Maintaining Freeshard's docs of record (agents.md, README.md, docs.freeshard.net) and writing anything user- or public-facing. TRIGGER on editing agents.md or README.md, "update the docs", "is this doc still true", contradictions between a doc and the code, the Portal/Freeshard naming question ("should I rename portal to shard?"), portal_controller.py vs freeshard_controller.py confusion, writing PR descriptions/commit messages/release notes, or any external claim about what Freeshard is or does. SKIP when the task is the mechanics of config options (use freeshard-config-and-flags), cross-repo type-sync (use freeshard-ecosystem-contracts), release *execution* (use freeshard-run-and-operate), or change gating/review policy (use freeshard-change-control).
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Maintaining Freeshard's docs of record (agents.md, README.md, docs.freeshard.net) and writing anything user- or public-facing. TRIGGER on editing agents.md or README.md, "update the docs", "is this doc still true", contradictions between a doc and the code, the Portal/Freeshard naming question ("should I rename portal to shard?"), portal_controller.py vs freeshard_controller.py confusion, writing PR descriptions/commit messages/release notes, or any external claim about what Freeshard is or does. SKIP when the task is the mechanics of config options (use freeshard-config-and-flags), cross-repo type-sync (use freeshard-ecosystem-contracts), release *execution* (use freeshard-run-and-operate), or change gating/review policy (use freeshard-change-control).
Freeshard docs and positioning
How to keep this repo's documentation truthful and how to write anything a human will read: PR descriptions, commit messages, docs pages, release notes, public copy.
Vocabulary used below: a shard is a personal cloud VM; shard_core (this repo) is the FastAPI backend running on each shard; the controller (repo freeshard-controller) is the central management service; web-terminal is the Vue 2 end-user UI; the project's pre-2025 name was Portal, and residue of that name is load-bearing (see "Naming discipline").
When NOT to use this skill
You are doing...
Use instead
Adding/changing a config option, env var, or flag
freeshard-config-and-flags
Syncing types from the controller, cross-repo contract changes
freeshard-ecosystem-contracts
Cutting a release, deploying, operating a shard
freeshard-run-and-operate
Deciding whether a change is allowed / how it gets reviewed
freeshard-change-control
Understanding the crypto/auth design itself (not documenting it)
freeshard-domain-reference
Setting up the dev environment the README describes
freeshard-build-and-env
Docs of record and authority order
When documents disagree, trust them in this order — and trust code over all of them:
Concept sections good; the dead dev-setup mechanics were fixed by issue #128.
5
docs.freeshard.net (sibling repo documentation/, MkDocs Material, deployed to GitHub Pages via .github/workflows/ci.yml)
End users + third-party app developers
User-level truth; not an engineering reference.
6
/home/ubuntu/projects/freeshard/agents.md (workspace-level, one directory up)
Agents working across repos
Useful repo map + cross-repo command reference. NOT in any git repo — exists only on the dev machine; do not cite it as a versioned source.
Standing rule — agents.md must not rot: when your PR changes anything agents.md describes (tech stack, architecture, key patterns, commands, file-path conventions), updating agents.md in the same PR is part of the change, not optional polish. Stale statements there cause real agent mistakes. The proof: agents.md claimed Ed25519 crypto for years while the code was RSA-4096/PSS, and the claim was believed until issue #128 checked it against crypto.py.
Known-stale docs (as of 2026-07-15)
Fixing a stale doc is in scope for any PR that touches the described area. You do not need a separate issue to correct a doc you just proved wrong — but a pure doc-fix PR is also fine and cheap to review.
#
Location
Stale claim
The truth
Evidence
1
documentation/README.md (sibling repo)
"Documentation for Portal app developers" + GitLab Pages boilerplate, badge to ptl-public.gitlab.io
Legacy scaffold README. The repo's agents.md is the current reference: MkDocs Material site at docs.freeshard.net.
documentation/README.md vs documentation/agents.md:1-10
2
documentation/agents.md:61
"Built and deployed via GitLab Pages"
Deploys via GitHub Pages: .github/workflows/ci.yml builds with mkdocs build --strict and deploys with actions/upload-pages-artifact / actions/deploy-pages on main. No .gitlab-ci.yml exists in the repo.
documentation/.github/workflows/ci.yml
Before repeating ANY doc claim in new writing, verify it against code first. Re-verification one-liners for this table are in "Provenance and maintenance".
House style: PRs, commits, issues
Conventions observable in merged history — follow them.
Commit messages
Narrative. Subject line in conventional-commit-ish form (fix(profile): ..., chore(deps): ..., perf(backup): ...); body explains symptom, cause, and why this fix — see commit e55ce51 for the model.
Rationale lives in the commit message, not code comments. Do not annotate code with "why this change exists" comments; a future reader gets the reasoning from git log/git blame. Comment code only when it is genuinely puzzling without one.
PR descriptions
Must cover all commits on the branch, not just the last one. Reviewers read the description as the change's contract.
Multi-file PRs include a "Recommended reading order" section listing changed files in dependency order (schema/config → leaf services → API handlers → tests). Precedent: PRs https://github.com/FreeshardBase/freeshard/pull/83, /pull/90, /pull/108 all carry one.
End agent-authored PR bodies per the harness convention (the Co-Authored-By / generated-with footer).
Issues
Reference issues and PRs by full URL (https://github.com/FreeshardBase/freeshard/issues/N), never a bare #N — bare numbers are ambiguous across the 13 FreeshardBase repos and break when text is copied elsewhere.
Issue/PR states are volatile: re-query with gh before asserting one is open/merged; never trust memory or an earlier point in the session.
Naming discipline: the Portal residue
The project was named Portal until the Feb 2025 rebrand (rename commit 88e5842, "rename lots of stuff from portal to shard"). The README's top NOTE acknowledges the leftover naming publicly. What remains is intentional-until-refactored — much of it is external contract, not sloppiness:
Residue
Where
Why you must NOT rename it opportunistically
Docker network portal
docker-compose.yml:2-3, every app compose template, shard_core/web/internal/call_peer.py:36-39
Every installed app's container joins this network by name. Renaming orphans running apps fleet-wide.
Apps in the store consume these headers for auth context. Wire contract.
{{ portal.* }} template variables
shard_core/service/app_installation/util.py:91-98 (render passes portal= and fs=)
Every published app template in app-repository uses them; docs.freeshard.net documents both portal.* and the newer fs.* sets with an explicit backwards-compat note (documentation/docs/developer_docs/includes/template_vars_portal.md, portal_name_info.md).
minimum_portal_size
shard_core/data_model/app_meta.py:125
Field name in the published app_meta.json format (schema v1.2). Renaming = format migration, not a refactor.
portal_controller.py module
shard_core/service/portal_controller.py
See trap below.
portal_core/ paths
Git history before 88e5842
Use both portal_core/ and shard_core/ prefixes when running git log --follow/churn analysis across the rename.
Rule: no mass-renames. A rename touching any row above is a cross-repo, fleet-affecting migration (app templates, published docs, running containers) and goes through freeshard-change-control as its own planned change. Never fold "cleaned up old Portal naming" into an unrelated PR.
Trap — two modules, one controller:shard_core/service/portal_controller.py and shard_core/service/freeshard_controller.py both call the SAME backend (settings().freeshard_controller.base_url = https://controller.freeshard.net) with different path conventions:
freeshard_controller.call_freeshard_controller does not — callers pass api/shards/self themselves (freeshard_controller.py:15). Handles the shared secret.
Adding a controller call to the wrong module produces /api/api/... or a missing /api/ prefix. Check which module the neighboring calls live in, and pass the path in that module's convention. (Note the naming trap that survives: the portal_controller.pymodule is live, but it reads settings().freeshard_controller — there is no [portal_controller] config section, since issue #128 removed the dead one.)
External positioning and claims discipline
What Freeshard publicly is: a personal cloud computer — the missing category between local software, the big-cloud web, and DIY self-hosting. One person (or household) owns the machine; sovereignty over data is the point. The public anchor phrasing is README.md:15: "a personal cloud computer that a consumer can rent for a small monthly fee (or selfhost) and which is as simple to use as a smartphone. Its aim is to restore people's sovereignty regarding the data they consume and produce."
Rules for anything public-facing (README, docs site, release notes, blog/newsletter copy, issue comments that will be read by users):
No performance, security, or privacy claim without a reproducible demonstration. If you cannot point to a script, test, or measurement someone else can rerun, do not state the claim. Write what IS demonstrable ("auth decisions are made on your shard, not in a central service") rather than superlatives ("bulletproof", "zero-knowledge", "blazing fast").
Status vocabulary is fixed: unshipped work is "planned"; a spike/PoC is a "prototype" (e.g. the embedded OIDC provider on branch spike/oidc-provider-poc is a prototype — unmerged as of 2026-07-03); merged-but-not-deployed is not "shipped" (deployment gates are freeshard-change-control's territory). Never present prototype results as product capabilities.
License precision. shard_core is licensed FSL-1.1-ALv2 (LICENSE.md) — Functional Source License, source-available, each version converting to Apache-2.0 two years after release (LICENSE.md:87-91). In external copy, describe shard_core as "source-available" / "Fair Source (FSL)" rather than unqualified "open source". The apps in the store, by contrast, ARE FOSS by policy (next rule).
FOSS-only app store is public policy. Third-party apps must carry a license from the allowlist — MIT, Apache-2.0, BSD-2/3-Clause, GPL-2.0/3.0, AGPL-3.0, LGPL-*, MPL-2.0, ISC, Unlicense, CC0-1.0 (app-repository/.claude/skills/add-app/reference/exit-criteria.md:38). Source-available licenses (BSL, SSPL, Elastic, O'Saasy) are a hard block regardless of technical fit — precedent: Fizzy, blocked 2026-05-18, recorded in app-repository/blocked_apps/fizzy.md. Rejections are documented in app-repository/blocked_apps/ so nobody re-litigates them; when writing about the store, you may state the policy plainly and cite that directory.
Third-party facts get verified, not remembered. Anything you assert about PayPal, OVH, Azure, Traefik, an upstream app, or any external product must come from a checked source or be labeled uncertain.
Release notes and user-visible communication
Version history: 0.x tags on this repo (git tag | sort -V); as of 2026-07-03 the latest tag is 0.39.4 (2026-06-24). Minor = feature, patch = hotfix. A move to full semantic versioning is decided-in-principle but not implemented — https://github.com/FreeshardBase/freeshard/issues/118 (OPEN as of 2026-07-03).
Patch bursts after a minor are normal, not a scandal. Every infra-level minor has produced a hotfix tail: 0.28.0→0.28.6 (7 releases in 3 days, 2024, rclone rollout), 0.38.0→0.38.4 (5 tags in 6 days, 2026, Postgres migration). Write the notes accordingly: state the fix plainly ("0.38.3 fixes backup validation failures on some shards"), never bury or euphemize it ("minor improvements"). Users on a broken version need to recognize their symptom in one line.
Don't infer "latest version" from the local working tree — the checkout may sit on a stale branch (this one pins 0.39.2 in docker-compose.yml while 0.39.4 is tagged). Query git fetch --tags && git tag | sort -V | tail -1 or the GitHub releases page.
Tag order ≠ commit topology here: 0.37.5 was tagged AFTER 0.38.0 as a deliberate old-line backport (branch origin/release/0.37.5). When writing "what changed between X and Y", diff the tags, don't assume linearity.
For prose style and voice of newsletters/blog posts, this skill only sets the facts-discipline; the maintainer has separate style guides outside this repo.
Doc-impact checklist (run for every PR)
Did I change anything agents.md describes (stack, architecture, patterns, commands, paths)? → update agents.md in this PR.
Did I touch an area with a known-stale doc (table above)? → fix that entry in this PR and remove it from the skill's table.
Did I add/rename/remove a just recipe, config option, or env var? → README dev section + agents.md Commands section; config specifics per freeshard-config-and-flags.
Does the change alter what third-party app developers see (template vars, headers, app_meta format)? → the documentation/ repo needs a matching PR (separate repo, own review).
Does my text reference issues/PRs? → full URLs, states re-checked with gh now.
Does my text make any public claim? → run it against the five claims-discipline rules above.
PR description: covers all commits, has Symptom/Cause/Fix or equivalent, has Recommended reading order if multi-file.
Provenance and maintenance
Written 2026-07-03; known-stale table updated 2026-07-15 after issue #128 fixed the agents.md Ed25519 claim, the README/justfile CONFIG mechanism, and the README GitLab API-doc link. Originally written against working tree at commit e55ce51 (branch fix-profile-billing-fields; origin/main then at 0a40684) plus sibling checkouts documentation/ and app-repository/ under /home/ubuntu/projects/freeshard/. Primary sources: repo files cited inline, git log/git tag, GitHub via gh (issues 41, 118; PRs 83, 90, 102, 108).
Drift-prone facts — re-verify before relying on them:
