docs-revamp #1

Merged

oleksii merged 3 commits from docs-revamp into master

2026-10-03 13:23:00 +00:00

oleksii commented

2026-10-03 13:20:11 +00:00

Owner

No description provided.

oleksii added 3 commits

2026-10-03 13:20:11 +00:00

Default the state backend to file, not redis a12237a2e0

Silence used to fall through to redis; it now falls through to file
with DefaultStoreDir. Explicit signals are unchanged: KPR_STORE is
still authoritative, KPR_STORE_DIR alone still selects file, and
KPR_REDIS_ADDR alone still selects redis. Files match what the tool
is for — one binary beside a registry, nothing else to run — and
docker-compose.yml already shipped the file stack as its default.

Two tests took their "no redis" condition from the old fallback
rather than asking for it, so both now point at a dead address:
TestOpenStoreNamesDeadRedis, and TestServeRunServesAndStopsOnSigterm
which would otherwise come up healthy on a file store and write a
kpr/ dir into the package directory.

Breaking for deployments running with neither variable set: they
come back on an empty file store with their rows still in redis.
Setting KPR_REDIS_ADDR restores the old behavior.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Restructure docs: split future work, extract the hexagonal study 68b010371b

Reference pages described shipped behavior and unbuilt designs in
one breath. Future work moves to <DOC>_FUTURE.md so it can go stale
on its own: ARCHITECTURE (current state, open items), GC (dangling
tags, token auth), BACKFILL (shadow reader, digest-less
enrichment), PROOFS, SENTINELS.

HEXAGONAL keeps the port map and gains a mermaid diagram of the
four bands plus a package table derived from the real import graph
— the ASCII block it replaces listed 6 of 18 packages and omitted
edge, backfill, lineage, proof and sentinel. The graph also shows
sentinel is a pure leaf, which no page said. The three misses and
their evaluations move to HEXAGONAL_STUDY.md; HEXAGONAL_WISDOMS is
untouched.

BACKFILL was an implementation plan in future tense for a command
that ships, down to a Steps section describing work already done;
rewritten as reference. That surfaced stale claims elsewhere:
ARCHITECTURE said kpr-backfill "will" sign rows, GOALS still listed
backfill as an open question.

docs/README.md is now the index, grouped by task, and the top-level
README links to it instead of carrying 25 rows. goals.md renames to
GOALS.md to match its neighbours.

Prose: the longest paragraphs become lists and tables. GC 599 words
to 188, BACKFILL 508 to 120, HEXAGONAL 396 to 109. Also fixes a
README reference to a --force flag that does not exist (it is
--accept-mode-flip) and a "make up" comment promising a redis
container the default stack has never had.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Revamp ADOPT and QUICKSTART file-first

CI / Test & Build (pull_request) Successful in 2m2s

Details

CI / Build All Platforms (pull_request) Has been skipped

Details

CI / Release (pull_request) Has been skipped

Details

0363c1071f

ADOPT's kpr service declared no volumes, but its first-run step
told you to run `kpr store unlock` — which resolves the store root
from /etc/distribution/config.yml and then proves that mount is the
store the registry serves. Neither the config nor the storage
volume was mounted, so the step could not have worked. Both are now
in the service definition, with a comment on why they are not
optional.

Both guides lead with the file backend. ADOPT drops from "three
wires (notifications, deletes, one redis DB)" to two wires plus the
shared volume, which was always the honest count: the volume is
mandatory, redis never was. Redis moves to a closing section for
when one host and one directory stop being enough, keeping the DB-4
convention.

Other stale content in ADOPT: backfill described as a known gap
citing a PLAN file that does not exist, and GC instructions for the
stock collector that predate `kpr gc`. Troubleshooting prose
becomes a symptom/cause table carrying the failure modes those
missing mounts produce.

QUICKSTART drops the now-redundant KPR_STORE=file, adds the uid to
the kpr service (prose said both writers must match, config set it
only on the registry), and drops the blobdescriptor cache so the
stack is trap-free: deletions are visible the moment gc finishes,
matching registry-config.file.yml.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

oleksii merged commit a3bc22b5a6 into master

2026-10-03 13:23:00 +00:00

oleksii deleted branch docs-revamp

2026-10-03 13:23:01 +00:00

oleksii referenced this pull request from a commit

2026-10-03 13:23:02 +00:00

Merge pull request 'docs-revamp' (#1) from docs-revamp into master

oleksii referenced this pull request from a commit

2026-10-04 19:18:08 +00:00

Merge pull request #1 from noroutine/dependabot/github_actions/codecov/codecov-action-7