You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Phase 6: Apply the pattern to the other repos #1378
nemar-cli goes first because it is the biggest, and whatever pattern it settles becomes the
pattern for the rest. This phase applies that pattern outward.
Repos in scope
Each has agent-facing material and at least some of it is admin-only under ADR 0057's recipe test:
nemarOrg/website (public): AGENTS.md plus a .context/ tree including 14+ ADRs
nemarOrg/docs (public): its own AGENTS.md, already partly corrected in phase 0
nemarOrg/nemar-tools (private): already private, so the source split does not apply, but the
retrieval contract does
Scope
Run phase 1's classification rule over each repo's agent files.
Move what the rule marks admin-only into the private content repo phase 3 creates.
Rewrite each AGENTS.md to hard rules plus docs URLs, carrying the same escalation instruction
phase 5 writes: a non-admin who hits an admin-adjacent problem tags the admin team rather than
being stuck.
Leave ADRs where they are. They are binding decisions about the code they sit next to, and the
docs link them at their public GitHub paths.
Note
nemar-observability commits CLAUDE.md, which the owner's standing convention keeps gitignored.
Worth resolving while touching that repo, but it is not this epic's decision to make.
Correction: this issue's scope still assumes a design ADR 0059 replaced.
The scope above says to "move what the rule marks admin-only into the private content repo phase 3 creates." Phase 3 creates no such repo. ADR 0059 settled on one docs repo, private at source and public at the URL, precisely to avoid a second content repo, a build-time content pull, a content credential, and a degradation path for when the private half is unreachable.
So for each repo in scope, admin-only material moves into nemarOrg/docs under /admin/, where the existing gate gives it exactly the protection nemar-cli's does. Everything else in this phase stands.
Two other things this phase should now carry, both learned after it was written:
Retrieval is already solved for agents. Phase 2 gave every docs page a .md mirror and phase 3 added nemar admin docs <path...>, which trades a CLI API key for a short-lived docs-scoped session (Phase 3 (part 1): the credential exchange, so an admin can read gated docs without a browser #1384). So the per-repo AGENTS.md rewrites can point at URLs and name a command that actually retrieves them, rather than describing an aspiration.
nemarOrg/docs is listed as public here. It is public today and becomes private at source in phase 3, so by the time this phase runs that line is wrong. Its own AGENTS.md still needs the pass.
Scope survey, 2026-09-14. The repo list in this issue is wrong in both directions.
Measured against the live org rather than from memory:
repo
visibility
AGENTS.md
CLAUDE.md
.context/
.rules/
website
public
32.8 KB
3.1 KB
8
9
nemar-citations
public
20.3 KB
?
?
?
docs
private
15.6 KB
374 B
5
7
nemar-observability
public
10.0 KB
374 B
5
8
nemar-infisical
private
5.1 KB
?
?
?
orcid-mock
private
3.4 KB
?
?
?
nemar-tools
private
none
none
none
none
nemar-tools has no agent files at all. Not an access problem: its root lists .gitignore, README.md, docs/, requirements.txt, scripts/, templates/ and nothing else. There is nothing here to reclassify or rewrite. Either drop it from scope or make the task "create one", which is a different piece of work and should be decided rather than assumed.
Three repos with agent files are missing from scope, and one of them is significant: nemar-citations carries a 20.3 KBAGENTS.md and is public, so ADR 0057's source-split rule applies to it exactly as it did to nemar-cli. nemar-infisical and orcid-mock are already private, so the split does not apply, but the retrieval contract and the escalation instruction do.
website is the real work. At 32.8 KB its AGENTS.md is larger than nemar-cli's was before phase 5 cut it from 669 lines to 313, and it carries 8 .context/ entries and 9 .rules/ entries on top.
One method correction carried from phase 4, because this phase repeats that exercise five more times: classify by grepping CONCEPTS, not code identifiers. A symbol-only search reported three times as many documentation gaps as were real, because documentation describes behavior in prose and rarely names the symbol. Acting on it would mean copying implementation detail into user documentation to close gaps that do not exist.
Suggested order, largest leverage first: website, then nemar-citations, then docs, then nemar-observability, then the two small private ones.
Plus nemarOrg/docs#39, which documented the website release process before any of it was deleted from website.
The scope in this issue was wrong in both directions and was corrected first.nemar-tools was listed and has no agent files at all; it has since been archived as stale (last push 2026-02-10, every reference to it documentary, its Zenodo flow superseded by EZID per ADR 0007). nemar-citations, nemar-infisical and orcid-mock had agent files and were missing from the list. nemar-citations mattered most: public, and the source-split rule applied to it in full.
Only one repository got shorter, and that is the honest outcome rather than a shortfall.
The pattern is "do not keep a second copy of the documentation site", not "make the file shorter". website was the only one carrying a real second copy, and it lost 200 lines. The other five describe their own systems, so applying the rule faithfully meant adding the two things they lacked: a retrieval contract and an escalation instruction. A pass that had cut them to look consistent would have destroyed information.
Two refinements this phase added to the pattern.
A section can mix material that moves with material that stays. website's deploy and staging sections looked like 133 lines to relocate, but the parts missing from the docs site turned out to be that repo's own code symbols (__EDGE_CACHE_NAMESPACE__, applySecurityHeaders, isNoindexHost, getCrossHostRedirect). Those are implementation detail; moving them would have pushed website internals onto a public documentation site. They stayed, and only the operational half became a pointer.
Classify by grepping concepts, not code identifiers. Carried from phase 4, where a symbol-only search reported three times as many documentation gaps as were real.
One consequence of the phase 3 flip surfaced here and was fixed.docs.nemar.org serves public pages from a repository that is now private, so a reader who spots an error cannot file against its source. The site footer now carries "Report a problem", pointing at this repository's public issue template, and support@nemar.org for a reader with no GitHub account. Both addresses already existed on the site; no new infrastructure was added.
Activity
neuromechanist commentedon Sep 14, 2026
Correction: this issue's scope still assumes a design ADR 0059 replaced.
The scope above says to "move what the rule marks admin-only into the private content repo phase 3 creates." Phase 3 creates no such repo. ADR 0059 settled on one docs repo, private at source and public at the URL, precisely to avoid a second content repo, a build-time content pull, a content credential, and a degradation path for when the private half is unreachable.
So for each repo in scope, admin-only material moves into
nemarOrg/docsunder/admin/, where the existing gate gives it exactly the protectionnemar-cli's does. Everything else in this phase stands.Two other things this phase should now carry, both learned after it was written:
Retrieval is already solved for agents. Phase 2 gave every docs page a
.mdmirror and phase 3 addednemar admin docs <path...>, which trades a CLI API key for a short-lived docs-scoped session (Phase 3 (part 1): the credential exchange, so an admin can read gated docs without a browser #1384). So the per-repoAGENTS.mdrewrites can point at URLs and name a command that actually retrieves them, rather than describing an aspiration.nemarOrg/docsis listed as public here. It is public today and becomes private at source in phase 3, so by the time this phase runs that line is wrong. Its ownAGENTS.mdstill needs the pass.neuromechanist commentedon Sep 14, 2026
Scope survey, 2026-09-14. The repo list in this issue is wrong in both directions.
Measured against the live org rather than from memory:
.context/.rules/websitenemar-citationsdocsnemar-observabilitynemar-infisicalorcid-mocknemar-toolsnemar-toolshas no agent files at all. Not an access problem: its root lists.gitignore,README.md,docs/,requirements.txt,scripts/,templates/and nothing else. There is nothing here to reclassify or rewrite. Either drop it from scope or make the task "create one", which is a different piece of work and should be decided rather than assumed.Three repos with agent files are missing from scope, and one of them is significant:
nemar-citationscarries a 20.3 KBAGENTS.mdand is public, so ADR 0057's source-split rule applies to it exactly as it did tonemar-cli.nemar-infisicalandorcid-mockare already private, so the split does not apply, but the retrieval contract and the escalation instruction do.websiteis the real work. At 32.8 KB itsAGENTS.mdis larger thannemar-cli's was before phase 5 cut it from 669 lines to 313, and it carries 8.context/entries and 9.rules/entries on top.One method correction carried from phase 4, because this phase repeats that exercise five more times: classify by grepping CONCEPTS, not code identifiers. A symbol-only search reported three times as many documentation gaps as were real, because documentation describes behavior in prose and rarely names the symbol. Acting on it would mean copying implementation detail into user documentation to close gaps that do not exist.
Suggested order, largest leverage first:
website, thennemar-citations, thendocs, thennemar-observability, then the two small private ones.neuromechanist commentedon Sep 14, 2026
Phase 6 is complete. Six repositories, six merged pull requests.
websitenemar-citationsdocsnemar-observabilitynemar-infisicalorcid-mockPlus nemarOrg/docs#39, which documented the website release process before any of it was deleted from
website.The scope in this issue was wrong in both directions and was corrected first.
nemar-toolswas listed and has no agent files at all; it has since been archived as stale (last push 2026-02-10, every reference to it documentary, its Zenodo flow superseded by EZID per ADR 0007).nemar-citations,nemar-infisicalandorcid-mockhad agent files and were missing from the list.nemar-citationsmattered most: public, and the source-split rule applied to it in full.Only one repository got shorter, and that is the honest outcome rather than a shortfall.
The pattern is "do not keep a second copy of the documentation site", not "make the file shorter".
websitewas the only one carrying a real second copy, and it lost 200 lines. The other five describe their own systems, so applying the rule faithfully meant adding the two things they lacked: a retrieval contract and an escalation instruction. A pass that had cut them to look consistent would have destroyed information.Two refinements this phase added to the pattern.
A section can mix material that moves with material that stays.
website's deploy and staging sections looked like 133 lines to relocate, but the parts missing from the docs site turned out to be that repo's own code symbols (__EDGE_CACHE_NAMESPACE__,applySecurityHeaders,isNoindexHost,getCrossHostRedirect). Those are implementation detail; moving them would have pushed website internals onto a public documentation site. They stayed, and only the operational half became a pointer.Classify by grepping concepts, not code identifiers. Carried from phase 4, where a symbol-only search reported three times as many documentation gaps as were real.
One consequence of the phase 3 flip surfaced here and was fixed.
docs.nemar.orgserves public pages from a repository that is now private, so a reader who spots an error cannot file against its source. The site footer now carries "Report a problem", pointing at this repository's public issue template, andsupport@nemar.orgfor a reader with no GitHub account. Both addresses already existed on the site; no new infrastructure was added.Epic #1336 has no phases left.