
sits in the code, and what is left is which repository and which document, which is
the only part a person knows.
Measured rather than assumed, because the three hosts differ. Gitea redirects its
short raw form to the resolved branch, so the lock records a branch without anybody
naming one. GitHub and GitLab accept HEAD and do not redirect, which would put a
moving ref in the lock — the hazard already recorded and nearly built anyway — so
those resolve the default branch with git ls-remote --symref first, one round trip
and no clone.
That supersedes the claim that list is the only command needing git, which is now
wrong for two of three flavors and would otherwise read as still true.
Adds loomctl config, which says which context is current and where the credential
came from without printing it, so that "my token is not being used" is answerable.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018UTxuSizozEA8yDitPuris
Writes the first real gaps: what an adopter could not tell from the documents,
recorded while the misunderstandings are fresh rather than after they stop being
visible.
cart — whether a response is an annotation or a reply in your own file, and the
discriminator that resolves it appearing nowhere; what to do when an open item has
no fallback, which breaks the mechanism rather than omitting a nicety; what a third
file means once one exists; and what must be extracted before converting, now that
v1 makes conversion destroy.
annotating — how a reader detects the freeze was violated, since the hash we check
is a local invention; that responding means a rewrite, which is named only in our
local protocol; and that under cart v1 a prompt's answer has no archive at all.
recording-decisions — the one that matters most: never revised and somebody must
be able to read it both hold and eventually conflict, and the document does not say
which gives. Provisional answer recorded, with the line compaction may not cross —
never drop a decline, a measurement, a falsified belief, or a supersession, because
those are the record of somebody changing their mind.
publication — whether publishing implies the audience can read it, and where
confidentiality lives given that what is safe to publish depends on who can read
the repository it sits in.
Also declines reference-only adoption with its reason: it is adoption minus
reconciliation-by-diff, not adoption minus offline reading, and the convention's
other exit is the one that actually occurred. The warning now names both exits
including the unbuilt one.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018UTxuSizozEA8yDitPuris
Under cart/v1 the cart is untracked, so converting deletes the dailies outright
rather than leaving them in history. That makes the write-ahead log the mechanism
rather than a convenience, and it sharpens the conflict already recorded in
cart.usages.md: the annotation protocol here says commit before dissolving because
git is the only archive, and an ignored cart has no archive.
Answers whether the cluster split removes the need for a PAT. For this repository
yes, and in general no: the confidentiality rule sorts the cases, so a public
repository may only adopt public documents and never needs a credential, while a
private one adopting private documents legitimately does. Needing a credential
stopped being a capability and became a signal, which is why add warns rather than
deciding — it sees that the source is private and cannot see who may read the
repository the copy lands in.
Downgrades the git-over-HTTPS assumption from load-bearing to unexercised. The code
did not change and its status did, which is the kind of thing only a log says.
Records that the completeness case has a mechanism after all — the usages facet the
tool points at when it stages a polad — and that it failed the first time because
the facet did not exist rather than because nothing could work.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018UTxuSizozEA8yDitPuris
Both parties reached the same deadlock independently: add refusing an existing
path, and check unable to ask about a document with no recorded origin. adopt was
the other way out and is not obviously worse.
Not taken because the byte comparison makes the verb unnecessary. Whatever the
command is called it must not trust the claim — it has to fetch and compare,
because the point is that nobody knows whether the local copy is still a copy — and
once it does that, it is add with an origin supplied. A verb earns its place by
naming an act rather than a state, and present-but-unlocked is a condition a
repository is temporarily in.
Recorded with the falsifier: if people and agents keep dropping documents into
externals by hand, and they will, it is a recurring category rather than a
migration, and then it deserves its own word in --help rather than a hint.
Also makes check's hint say that add supplies the origin without rewriting the
local copy, since the previous wording left that to be inferred.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018UTxuSizozEA8yDitPuris
Restores something the specimen said and the round that discarded the specimen
lost with it: a changed external becomes a polad in the cart, and somebody
decides. check now stages what moved into .loom/cart/current/polad/ with the ETag
that was served alongside the bytes, and prints the .usages.md beside it, because
reconciliation runs the other way — the facets usually survive and what moves is
the code a usage named. It says so when there is no usages file, which is its own
finding.
With no cart open, check reports and stages nothing. The tool does not open a
round: a cart is a bounded exchange between two presences and starting one is
somebody's act, not a side effect of asking about freshness.
add now adopts what is not here and refuses what is already adopted, superseding
the entry that had it announce an overwrite — it no longer overwrites at all. The
one exception is the only way out of a dead end: a document present but unlocked
was fetched by hand, nothing records its origin, and the path does not round-trip,
so check cannot ask about it and a refusal would strand it forever. add accepts it
and the bytes decide — identical locks it without rewriting anything, which makes
the lock's assertion verified rather than assumed, and different stages a polad and
leaves the local copy alone because it is the only evidence anything moved.
apply exists because the lock is the half a person forgets: moving a polad by hand
leaves a lock describing the copy you just replaced. Recorded with its limit —
for an external, discard does not mean the change goes away, so discarding is
really knowingly stale and nothing yet records that choice.
Measured end to end on this repository: eight hand-fetched documents, all eight
locked, nothing rewritten.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018UTxuSizozEA8yDitPuris
The upstream cart.md change was planted to test whether the tool would notice, and
the honest sequence is that check did not catch it. Every document was unlocked,
and unlocked means I cannot tell you. add revealed the change by overwriting the
file, and the only reason the diff was visible is that git happened to be watching
the working tree.
So the mechanism is proven and the workflow is not: a document nobody has locked is
a document nobody is checking, silently.
Fixes the defect that exposed. add reported only "adopted" while overwriting local
content that differed from what the publisher serves — and a differing copy is the
only evidence that something moved while the document was unlocked. It now says so.
Recorded with its own limit: add still overwrites rather than stashing, so outside
a git working tree the note says something was lost without letting you see it.
Also records the unowned half. Nothing decides when check runs, and every answer —
a hook, CI, an agent's session start, a schedule — is outside the tool. Filed as a
need with no owner rather than built, because choosing one would put a scheduler
inside a fetcher and a comparator, with the falsifier that if nobody ever runs
check unaided then a tool that reports only when asked reports nothing.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018UTxuSizozEA8yDitPuris
Go, standard library only, with git shelled out for list alone. list enumerates a
publisher's .loom/published by partial clone and ls-tree; add fetches one document,
writes it under .loom/externals and records the resolved origin and the publisher's
ETag in .loom/externals/.locks; check asks conditionally and reports.
The first real run did what the tool exists for. All eight documents adopted by
hand before it existed reported unlocked — the tool refuses to invent a lock by
adopting whatever the remote currently serves, since that would assert the local
copy is the one being served, which is the thing it was about to check. Locking
them fetched two that had moved: bedrock/starting.md, which now says the worked
example is private and will not link to something you cannot fetch, and cart.md,
which went to v1.
cart v1 changes a role we cast: a cart is not committed, because a committed cart
grows a third file by itself — version control does not require anybody to ask, so
the two-file rule is never invoked — and because ignored, gone means gone. Adds
.loom/cart/ to .gitignore and supersedes the isolation entry rather than editing
it. osprey and marmalade are already in history and are left there: rewriting to
honour a rule adopted afterwards costs more than it buys.
Records the conflict this creates rather than settling it: the annotation protocol
here says commit before dissolving because git is the only archive, and an ignored
cart has no archive, so dissolving would destroy the annotations outright.
Credentials are read-only, per host, and passed to git through the environment
rather than argv, because argv is visible to every process on the machine.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018UTxuSizozEA8yDitPuris
Moves the wal into the log and closes the round. Nothing was unresolved: the one
open item — where the credential lives — is recorded as decided by fallback rather
than shelved, since cart says silence is a usable reply and that a fallback which
quietly becomes the decision is worth being able to find later.
The round decided the shape (loomctl, nouns, no plugin seam until somebody reports
a gap, no audience declaration), declined published check on the registry argument
with both its falsifiers, kept the insight that outlives it, settled list on git
with the three-host measurement behind it, established that nothing needs ssh and
nothing needs write access anywhere, chose Go, and recorded git-on-PATH as the
consequence.
Two entries carry things nobody decided. One is the untested assumption that a
read-scoped PAT authenticates git over HTTPS and not only the API and raw files,
which is the first thing to run against the next token and the only argument that
would bring ssh back. The other is a wrinkle in the isolation role: it assumed work
happens during a round, and both rounds so far have decided work that happens
after, so a branch is named for the round that authorised it and outlives the cart.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018UTxuSizozEA8yDitPuris
Jeff's correction breaks published check rather than denting it. homelab-cluster
publishes six documents and is private deliberately; publication.md says others,
not everyone, and I read one word as the other. The tool as logged would have
reported a correct repository as broken, so a superseding entry is added rather
than revising the osprey entry. What survives is that bedrock's public page
promising a private repository as the better answer was a real defect, and it was
never about homelab-cluster.
So published check must not report a verdict. It reports the identity it used, and
warns when that identity is the repository's own owner — publishing is not an act
you can complete alone, and a tool that verifies with your own credential has
helped you complete it alone and told you it was fine.
Also corrects me: I confirmed the specimen's token claim enthusiastically without
testing the case that would show it wrong. A refused write names the token's
scopes in the error body, so a token cannot be asked what it may do but can be
told by being refused. Verified that no other channel exists — the tokens endpoint
rejects token auth, and no scope header comes back on success.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018UTxuSizozEA8yDitPuris
The round is over and its artifact is .loom/event-log.md — fifteen entries, each
carrying the belief that could show it wrong, all tagged osprey so it stays
findable what else was in the room.
No spec is written. A spec flattens everything to equal confidence, which is how
the specimen managed to be wrong with a straight face within a day of being
proposed. The specimen is discarded rather than promoted; the story of its being
wrong is in the log, where it can be found, and the artifact would only mislead
somebody who read it without the round beside it.
Reconciles externals.md, which check flagged as the one drifted document of eight.
The new copy carries the 404 row, so the gap filed against it is closed and
externals.gaps.md is deleted — at reconciliation rather than at repair, because a
facet describes the local pair and the gap stayed true of the copy we held until
we fetched. The entry recording that our 404 behaviour is no longer provisional is
the one that would have been easiest to skip and is the reason the log exists.
No shelf: nothing was unresolved.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018UTxuSizozEA8yDitPuris
jeffryandClaude Opus 5