jeffryandClaude Opus 5
101b0280c4
bedrock publishes venues.md now, so the four-line section here had become the second saying of a rule owned elsewhere. Kept the one sentence that is genuinely this document's: put a choice where reconciliation will look for it, because a .usages.md is what somebody opens when a document moves. The old heading — "venues, for things you cannot fetch" — was wrong and had been since it was written. A venue is not defined by being unfetchable, it is defined by your having decided something. Kafka's documentation is extremely fetchable. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
9.1 KiB
Event log
Decisions about this convention. Appended, newest last, never revised. An entry states what was decided, what it is believed to advance, and the belief that could turn out false.
Decided (2026-09-07, loom + claude-substrate — discovered, cart: osprey):
404 is named in the convention as two answers wearing one status, and a
client reports both readings rather than picking one.
Advances a consumer not being told the wrong one of the document was withdrawn and you no longer have access.
Because it was filed as a gap against this document by loom-cli, using this
document's own test — could you say whose job it is? It was ours. The
table listed 304, 200 and 410 and not 404, while a host returns 404
for both readings by design: one that distinguished them would leak the
existence of things you may not see.
Over ssh they are distinguishable — permission denied against repository not found — so a client with both transports should say which it used.
Decided (2026-09-07, loom + claude-substrate — discovered, cart: osprey):
a facet describes the local pair. A gap is true of the copy you hold, not of
the document upstream, so a publisher fixing their end does not close it.
Advances a consumer's tree never describing a document that is not in it.
Because we nearly deleted a consumer's .gaps.md on the grounds that the
upstream document had been repaired. Their copy still lacked the row. The
second presence in osprey caught it, and the general form follows from the thing
this whole convention rests on — the copy is theirs and everything beside it is
ours — which we had not followed through to reconciliation.
So a gap closes at reconciliation, not at repair.
And the consequence we would not have reached: what survives is not the gap, it is what the gap justified. A workaround is often not retired by a fix — it stops being a workaround and becomes the specified behaviour, unchanged in the code and entirely changed in status. That change is invisible where the code is, so it is an entry in the consumer's own log — otherwise somebody inheriting the workaround goes looking for the gap that justified it and finds nothing.
A closed gap is not a decline: a decline is what you considered and did not do; a closed gap is what you needed and got. Opposite sign, and filing one as the other puts a thing you wanted into a list of things you rejected.
Decided (2026-09-07, loom + claude-substrate — discovered, cart: osprey):
the path does not record the origin. A lock does, holding the resolved URL and
the publisher's ETag, in .loom/externals/.locks.
Advances the one command that has to send a request being able to construct one.
Because loom-cli's builder tried to run check against a real tree and
could not. The stored path does not round-trip to a URL — we hold
git.hypertheory-labs.dev/loom/externals/externals.md and the document is served
from /loom/externals/raw/branch/main/.loom/published/externals.md. It 404s,
measured.
Two segments are dropped and the second is not routing:
- the host's raw route and the branch — recoverable only by knowing that host's URL shape and guessing a branch name
.loom/published/— whichpublication.mdsays is the entire contract. So a consumer's tree did not record whether a copy came from a published surface or from a file its owner may rename at will.
And the short form is worse than incomplete: a host may redirect it to whatever the default branch is at the time you ask, so a lock holding one is locked to a moving target and a branch rename reports as a change in the document.
This is the
ETagmistake one layer down, and the builder named the appetite behind both: the design is beautiful when nothing is written down, and both times what made it possible was a property of one host.
Second occurrence, so it is a pattern rather than an incident: when a design claims something is derivable and therefore need not be recorded, check whether the derivation is a fact about the general case or about the host in front of you.
Finding (2026-09-07, claude-substrate — cart: osprey): gitea's ETag
being the blob hash is a migration aid, not a mechanism.
loom-cli's builder ran check across eight documents with no locks at all,
by hashing local copies and comparing — the exact thing this convention
forbids — and it worked, because on this host the two coincide.
It is worth doing once, to lock what was fetched by hand before a tool existed. It is worth nothing after that, and it works on gitea only.
Recorded because the next person to notice the coincidence will think they have found the good idea again, and there is now a log entry saying it was found twice and rejected twice.
Seven of the eight were byte-identical to upstream, so this host's raw serving normalises nothing — which is a fact about this host and not a licence.
Decided (2026-09-07, loom + claude-substrate — discovered, cart: marmalade):
do not adopt from a source less readable than the repository you are adopting
into. Reference-only, or ask them to publish.
Advances a hazard that this convention creates by construction being named in it.
Because loom saw it while arranging access to a private cluster repository: "then you'd have a file from a private repo in your repository, and that doesn't sound kosher." It is not.
Adopting is copying, and confidentiality does not travel with the copy. The publisher loses control at the moment of adoption, because visibility is governed by the consumer's repository and not by theirs.
Nothing in a tree marks a copy as having come from somewhere private. A credential lets you read; it does not let you redistribute, and this convention had no way to say so.
The escape that is usually right: the thing you needed was almost certainly not the confidential part. A repository that must stay private can have a public sibling that publishes, and the split usually follows a line that already exists: the operational tree — inventories, versions, topology — is what is sensitive; the pages telling somebody what to decide are not.
Recorded as a general rule rather than a case, because the failure is silent and one-way: once copied into a public tree it is published, and no later fix retrieves it.
The lock grows a fourth column: visibility
Decided. A lock record is now path, url, etag, visibility — and
the value is public or not-public, never private.
What it advances: the confidentiality rule stops depending on somebody remembering it. Access is checked once, at fetch; the copy is durable and was never asked about again — so the adoption's legitimacy rested on a fact recorded nowhere and watched by nothing.
Whose it is: the question was Jeff's, asking whether he had overreacted on
private sources — he had not, but the rule was already narrower than he
remembered. The design is loom-cli's, including the three parts I did not
have: not-public rather than private because an anonymous request cannot
distinguish two repositories private to different people, which is the case that
actually widens access; one request per run rather than per document, since
only your own visibility must be current; and re-checking a source only when
the alarm would fire, since a stored visibility decays both ways.
Declined: refusing the adoption. loomctl warns and proceeds. Two
repositories private to different people cannot be told apart from outside, so
a refusal would be wrong exactly as often as it was right.
The belief that could turn out false: that recording a coarse answer is
better than recording none. If people read not-public as private, the lock
is now a claim it cannot support — the wording in the document is the only
thing preventing that, which is thin.
Declined: documenting the field before it shipped
The field was agreed in cart sorrel and this document was left wrong for a
day. Documenting a format that does not exist yet is the same mistake as
publishing into a repository nobody can read — it passes every check available
to the writer.
Declined: keeping the venue section
bedrock publishes venues.md now, so this document's four-line section had
become the second saying of a rule owned elsewhere.
Kept: the one sentence that is genuinely about externals — put a choice where
reconciliation will look for it. That is a claim about .usages.md, which is
this document's, and it happens to be the sentence that tells you when not to
write a venue file.
Everything else went, including the "things you cannot fetch" framing. That heading was wrong and had been since it was written: a venue is not defined by being unfetchable — it is defined by your having decided something. Kafka's documentation is extremely fetchable.