The AI "Ask" affordance (selection tooltip + prompt bar) had no response
surface on an entry's canonical `main` view: RFCView renders the human-
discussion panel there, not the chat panel, so a chat turn posted from the
reading view had nowhere to render — asking-while-reading silently did
nothing. AI chat is an editing activity (a turn can emit <change> proposals)
and only runs on an edit branch.
Option B: when Ask is invoked from main, transparently cut an edit branch via
the same dispatch as Start Contributing (promote-to-branch for active,
start-edit-branch for super-draft; idempotent — reuse an existing edit
branch), navigate onto it so the chat panel mounts, and run the question
(text + selected quote) as the branch's first chat turn once its view and
main_thread_id resolve. The turn fires from the message-load effect's
continuation (not a parallel effect) so a late message load can't clobber the
optimistic turn; a live ref keeps the latest submitChatTurn closure in reach.
Degrades gracefully: signed-out → login; a viewer who can't contribute has the
cut rejected server-side and the error surfaced (no spurious branch); asking
from a branch already, and Flag on main (human discussion thread), unchanged.
Frontend-only; the branch chat endpoints already work on a branch. SPEC §8.3
records the behavior; new RFCView unit tests + an e2e spec cover the flow.
backend 677 / frontend 66 / e2e 5 green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
§22 migrated the READ/catalog path to the three-tier (project/collection)
model but the WRITE/branch/body subsystem still hardcoded the default
project's default collection — resolving every meta-resident entry to the
default content repo at rfcs/<slug>.md, ignoring the entry's project (its own
content repo) and collection (a <subfolder>/rfcs/ prefix). An entry outside
the default collection rendered a blank canonical body and every edit/PR/
body-write path hit the wrong file.
- New single resolver: projects.content_repo_for_collection() +
projects.entry_location(config, cid, slug) -> (org, repo, md_path)
(collection -> project -> content_repo + subfolder; falls back to the
default repo for a legacy/unknown collection).
- Every entry write/read path resolves via it: api_branches (body GET + all
branch write paths), api_prs (pr-draft/open/merge/withdraw/review/
resolution-branch), api_graduation (graduate/claim/retire/unretire +
orchestrator + state-flip), api.py mark-reviewed + idea-PR merge/decline/
withdraw + proposal preview, api_metadata. bot.open_metadata_pr and
bot.mark_entry_reviewed gained a file_path param. refresh_meta_branches,
the webhook corpus-refresh dispatch, and the hygiene branch-delete now
span every project's content repo, not just the default.
- New additive collection-scoped body-read routes:
GET /api/projects/{pid}/collections/{cid}/rfcs/{slug}/main and
.../branches/{branch} disambiguate a slug across collections (G-5) and read
the entry's own repo. Slug-only routes kept (now collection-aware via the
cached row). Frontend getRFCMain/getBranch take optional pid+cid; RFCView
threads them (mirrors the v0.52.1 entry-detail fix).
No migration, no config change. Existing default-collection entries
unaffected. Tests: backend resolver + collection-scoped branch/body + graduate-
in-subfolder write path; frontend api unit. backend 677 / frontend 60 green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The top-level catch-all silently redirected unknown paths to "/", and the
nested /admin/*, /p/:projectId/*, and /docs/* route groups had no catch-all
(invalid subpaths rendered a blank pane). Add a shared NotFound component and
wire it into all four route groups so a bad/typo'd URL shows a clear 404 with
a link home. Client-side 404 UI (SPA still serves HTTP 200).
Patch 0.52.2 → 0.52.3; CHANGELOG updated. Caught by the operator on PPE.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The /admin left-rail NavLinks used relative targets (to="users", etc.). Under
the /admin/* nested route a relative link resolves against the current URL, so
each click appended a segment — Users→Allowlist→Graduation yielded
/admin/users/allowlist/graduation instead of /admin/graduation. Use absolute
targets (/admin/<tab>) + `end` for exact active matching.
Patch 0.52.1 → 0.52.2; CHANGELOG updated. Caught by the operator on PPE.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The §9 deployed-environment E2E harness (0.52.0), run against a PPE host
with per-collection-isolated content, surfaced a latent multi-collection
bug: RFCView computed the collection id from the route but called
getRFC(pid, slug) without it, so a named-collection entry was always
fetched via the project default-collection route — which 404s for an entry
that exists only in a named collection ("Error: Not found"; metadata panel
absent). Local/Tier-1 stacks masked it (same slug also reachable via the
default collection). Thread cid through all three getRFC call sites; re-run
the load effect on collection change.
Harness/test-infra (not in the deployed artifact):
- e2e: pre-record cookie consent via addInitScript (lib/fixtures.js) so the
bottom-fixed consent banner can't intercept catalog row-select clicks on
the slower deployed edge.
- testing/seed-ppe.sh: fail loudly on any non-2xx Gitea response (a
swallowed 403 org-repo create had reached the deploy as a 502).
- testing/ppe-deploy-and-test.sh: seed via the Keychain admin token
(write:organization needed to create the PPE repos); store the E2E secret
newline-free; read EXPECT_VERSION from VERSION.
Patch bump 0.52.0 → 0.52.1; CHANGELOG updated.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Runs the whole §9 PPE stage non-interactively after the operator's gcloud
reauth: bot-token-seed the PPE repos -> ensure E2E secret -> deploy ->
wait for health=0.52.0 + bdd-collection sync -> run metadata.spec.js
against the deployed host. Idempotent; secrets fetched from SM, never
echoed. Test/ops infra (not in the deployed artifact).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Test-infra only (e2e/ is not in the deployed artifact). Refines the
v0.52.0 deployed-env harness:
- retries:2 + on-first-retry trace now meaningful (was dead: retries
defaulted to 0); de-risks timing flakes over the public edge.
- workers:1 — the metadata specs run in order and write real commits;
parallel workers would race on shared state and concurrent bot pushes.
- deployed timeouts bumped (60s/20s) when E2E_TEST_AUTH_SECRET is set.
- dismissCookies forcibly removes any lingering consent-banner node so it
can't intercept catalog-footer checkbox clicks (the recurring flake).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The §9 pipeline's PPE+E2E stage was unreachable: e2e/metadata.spec.js was
bound to Tier-1-only scaffolding (docker-seeded faceted collection,
SQLite-injected owner, Mailpit OTC sink). This makes the same suite run
against a deployed host.
- backend: POST /auth/test/login — fail-closed, secret-gated, single-
identity owner test-login (404 unless E2E_TEST_AUTH_SECRET +
E2E_TEST_AUTH_EMAIL both set; constant-time compare; loud startup warn).
6 vertical tests; backend 665 green. Documented in backend/.env.example.
- e2e/lib/auth.js: branch on E2E_TEST_AUTH_SECRET (deployed test-login vs
local Mailpit OTC); OWNER_EMAIL from E2E_OWNER_EMAIL. Spec unchanged so
the localhost Tier-1 path keeps working.
- testing/seed-ppe.sh: seed a dedicated, prod-untouching PPE registry +
content repo (faceted bdd collection) — real OHM content never touched.
- docs/design/2026-06-07-deployed-env-e2e-harness.md; CHANGELOG; VERSION
+ frontend/package.json -> 0.52.0.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Modernize the Tier-1 stack to the three-tier app and add browser coverage for
the §22.4a metadata UI, closing the E2E gap deferred across SLICE-3/4/5:
- seed-gitea.sh: create a REGISTRY_REPO with projects.yaml + a faceted named
collection (.collection.yaml fields: priority enum + tags) seeded with three
metadata-bearing entries; register content+registry webhooks; self-guarding
(skip if a prior token still works) so a dependency-triggered re-run can't
remint and invalidate the backend's token.
- .env.tier1: REGISTRY_REPO/DEFAULT_PROJECT_ID; disable OTC cooldown + lift the
per-IP auth limiter for the single-IP test runner.
- docker-compose: pin backend image; backend-seed inserts a granted owner the
OTC path can sign in as (write paths need contributor+).
- Makefile: two-phase tier1-up (seed to completion, then create backend so it
reads the populated token env); robust down; e2e-fresh = down+up+e2e (the
canonical run, since the edit/bulk specs mutate the seeded corpus).
- metadata.spec.js: SLICE-3 faceted filter (anon), SLICE-4 edit panel (owner),
SLICE-5 bulk bar (owner). 4 passed against a fresh stack.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add RATELIMIT_OTC_REQUEST_MAX / RATELIMIT_VERIFY_MAX / RATELIMIT_CHECK_MAX
(default to the existing secure values; non-positive/unparseable → default).
Lets a test/PPE stack that drives the auth endpoints repeatedly from one IP
lift the budget; production leaves them unset. Mirrors the existing
OTC_REQUEST_COOLDOWN_SECONDS knob.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
useCollectionId read useParams().collectionId, but the Catalog renders at
/p/:projectId/* — outside the c/:collectionId route — so it always fell back
to the default collection. The faceted filter (SLICE-3) and bulk action bar
(SLICE-5) therefore never scoped to a named (fields-bearing) collection in the
deployed app; the Catalog unit test had mocked useCollectionId, hiding it.
Resolve the /c/<id>/ segment from the pathname when the route param isn't in
scope. Caught by the new Playwright metadata suite.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
On a deployment that already had >=2 projects when migration 029 ran, 029
seeds the default project's collection id as the project id (e.g. 'ohm'),
but the registry mirror expects 'default' -> it inserted a duplicate empty
'default' collection, orphaning the entries.
Adds projects.reconcile_default_collection_id() -- the collection-grain twin
of restamp_default_project -- run at startup BEFORE the mirror so it merges
onto the canonical 'default' collection instead of duplicating. Renames the
divergent collection id and cascades collection_id across all keyed tables
(FK-off atomic rename + foreign_key_check). Idempotent; no-op on fresh /
single-project / already-aligned deploys. Seam tests show the duplicate forms
without the fix and merges with it.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Code-review follow-ups:
- Validate the raw submitted value before apply_values coerces it, in both
the single-edit and bulk endpoints — a scalar set onto a tags field now
rejects (422 / rejected) instead of silently char-splitting into a list.
- Catalog prunes bulk selections to the currently-visible (filtered) entries
after each list fetch, so the bulk bar can't act on rows the user has
filtered out of view.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Faceted catalog rows are selectable for contributors; a sticky bulk bar
(BulkActionBar) drives set/add/remove gestures from the collection fields
schema, calls bulkEntryMeta (one commit), toasts applied/skipped counts,
and re-fetches. Non-contributors and legacy (no-fields) collections see
no selection (INV-5).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
MetadataFieldsPanel renders one control per declared field (enum→select,
tags→chips, text→input), read-only without contribute access, saving changed
values via saveEntryMeta (direct sidecar commit). Wired into the canonical
main view, gated on the collection declaring fields (INV-5). saveEntryMeta
API client added.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
POST .../collections/{cid}/migrate — now safe to ship since all write paths
are sidecar-aware. Idempotent, one commit per collection.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Direct-commit to the sidecar (D7), contributor+ gated (INV-4), schema-validated
at the write boundary, lazy-migrates a legacy entry, re-ingests. GET RFC now
returns the per-entry meta mapping and a can_edit_meta capability.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
graduate, claim, retire/unretire, _read_meta_entry, mark_entry_reviewed,
body extract/wrap (api_branches + api_prs replay) now dual-read and write
metadata to the sidecar via write_entry_files + bot.commit_entry_files/
open_entry_pr — a migrated body-only .md no longer crashes entry.parse or
re-grows frontmatter; legacy entries lazy-migrate on first metadata write.
Existing tests updated to assert the sidecar (INV-2 clean docs).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
apply_values, EntryGitState, read_entry_from_git, write_entry_files, sidecar_path_for.
Plus the SLICE-4 implementation plan.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Collections declare a `fields:` schema in `.collection.yaml`; entries carry
typed metadata (enum/tags/text). New central `metadata_schema` module parses
the schema leniently (INV-3) and validates entry values — advisory at read
(corpus mirror flags violations as `metadata_malformed` without hard-failing),
the enforcement point for the write boundary (edit endpoints land SLICE-4/5).
- app/metadata_schema.py: parse_fields (lenient/normalizing) + validate
- registry.parse_collection_manifest reads `fields:` into collection config
- collections.get_collection unpacks `fields`; served by the collection API
- cache._refresh_collection_corpus validates each entry advisory-only
Non-breaking, opt-in: no `fields:` → unchanged (INV-5); default `document`
collection declares none (N=1 unchanged). No DB migration — schema rides in
collections.config_json. SLICE-2 of
docs/design/2026-06-06-configurable-collection-metadata.md §7.2.
Backend suite green (601 passed).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
§22.4a SLICE-1 of docs/design/2026-06-06-configurable-collection-metadata.md
(§7.2). Entry metadata can live in a per-entry `<slug>.meta.yaml` sidecar with
the `.md` kept as pure prose (INV-2). Additive and non-breaking — with no
sidecars present every corpus stays on the legacy frontmatter path,
byte-identical (N=1 unchanged).
- Dual-read (app/metadata.py `read_entry`) — sidecar-else-legacy-frontmatter,
identical records (INV-6); unknown/forward-compat keys ride along through
parse→serialize and migration (INV-7, `Entry.extra`). A degenerate sidecar
(malformed/empty/slug-less) never drops the entry — slug backstopped from the
filename stem, flagged not lost (INV-3).
- Migration tool (`metadata.migrate_collection`) — idempotent, one ChangeFiles
commit per collection (new `gitea.change_files`). Tested as a function; its
Owner-gated operator trigger is DEFERRED to SLICE-4 (write paths must become
sidecar-aware first — see the design's SLICE-4 note + INV-8). No production
trigger ships here, so no corpus is rewritten.
- Malformed flag — migration 033 adds `cached_rfcs.metadata_malformed`
(additive); the corpus mirror derives it; catalog list + entry-detail APIs
surface `metadata_malformed`.
- INV-7 at graduation — graduation now carries `Entry.extra` through the rebuild
instead of dropping forward-compat keys.
Gate: backend 575 passed (28 new: test_metadata / _migration / _cache +
graduation extra-preservation). Frontend untouched. CHANGELOG 0.47.0 +
upgrade-steps; VERSION + frontend/package.json -> 0.47.0.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Ben Stull