* Curated marketplace enrichment via agnes-metadata.json + curator metadata
Adds a second well-defined metadata file `.claude-plugin/agnes-metadata.json`
that upstream marketplace repos can opt into, providing per-plugin (and
per-skill / per-agent) cover photo, demo video URL, doc links, and
category override. The Claude Code marketplace contract is untouched —
agnes-metadata.json + the convention `.agnes/` directory are stripped
from the synthetic Claude Code marketplace served via /marketplace.zip
and /marketplace.git/*, so user instances see a clean Claude Code repo
with no Agnes-only metadata.
Highlights:
- DB schema v32 — adds curator_name + curator_email on marketplace_registry,
cover_photo_url + video_url + doc_links on marketplace_plugins.
- Mandatory curator at marketplace registration, editable later through
the admin UI; surfaces on cards + detail pages in place of owner_todo.
- External-asset mirror cache at ${DATA_DIR}/marketplace-cache/<slug>/
with conditional GET, 60s timeout, 10 MB body cap, SSRF guards, and
Wikipedia-policy-compliant User-Agent.
- Strict drop semantics — anything Agnes can't deliver as a real PDF /
Markdown / plain text doc, or a real PNG / JPEG / WebP cover, is
dropped from the served metadata; UI looks identical to no-entry case
(gradient placeholder for missing covers, no row in the doc list).
- Doc allowlist + image allowlist enforced on both the curated mirror
flow and the Flea upload flow (/store/new); shared module
src/marketplace_assets.py.
- New /api/marketplace/curated/{mp}/{plugin}/{asset,doc,mirrored}/...
endpoints with path-traversal guards + RBAC + Content-Disposition
attachment for docs.
- Curator-focused format guide at /marketplace/format-guide; canonical
source is docs/curated-marketplace-format.md, also linked from the
admin /admin/marketplaces page next to + Add Marketplace.
See CHANGELOG.md under [Unreleased] for the full breakdown.
* Fix format-guide test assertion to match shortened disclaimer
The 'Flea Market' phrase was trimmed out of the disclaimer in
docs/curated-marketplace-format.md after the curator-focused rewrite.
Update the rendered-HTML test to assert the channel-scoping phrase
that's actually present ('Curated Marketplace channel only') rather
than the 'Flea Market' contrast that's no longer in the doc.
* Drop unused 'version' field from agnes-metadata.json schema
The parser never read it; it was a YAGNI placeholder for future
schema evolution. Curators don't need to wonder what to put there
when adding the file for the first time. Will be re-added if and
when we actually introduce a backwards-incompatible schema change.
* Harden asset mirror against SSRF via redirect + DNS rebinding
The pre-flight _is_safe_url check validated only the initial URL;
urllib.request.urlopen then followed redirects and re-resolved DNS for
the actual connection — both bypassable. Attacker-controlled origin
could 302 to http://169.254.169.254/... and exfil cloud metadata;
attacker-controlled DNS could return public IP first / 127.0.0.1 second.
Replace urlopen call with a shared OpenerDirector wired through three
custom handlers: _SafeRedirectHandler re-runs SSRF allowlist on every
redirect Location (max 5 hops, down from urllib's 10), and
_PinnedHTTPHandler / _PinnedHTTPSHandler connect to the IP that passed
validation rather than re-resolving the hostname. TLS SNI + cert verify
stay bound to the original hostname.
_resolve_safe returns the validated IP (the existing _is_safe_url
2-tuple wrapper stays for backwards compatibility) and rejects round-
robin DNS that mixes a public + private record. _UnsafeRedirectError
is a typed exception so _fetch_url can map redirect blocks to terminal
'rejected' status (not transient 'failed'). _http_open is the single
call site so tests can mock at one well-defined seam.
Tests cover redirect blocking (link-local, loopback), redirect-error
unwrapping inside URLError, pinned-IP connection target, and the
end-to-end DNS-rebinding scenario. Existing tests that mocked
urllib.request.urlopen are migrated to mock _http_open.
* Harden /asset/ endpoint against stored XSS
The endpoint served any file in the cloned marketplace repo with
stdlib-detected Content-Type, so a curator who landed evil.html (or a
renamed evil.png carrying HTML bytes) in the working tree got a
same-origin XSS — the response shares cookie scope with /admin and
/api/me/*.
The asset endpoint is image-only by contract (cover photos referenced
from agnes-metadata.json + inner skill / agent cards), so applying the
same allowlist + magic-bytes pattern that /doc/ already uses closes
the gap without breaking any legitimate use case. Three layered
checks: extension in IMAGE_EXTENSIONS (.png/.jpg/.jpeg/.webp; SVG
excluded — <script> inside SVG executes), validate_image_file magic
bytes (defeats rename-extension attack), Content-Type pinned from the
validated extension (never stdlib mimetypes).
Defense-in-depth: X-Content-Type-Options: nosniff stops browser MIME
sniffing; Content-Security-Policy: default-src 'none' blocks script /
iframe execution even if a future regression let HTML through.
Tests cover the .html extension reject, the renamed-HTML-as-PNG magic-
bytes reject, the .svg reject, and the happy-path PNG with security
headers attached. The pre-existing path-traversal test seeds a real
PNG instead of ok.txt now that the endpoint is image-only.
* Enforce mandatory curator on marketplace PATCH
The POST handler enforced curator_name + curator_email at create time,
but PATCH treated empty / missing curator inputs as 'no change'. Legacy
rows that pre-date v32 (curator_name=NULL) could be edited indefinitely
without ever filling the curator gap, and OWNER_TODO_PLACEHOLDER lingered
on every /marketplace card.
Reject the PATCH with 400 when the post-merge row would persist with
empty curator. The check fires after the existing field-merge logic, so
once-filled rows that don't touch curator still pass through (their
existing values fall through from the DB row). DB column stays nullable
so untouched legacy rows continue to coexist — the gate fires only the
moment an admin opens the edit modal.
Existing PATCH semantics preserved: empty-string input still means 'leave
existing value alone', and once-filled curator can't be cleared (those
test cases pass unchanged). New test seeds a legacy row directly via the
repository, then exercises url-only PATCH (rejected), partial-fill PATCH
(rejected), and full-fill PATCH (succeeds); a follow-up no-curator PATCH
on the now-formed row also passes.
* Drop unused curated-marketplace helpers (PR #234 review)
* build_db_payload — imported by src/marketplace.py but never called.
The strict-drop semantics it would have implemented were re-written
inline in _refresh_plugin_cache (see the comment block there). The
standalone helper still carried the old fall-back-to-original-external-
URL-on-mirror-failure behaviour, which contradicts the documented
drop-when-can't-deliver contract — a future contributor who re-wired
it would have introduced a silent regression. Delete with the helper
+ the import + the comment that referenced it.
* _resolve_marketplace_name — one-line shim with no remaining call
sites. Callers use _resolve_marketplace_meta which returns name +
curator together, avoiding the double DB hit the shim exists to
hide.
* '# noqa: F401 Optional kept for forward-compat' was wrong — Optional
IS used in src/marketplace.py (line 70 and line 238). Drop the noqa
comment so a future ruff run doesn't try to remove a real import.
Removing build_db_payload also drops the only remaining use of Optional
in src/marketplace_metadata.py, so the import comes out there too.
* Cap agnes-metadata.json size + catch RecursionError on parse
The reader is invoked once per marketplace per sync and the file is
curator-controlled. Two failure modes were unguarded:
* Multi-GB JSON: path.read_text() pulled the whole file into memory
before json.loads even ran. A curator with commit access to an
upstream repo could OOM the sync worker.
* Deeply-nested JSON under any size cap: cpython's recursive object /
array parser raises RecursionError at ~1000 levels of depth.
RecursionError is a RuntimeError, not ValueError, so the existing
catch let it propagate up and abort the entire sync — every other
marketplace in the same pass got skipped.
Add AGNES_METADATA_MAX_BYTES = 1 MiB (a real metadata file with covers,
docs, categories for ~50 plugins fits in <100 KB so the cap is
generous) and gate the size check on path.stat().st_size before the
body read. Broaden the parse except to (ValueError, RecursionError)
with a unified log line. Both failure modes degrade to the same
empty-dict fall-back the malformed-JSON path already used, so one bad
upstream never aborts the rest of the sync.
Tests cover the size cap firing before json.loads (whitespace-padded
valid JSON exceeding the cap) and the recursion path (5000 nested
arrays — past cpython's default recursion limit but well under the
size cap).
* Persist asset-mirror manifest per body write, before unlink
sync_assets wrote each body atomically (tmp + rename) but persisted
the manifest only at the end of the batch. A kill -9 mid-Phase 2 left
on-disk files the manifest never referenced. Once a curator dropped
that URL from agnes-metadata.json, Phase 3's cleanup had no record of
the file and the orphan stayed forever — there's no GC pass walking
the cache dir today, so disk would slowly bloat.
Phase 2 (body-write iteration): after the in-memory manifest mutation,
persist BEFORE unlinking the previous body. The crash window narrows
from 'all of Phase 2' to 'between persist and unlink' (microseconds).
A persist failure mid-batch keeps the previous body on disk — the on-
disk manifest still references it, and a stale-but-existing file beats
a 404. Cost: one extra tmp+rename per body write; manifest is a few KB
so the overhead is negligible vs. the HTTP fetches.
Phase 3 (curator-removed URLs): same discipline. Collect the to-delete
relpaths, persist the manifest with the entries already gone, THEN
unlink. A crash mid-cleanup leaves at most a microsecond window where
files exist despite the manifest no longer naming them. The next sync
reads the (correct) manifest and the orphan stays orphaned, but the
served state is consistent.
Tests cover per-body persist call count, the post-update on-disk
manifest content, and Phase 3 ordering verified by reading the on-disk
manifest from inside Path.unlink.
* Consolidate marketplace video embeds + format-guide CSS
The YouTube nocookie / Vimeo / <video> / link-fallback detection logic
was duplicated verbatim in marketplace_plugin_detail.html and
marketplace_item_detail.html (~40 JS lines each, with subtly-different
inline styles). Both templates now {% include %} a single
_marketplace_video_embed.html partial inside their IIFE so the regex,
the nocookie attribute set, and the unknown-host link fallback live in
ONE place — future tweaks (new host, new attribute, fixed sandbox flag)
no longer need to be applied twice in lockstep.
The .video-wrap selectors (one inline <style> rule in plugin_detail,
one inline style='...' attribute in item_detail) are replaced by the
existing .video-embed 16:9 wrapper in style-custom.css, with new
.video-embed video / .video-embed a child rules added so the wrapper
handles all four embed shapes uniformly without per-template
positioning.
The 60-line inline <style> block in marketplace_format_guide.html
moves verbatim to style-custom.css under a new 'Marketplace format
guide page' section, scoped to .format-guide so other pages aren't
affected.
No user-visible behaviour change: the rendered HTML for valid
YouTube / Vimeo / mp4 / external links is byte-identical to before,
and the format-guide page renders the same.
* Maintainability cleanup batch (PR #234 review)
#10: drop _path_under from app/api/marketplace.py — it was a byte-
equivalent clone of _safe_join (same Path.resolve(strict=True) +
relative_to() containment check). The three v32 endpoint handlers
(/asset, /doc, /mirrored) now share the existing helper.
#14: rename src/marketplace_assets.py → src/marketplace_asset_validation.py
so the file's purpose is obvious from the name and the previous
overlap with src/marketplace_asset_mirror.py is gone. Six call-site
imports updated in lockstep; CHANGELOG references under [Unreleased]
updated to track the new path.
#11: consolidate the URL builders that resolve
/api/marketplace/curated/<slug>/<plugin>/{asset,doc,mirrored}/...
paths. _internal_asset_url / _internal_doc_url / _mirrored_asset_url
lived in src/marketplace.py, while a copy named _mirrored_url lived
in app/api/marketplace.py with a 'must stay aligned' comment. New
module src/marketplace_urls.py is the single source of truth — both
call sites import from it and a future URL-format tweak only needs
to change one file. The _ROUTE_PREFIX constant collapses the per-
function f-string repetition. The route-handler endpoints themselves
still own the path string literals (keeping the builders identical
to the route declarations remains a checklist item, not a runtime
guarantee).
* Re-key asset-mirror manifest by (plugin, url) + dedup HTTP fetches
The manifest used to be keyed by URL alone, so two plugins in the
same marketplace referencing the same external image (a shared CDN
icon, a common cover) collided on entry.plugin_name — last writer
won. The DB row for the losing plugin then stored a served URL
pointing under the winning plugin's tree, and require_resource_access
denied legitimate access on one side and let the other plugin's user
reach the wrong asset.
In-memory: Dict[Tuple[str, str], MirrorEntry] keyed (plugin_name, url).
On disk: format flips from {url: entry} dict to [entry, ...] list of
self-describing entries (each carries plugin_name + url + the
previous fields). JSON keys can't be tuples; encoding 'plugin::url'
would just shift the parsing burden.
Phase 1 of sync_assets deduplicates fetches by URL — three plugins
sharing one URL share one HTTP request. The conditional-GET prior is
picked from any owning plugin's prior entry; if their etags diverge
(rare) we miss one 304 and pay for a full re-download instead.
Phase 2 still creates a per-(plugin, url) manifest entry pointing
under the plugin's own subdir, and Phase 3 cleanup is keyed the same
way so dropping a URL from one plugin's metadata doesn't disturb
another plugin still referencing it.
Body files stay per plugin (RBAC-clean isolation: deleting plugin A's
cache can't strand plugin B). Bandwidth saved by fetch dedup.
Consumer code re-keyed: src.marketplace._refresh_plugin_cache rebuilt
served_url_for / mirror_status as composite-keyed maps;
app.api.marketplace._resolve_external_via_mirror /
_curated_inner_cover / _curated_inner_enrichment look up by
(plugin_name, url).
Tests cover per-plugin manifest entries with shared URL, the single
HTTP fetch for N plugins, and Phase 3 drop-one-keep-other. All
existing tests migrated to composite key access; v2 list format
assertions verify on-disk shape.
* Migrate asset mirror from urllib.request to httpx
The asset mirror was the only HTTP call site in Agnes still using
urllib.request; every other module (CLI, Jira / OpenMetadata / OpenAI
connectors, scheduler, Telegram bot) already used httpx. The asset
mirror was added in this PR's base commit, so this is the only chance
to bring it into convention before someone copies it as 'the pattern
for HTTP fetches in Agnes'.
Three concrete benefits beyond consistency:
* SSRF defence collapses from five urllib classes
(_PinnedHTTPConnection, _PinnedHTTPSConnection, _PinnedHTTPHandler,
_PinnedHTTPSHandler, _SafeRedirectHandler) into one
_SSRFGuardTransport. httpx invokes handle_request() on every redirect
hop, so re-validation is free — we don't need a custom redirect
handler at all.
* DNS-rebinding defence: the transport rewrites request.url.host to the
SSRF-validated IP before delegating to super().handle_request().
httpcore connects to whatever URL.host says, so this pins the
connection without subclassing HTTPSConnection. The original hostname
goes into the Host header + the sni_hostname extension so TLS / vhost
routing still bind to the curator-supplied hostname.
* Error handling: one httpx.HTTPError catch-all for transport errors,
plus specific httpx.TimeoutException / httpx.TooManyRedirects branches
for clearer diagnostics. Matches the _translate_transport_error shape
in cli/client.py.
The shared httpx.Client is built lazily at module load (same pattern as
cli/client.py:_get_shared_client) with follow_redirects=True,
max_redirects=5, timeout=HTTP_TIMEOUT_SEC, and our custom transport.
Externally observable behaviour is unchanged: same FetchOutcome
statuses, same manifest format, same conditional GET semantics, same
body-size cap.
Tests migrated from urllib-shaped fakes to httpx-shaped (status_code,
iter_bytes, context manager). Five urllib-specific tests replaced with
httpx equivalents — three transport unit tests + one DNS-rebinding
integration test that verifies host rewrite via monkey-patched
super().handle_request. One test deleted without replacement
(unwrap-URLError-wrapping-an-_UnsafeRedirectError — urllib-specific,
not applicable to httpx).
* Surface curated agnes-metadata enrichment on My Stack tab
GET /api/marketplace/items?tab=my built each curated row from the
on-disk marketplace.json by way of resolve_allowed_plugins, which
doesn't carry the agnes-metadata enrichment columns
(cover_photo_url, video_url, category override, doc_links). The
handler then hard-coded cover_photo_url=None on the synthetic row.
Result: once a user clicked '+ Add to my stack' on a curated card,
the same plugin in tab=my rendered with the gradient placeholder
instead of its cover photo — confusing parity break vs. the curated
tab where the same row goes through MarketplacePluginsRepository
and gets the enriched columns.
Pre-load the enriched marketplace_plugins rows for every marketplace
the user is subscribed to, then look each granted+subscribed plugin
up by (marketplace_id, plugin_name). Fall back to the on-disk
synthetic shape only when the DB row is missing — happens during
the rare race where RBAC is granted before the first sync cycle
ingests the plugin. RBAC gating (granted set from
resolve_allowed_plugins) is unchanged so this fix can't widen
visibility; it just upgrades the data shape behind cards the user
was already going to see.
Per-marketplace list_for_marketplace beats N gets — typical user is
subscribed to <5 marketplaces, so this is at most a handful of
queries vs. one per subscribed plugin.
Regression test seeds a plugin with cover_photo_url + category
override, subscribes the user, hits /api/marketplace/items?tab=my,
and asserts photo_url + category come through. The misleading
'fall through to gradient until the user re-visits the curated tab'
comment is gone.
---------
Co-authored-by: Minas Arustamyan <arustamyan.minas@gmail.com>
The original list-form _V34_TO_V35_MIGRATIONS ran four ALTER
statements in sequence:
ADD _vis_v35 → UPDATE _vis_v35 = visibility_status →
DROP visibility_status → RENAME _vis_v35 TO visibility_status
If the RENAME failed for any reason after the DROP succeeded — DuckDB
lock contention at startup, scheduler-vs-app race opening
system.duckdb, container kill mid-migration, etc. — the DB was
stranded with _vis_v35 populated and visibility_status missing. The
schema_version row never bumped because the UPDATE at the bottom of
the migration ladder runs only when every step succeeded. Subsequent
restarts then hit DROP visibility_status again with no IF EXISTS
guard and looped on the same error; the only recovery was hand-
editing the DB.
Replace the list with a Python function _v34_to_v35_migrate that
inspects the table's columns up front and dispatches into one of
three paths:
* clean v34 (visibility_status present, _vis_v35 absent) — run the
full rebuild
* partial v35 (_vis_v35 present, visibility_status absent) — finish
the RENAME alone, data is already in _vis_v35 from the prior
UPDATE
* both columns present (rare; aborted before DROP) — drop the temp
and keep visibility_status
The audit columns (archived_at, archived_by) ship first behind
IF NOT EXISTS so they're safe in all states. Operators stranded by
the original bug now recover automatically on next startup.
Tests cover the three direct paths plus an end-to-end scenario where
_ensure_schema walks a schema_version=32 DB with the half-applied
state up through to v36.
Co-authored-by: Minas Arustamyan <arustamyan.minas@gmail.com>
* feat(store): flea-market upload guardrails + soft delete + JOIN-based admin queue
Adds an end-to-end guardrails pipeline for store uploads (manifest +
static-security + LLM review), persists blocked bundles for forensics,
introduces soft-delete (Archive) semantics, consolidates the legacy
/store/{id} surface into /marketplace/flea/{id}, and reworks the admin
queue so lifecycle filters read live entity visibility via LEFT JOIN
rather than a denormalized submission column.
Schema v29 → v35:
* v29 store_submissions table + store_entities.visibility_status
* v30 file_size, bundle_sha256, bundle_purged_at on submissions
* v31 reshape store_submissions (drop legacy unique on entity_id)
* v32 store_entities.archived_at/by + 'archived' visibility value
* v33 drop store_submissions.retry_count (unused)
* v34 ensure idx_store_submissions_entity exists post column-drop
* v35 broaden visibility_status enum + JOIN architecture cutover
Pipeline (src/store_guardrails/):
* Inline checks: manifest_check, static_scan, quality_check
* LLM review configurable haiku|sonnet|opus (default haiku)
* BackgroundTasks-driven async path with structured-output JSON
* Per-submitter daily quota (default 50)
* 30-day TTL purge job (POST /api/admin/run-blocked-purge)
* Bundle SHA256 + size persisted; sha256 survives purge for forensics
Visibility model:
* pending | approved | hidden | archived
* _enforce_visibility returns 404 (no leak) for non-owner non-admin
* Owner sees own non-approved entries via include_owner_id widening
* Install refused with 409 entity_not_approved when not approved
Soft-delete (DELETE /api/store/entities/{id}):
* Default = soft (visibility_status='archived'); existing installs
keep getting served the bundle so users don't lose the plugin
* ?hard=true admin-only: drops bundle + cascades user_store_installs
* Hard-delete preserves entity_id on submission as tombstone so
audit_log linkage survives for the activity timeline
Admin queue lifecycle (the JOIN refactor):
* Verdict (store_submissions.status) is immutable forensic record
* Lifecycle (store_entities.visibility_status) is live state
* /admin/store/submissions Archived chip translates to
`e.visibility_status='archived'` via LEFT JOIN — any path that
flips visibility surfaces in the queue immediately
* Detail page renders Status (verdict) and Entity lifecycle side by
side so admins see "approved at review, now archived" at a glance
URL consolidation:
* /store/{id} deleted (no redirect, stale bookmarks 404)
* /marketplace/flea/{id} is the canonical detail surface
* Three in-tree callers (upload-success, my-stack card, store
listing card) updated to point at the new URL
* Quarantine banner extracted to _quarantine_banner.html partial,
self-guarded, included from both flea detail templates
* Banner JS auto-refreshes when the verdict lands by polling
/api/marketplace/flea/{id}/detail (visibility_status +
submission_status — the latter is needed because blocked_llm
keeps the entity at visibility_status='pending')
Audit log resource format:
* runner.py emits prefixed `store_submission:{id}` (post-fix)
* Detail-page timeline query handles three patterns: prefixed
submission, helper-emitted `store_entity:{sub_id}`, and bare-id
legacy rows — all surface in the activity timeline
UX fixes:
* Owner sees Under review / Quarantined / Hidden banner with status
* Install button gray-disabled (not blue) when non-approved
* Owner cannot delete quarantined entries (403); admin can
* Admin queue: filter chips, sortable columns, paging, page-size
* Auto-refresh queue every 5s while pending rows are visible
* Store upload page file picker no longer opens twice (label →
input default action collided with explicit JS handler)
Tests: 168 passed across the guardrails suites (admin submissions,
store API, inline / LLM / purge guardrails, store repositories,
marketplace filter, schema version). New regression coverage
includes: archive surfaces via JOIN even when API path is bypassed;
deleted submission renders activity timeline (tombstone); flea
detail surfaces submission_status only for owner/admin; detail page
renders Entity lifecycle row; audit log resource format covers both
helper and runner paths.
* fix(store-guardrails): PR #233 follow-up — prompt injection, atomic PUT, BG race, schema, reaper, sort whitelist
Addresses 9 of the 23 findings from the PR #233 review (spec at
docs/superpowers/specs/2026-05-09-pr233-guardrails-fixes-spec.md).
Merge-gate items #1-#6 plus high-value mediums #7, #9-#12, #23.
Architectural items (#8 enum split, #14 factory) and pure
maintainability (#15-#22) deferred to follow-ups.
Security:
* #1 prompt injection — SYSTEM_PROMPT now passed via the SDK's
dedicated system= parameter; bundle wrapped in <bundle>...</bundle>
sentinels declared data-only by the system prompt; literal
sentinel strings in user content are escaped so an adversarial
README can't forge a close tag.
* #6 static scan honesty — module docstring + admin copy + docs
declare static scan as signal not gate; .md/.txt/.rst/.html/.json/
.yaml/.yml/.toml skipped to avoid false positives on prose.
AST mode for Python deferred (separate flag, FP comparison work).
Correctness:
* #2 PUT atomicity — bundles bake into plugin.staging-<rand>/
alongside live, atomic-rename on success; failed checks leave
live tree byte-for-byte intact.
* #3 BG-task race — set_visibility_if_pending guards verdict flips
to the (pending, hidden) review window; admin archives during
review survive; skipped flips audit-logged.
* #4 v35 NOT NULL/DEFAULT — schema v35→v36 re-applies them on
store_entities.visibility_status. CHECK constraint enforced
application-side (DuckDB ADD CHECK on existing column unsupported).
* #7 stuck-review reaper — reap_stuck_llm_reviews flips pending_llm
rows older than guardrails.stuck_review_grace_seconds (default
1800) to review_error. Scheduler runs every 15 min via new
/api/admin/run-reap-stuck-reviews. Set knob to 0 to disable.
* #9 quota counter — count_blocked_for_submitter_since now counts
blocked_inline + blocked_llm + review_error so a submitter
triggering only LLM-blocked verdicts is bounded.
* #10 missing risk_level — surfaces as review_error with
error='missing_risk_level' instead of silently defaulting to
'medium' (which looked like a model-decided block).
* #11 archived_at clear — set_visibility nulls archived_at +
archived_by when transitioning out of 'archived' so a future
read doesn't show stale archive forensics on an approved row.
Maintainability:
* #12 FSM doc comment — accurate insert/transition/lifecycle
description in src/db.py near store_submissions schema.
* #23 sort-key whitelist — admin queue rejects unknown sort keys
with 400 invalid_sort_key; substring-replace footgun removed.
Deferred (separate PRs):
* #5 quota race — proper fix requires asyncio.Lock spanning the
full pipeline; threading.Lock blocks event loop, DuckDB MVCC
doesn't help. API-level slowapi bounds worst case for now.
* #6 part 3 (AST static scan), #8 (enum split), #13 (import
bundle docs), #14 (factory consolidation), #15-#22 (maint).
Tests:
* New: tests/test_store_guardrails_prompt_injection.py (corpus +
trust-boundary invariants), tests/test_store_put_atomic.py,
tests/test_store_guardrails_reaper.py.
* Extended: test_store_guardrails_llm.py (system param, missing
risk_level, BG race), test_admin_store_submissions.py (quota
counter widening, sort whitelist 400), test_store_repositories.py
(un-archive metadata clear), test_db_schema_version.py (v36).
* Full suite: 3738 passed; 17 pre-existing baseline failures
unchanged (db migration tests, cli binary rename, catalog export,
user mgmt v5 backfill — confirmed by stash + rerun on clean tree).
* Extract session pipeline framework, refactor verification, add UsageProcessor skeleton
Pluggable framework under services/session_pipeline/ (contract + lib + per-processor
runner) so multiple processors can read /data/user_sessions/<key>/*.jsonl on their
own cadence with full failure isolation. Verification flow becomes the first plugin;
a no-op UsageProcessor reserves the second slot pending a separate brainstorm on
extraction logic + storage shape.
Schema v28→v29: rename session_extraction_state → session_processor_state with
composite PK (processor_name, session_file). Existing rows copied over with
processor_name='verification'; legacy table dropped. Migration is idempotent and
no-ops the copy step on fresh installs that came up at the new schema.
Endpoint: /api/admin/run-verification-detector replaced by parametrized
/api/admin/run-session-processor?processor=<name>. Audit action format follows.
Scheduler JOBS: verification-detector entry split into session-processor:verification
+ session-processor:usage. SCHEDULER_VERIFICATION_DETECTOR_INTERVAL retained for
operator compatibility (drives both cadence and health-check grace window);
SCHEDULER_USAGE_PROCESSOR_INTERVAL added.
* Address PR #232 review: scan dead branch + per-processor lock
- `SessionProcessorStateRepository.scan_unprocessed_for` dead else: both
branches surfaced every jsonl, the SELECT was unused, runner MD5-rehashed
every stable session per tick. Replaced with an mtime precheck — stable
sessions (mtime <= processed_at) are filtered at scan; modified files
still surface for the runner's authoritative `file_hash` invalidation.
Naive-local comparison matches the existing health-check idiom (DuckDB
TIMESTAMP strips tz on storage).
- Per-processor advisory lock around `_run_processor` in
`/api/admin/run-session-processor`. Scheduler tick + manual admin POST
could otherwise both run, both call create_evidence on overlapping
detections, and accumulate duplicate verification_evidence rows (the
dedup short-circuit only covers create+contradiction, not evidence per
ADR Decision 3). Non-blocking acquire → 409 Conflict on concurrent
invocation; release in finally so a runner exception doesn't wedge the
processor.
Tests: two new scan unit tests (mtime filter + post-mark mtime bump), 409
endpoint test, lock-released-on-exception test. Two existing tests updated
for the new "filtered at scan" stat shape (previously asserted skipped == 1,
now scanned == 0).
* Address PR #232 review #2: parallel scheduler tick + last_run on terminal state
Two pre-existing scaffold bugs in services/scheduler/__main__.py amplified
by adding more session-pipeline jobs:
1. Serial for-loop over jobs with synchronous httpx.post(timeout=900) — a
10-minute verification run blocked every other job (data-refresh,
health-check, usage, corporate-memory) for the whole window. The PR's
stated isolation guarantee held inside the runner but broke at the
scheduler dispatch layer.
2. last_run advanced only when _call_api returned True. Permanent-failure
jobs hot-looped on every tick (30s) instead of cadence (15min).
Fix: ThreadPoolExecutor.submit per due job + per-job in_flight set so a
long-running job can't be re-launched on subsequent ticks. last_run
advances unconditionally in finally; errors still surface via _call_api
logging + audit_log on the receiving side.
_run_job extracted to module-level for unit testing. New tests:
- TestRunJobBookkeeping: advances on success / failure / unhandled raise
- TestRunLoopParallelism: in_flight protection prevents duplicate
launches across ticks for a single slow job
---------
Co-authored-by: Minas Arustamyan <arustamyan.minas@gmail.com>
* feat(home+news): state-aware /home + /news + admin-edited news section
Squash of the vr/home-page feature work for clean rebase onto main.
Original 18-commit history preserved in branch backup/vr-home-page-pre-rebase.
What's in this PR:
**State-aware /home page**
- New `/home` route with hero + auto-mode + connectors (Asana / GWS /
Atlassian) + lookarounds. Onboarded vs not-onboarded state-machine
branches a single template (`home_not_onboarded.html`); the install
steps, "Setup a new Claude Code" CTA (90-day PAT mint), and per-
connector setup prompts hide once `users.onboarded=TRUE`. A
completion badge replaces them.
- "Mark me as offboarded" button reverses the flag without an SQL UPDATE.
- `users.onboarded BOOLEAN` column added; default FALSE; flipped by the
CLI's `agnes init` post-success POST and the `/admin/users` API.
- Connector setup prompts pre-check whether the tool is already
installed/connected before re-running setup.
- GWS scope set widened to include Google Chat (`chat.spaces`,
`chat.messages`).
**Single template + design tokens**
- `dashboard.html` now extends `base.html` via the new
`{% block layout %}` opt-out (full-width pages skip the 800px
`.container`). Net: every page shares one shell.
- `style-custom.css` `:root` extended with `--space-{7,9,10,12}`,
`--radius-2xl`, `--shadow-{card,elevated}`, `--text-{muted,disabled}`,
`--focus-ring`, `--transition-*`, `--width-{narrow,app,wide}` so
inline page styles can migrate incrementally.
**Auth redirects honor AGNES_HOME_ROUTE**
- `safe_next_path` resolves the configured home route when no `default=`
is passed; OAuth callbacks, magic-link clicks, password form, and
LOCAL_DEV_MODE shortcuts now land on `/home` (or whatever the operator
picked) instead of always /dashboard.
**News section + /news permalink + /admin/news editor**
- Schema-bumped `news_template` table (single versioned entity, draft +
publish gate). `published BOOLEAN` distinguishes draft from public;
monotonically-increasing `version` per save; rows >30d pruned on
save except the currently-displayed published version.
- `/home` bottom-of-page renders the latest published intro with a
"Read more →" link to `/news` (which renders the full body).
- `/admin/news` editor with sandboxed live preview, versions table,
per-row Unpublish, Format-help cheatsheet.
- `agnes admin news show / draft / edit / publish / unpublish /
versions / export` (CLI). Talks to the live server via the
`/api/admin/news/*` endpoints (PAT-authed) — no direct DB access
so it coexists with a running uvicorn.
- **Optimistic-lock guard**: `agnes admin news publish --version N` and
PUT/PATCH endpoints accept `expected_version` and 409 with structured
`{error: "version_conflict", expected, actual, actual_by}` when a
concurrent admin replaced the draft. Edit refuses to overwrite a
draft authored by someone else without `--force` or
`--expect-version`.
- nh3 (Rust-backed ammonia) HTML sanitizer; iframe pre-pass strips
any iframe whose src is not on the YouTube/Vimeo/Loom allowlist;
javascript:/data: schemes blocked everywhere.
- Author CSS vocabulary: `.news-hero` (blue gradient hero block),
`.callout`/`.callout-{info,warn,success,danger}`,
`.video-embed`, `.news-section`, `.news-grid-{2,3}`, `.news-cta` —
all consolidated in `style-custom.css` under "News content
vocabulary (shared)" so /home perex, /news body, and /admin/news
preview share one source of styling.
- Code-inside-`<pre>` contrast fix (was unreadable amber-on-silver).
- `.news-content` table styling (border, header band, row-hover).
**`scripts/dev/run-local.sh`** — local uvicorn launcher. Pulls Google
OAuth client id/secret from GCP Secret Manager
(`AGNES_OAUTH_GCP_PROJECT`-driven, no vendor defaults), points
`AGNES_CLI_DIST_DIR` at `./dist` so the wheel endpoint resolves, and
`--dev` flips `LOCAL_DEV_MODE=1` + `AGNES_HOME_ROUTE=/home` for one-
command iteration. `LOCAL_DEV_MODE=1` also enables the FastAPI debug
toolbar.
**CLAUDE.md "Run tests before every push" section** codifies
`pytest tests/ -n auto -q` as non-negotiable before each push.
**Tests**: 51 + 14 + 8 = 73 new tests across news-template repo,
sanitizer, API, web, CLI; plus updated home/auth/template tests for
the new shared-shell architecture.
Origin docs (gitignored, customer-fork content):
docs/brainstorms/home-page-requirements.md,
docs/plans/2026-05-07-001-feat-home-page-plan.md.
* feat(cli): agnes onboarded {on,off,status} — self-scoped flag toggle
User-facing equivalent of the in-page "Mark me as (off)boarded" button
on /home. POSTs /api/me/onboarded with {onboarded, source}; --source
overrides the audit-log marker so flips made from the CLI vs the web
button vs agnes init automation stay distinguishable.
`status` reads via /api/me/profile (when present); falls back to a
quick body-marker scan of /home so the read path doesn't write an
audit_log row. PAT-authed via cli.client.api_post — same convention
as agnes admin news / agnes admin add-user etc.
Tests: 5 covering on/off/status round-trip, idempotency, and
audit-log source recording. Full suite holds at 12 pre-existing
failures (same set as before).
* ui(nav+home): primary nav reorg + green What's new band + /marketplace link fix
Primary nav (post-rebase audit + per-user feedback):
- Items: Home → Marketplace → Data Packages → Memory. Admin dropdown
for admins only. The "Dashboard" label was renamed Home — point still
resolves through `home_route` so customer instances on /dashboard
still land there.
- Activity Center moved into the Admin dropdown. Per-team adoption
analytics is admin-consumed in practice; the route still allows
any authed user for direct deep-links so existing /home tile +
bookmarks keep working.
- Memory link added (→ /corporate-memory) — was previously buried in
the /home "Look around" tiles.
- Setup local agent + My Stack dropped from main nav. Setup is the
/home install flow's home now; My Stack lives as a tab inside
/marketplace.
/home tweaks:
- Plugin marketplace tile now points at /marketplace (was /store —
legacy from before the marketplace rebrand landed in #230).
- "What's new" section header gets a green band (success-flavored
D1FAE5 background, A7F3D0 border, darker green title) so the
bottom-of-page news block visibly distinguishes from the blue
install-hero at the top. Header strip only — body stays white.
Test fix: test_home_route_resolution renamed `dashboard_link_uses_home_route`
→ `home_link_uses_home_route` and asserts `href="/home">Home` instead
of `href="/home">Dashboard` after the label change.
* fix(home): decouple Step 3 + Connect-tools collapse from server onboarded flag
The server-side `users.onboarded` flip happens through two paths:
1. Explicit user click on "Mark me as onboarded" or `agnes onboarded on`.
2. Implicit `agnes init` POST → /api/me/onboarded on success.
Path 2 produced a UX surprise: an analyst running `agnes init` mid-flow
reloaded /home and saw Step 3 (auto-mode) + Connect-your-tools auto-
collapse to summary bars. They were actively working through those
sections — the install POST never signalled "I'm done with the rest
of setup", just "Agnes itself is installed".
Decouple the section-collapse decision from the server flag:
- Step 1 + Step 2 install blocks: still hidden on `onboarded=TRUE`
(their completion is a hard server signal — Agnes IS installed).
- Step 3 + Connect-your-tools: render flat by default in BOTH states.
Wrapped in `<details class="setup-collapsible" open>` so the
browser's native disclosure handles per-section toggle without JS,
but the `<summary>` is CSS-hidden until the page-level
`data-setup-minimized="1"` attribute is set on `.home-mock`.
- New "Minimize setup view" toggle inside the blue install-hero,
rendered only when onboarded. Click flips the data-attr on
`.home-mock` AND removes the `open` attribute from each
`<details>`. State persists in `localStorage["agnes_home_setup_minimized"]`
so the choice survives reloads but is per-device.
- "Show full setup view" (the same button when minimized) re-opens
both `<details>` and clears localStorage.
When minimized, each `<details>` still has its own native expand/
collapse — click the gray summary bar to peek at one section without
toggling the page-level minimize off.
Tests:
- test_step3_and_connectors_render_flat_when_onboarded_by_default —
asserts `<details class="setup-collapsible" ... open>` for both
sections post-onboarding and the absence of any server-rendered
`data-setup-minimized` attribute on the `.home-mock` root.
- test_minimize_toggle_visible_only_when_onboarded — toggle button
rendered only when onboarded.
Full pytest holds at 12 pre-existing failures (same set).
* feat(observability): optional PostHog integration (errors, LLM traces, replay, flags)
Off by default. Activates when POSTHOG_API_KEY is set in env. Defaults
to PostHog Cloud EU; override host for US Cloud or self-hosted.
Coverage:
- FastAPI 500 handler captures unhandled exceptions
- src/orchestrator.py rebuild + rebuild_source failures
- services/scheduler/ HTTP-job failures
- cli/main.py uncaught CLI errors (Typer.Exit/SystemExit/KeyboardInterrupt
skipped; flushes before re-raise so short-lived CLI invocations don't
drop events)
- connectors/llm/anthropic_provider.py + openai_compat.py emit
$ai_generation events with provider, model, latency, token counts
(prompt/completion bodies stay off unless POSTHOG_LLM_PAYLOADS=1
because LLM prompts here routinely include customer SQL/data)
- Browser snippet injected into every text/html response by
PosthogInjectionMiddleware — registered inside the GZip layer so it
sees uncompressed HTML before compression. Many templates are
standalone (their own DOCTYPE) and never extend base.html, so a
per-template include would miss them.
- Frontend: $pageview, $pageleave, JS error capture via window.error
and unhandledrejection handlers, masked session replay
(maskAllInputs: true plus CSS-selector mask for known data surfaces),
feature flags (browser posthog.isFeatureEnabled + server-side
feature_enabled with fallback for older SDKs).
Identification mode operator-configurable: none / id / email / full.
Default email ships user.id + email but never name. CLI entry point
moves from cli.main:app to cli.main:main (Typer wrapper).
Files:
- src/observability/posthog_client.py — lazy singleton, no network
when disabled, single-process flush on shutdown
- src/observability/llm_tracing.py — trace_generation context manager
- app/middleware/posthog_inject.py — HTML rewrite middleware
- app/web/templates/_posthog.html — browser snippet template
- docs/observability.md — operator guide
- config/.env.template — documented POSTHOG_* knobs
- tests/test_posthog_disabled.py + tests/test_posthog_client.py +
tests/test_llm_tracing.py — 18 tests covering disabled state,
identify-mode payloads, $ai_generation shape, error variant.
CHANGELOG entry under [Unreleased] Added.
* feat(observability): tag every PostHog event with environment + release
Splits PostHog dashboards cleanly between localhost / dev / staging /
production without manual tagging on every capture call.
- POSTHOG_ENVIRONMENT explicit override; auto-resolves to "local" when
LOCAL_DEV_MODE=1, else RELEASE_CHANNEL, else AGNES_DEPLOYMENT_ENV,
else "unknown".
- AGNES_VERSION → RELEASE_CHANNEL fallback feeds the `release` property
for "is this error new in this release?" cohorting.
- Backend gets both via the PostHog SDK's super_properties constructor
arg (every captured event picks them up automatically).
- Browser snippet calls posthog.register({environment, release}) inside
the loaded callback so $pageview, $exception, autocapture, etc. all
carry the same labels.
- request.state.user now populated by auth dependencies so the snippet
can actually call posthog.identify(user_id, {email}) for logged-in
users (previously the user block always resolved to None because
nothing wrote to request.state.user).
4 new tests cover env resolution: explicit > LOCAL_DEV_MODE > channel
> unknown, plus super-properties forwarding into the SDK constructor.
* feat(observability): inline user attrs on every PostHog event + debug throw route
PostHog's UI shows person properties on the Person profile page, not
inline on each event — so a reviewer triaging an exception couldn't tell
which user hit the bug without clicking through. Fix it on both sides.
- Backend capture_exception merges user_id / user_email / user_name into
the event properties (gated by POSTHOG_IDENTIFY_PII: none/id/email/full).
Backed by a new _user_props_for_event helper on PosthogClient.
- Browser snippet registers user_id + user_email + user_name as super-
properties via posthog.register({...}) so every $exception, $pageview,
and custom event coming from posthog.captureException() carries them
inline. Mirrors the backend so cross-referencing client/server events
doesn't require a person-profile lookup.
- /api/debug/throw — debug-only endpoint gated by DEBUG=1 (404 in prod).
Runs Depends(get_current_user) first so request.state.user is set when
the unhandled-exception handler captures the event. Lets operators
exercise the full observability path end-to-end without hand-rolling
a TestClient script. Configurable via ?kind=ValueError&msg=...
7 new tests cover: backend user-attr merge across identify modes,
anonymous request fall-through, browser snippet super-prop emission for
logged-in / anonymous / id-only / full-name cases.
* fix(observability): address minasarustamyan PR #231 review
Two bugs caught in review.
1. PosthogInjectionMiddleware dropped Response.background on every
return path. BaseHTTPMiddleware materialises the body and asks
subclasses to return a fresh Response — three paths in dispatch()
omitted background=, silently cancelling any BackgroundTask /
BackgroundTasks the route attached (audit logging, async webhooks,
email sends) with no log line. Fix: route every return through a
_passthrough() helper that forwards background.
Also adds a _MAX_BUFFER_BYTES (4 MB) cap so a streamed-HTML response
can't balloon RSS during buffering. Bigger bodies short-circuit
through with a warning rather than being injected.
Regression tests in tests/test_posthog_inject_middleware.py exercise
four return paths (snippet present, render-fail, double-injection
guard, non-HTML passthrough) plus the streaming-guard short-circuit.
2. $ai_input / $ai_output_choices were emitted without truncation, so
POSTHOG_LLM_PAYLOADS=1 silently dropped events past PostHog's ~32 KB
per-event ingest limit — exactly the calls (large prompts with
schemas / sample rows / SQL) an operator would want to inspect.
Fix: clip both at POSTHOG_LLM_PAYLOAD_MAX_CHARS (default 30000) with
an explicit "…[truncated N chars]" marker so readers don't mistake
truncated captures for complete ones. Metadata (provider, model,
tokens, latency, error) flows regardless. Three new tests cover
default-cap clipping, env-override, and pass-through under the cap.
37 PostHog tests pass.
* Add /marketplace browse page + Model B opt-in stack composition
New /marketplace browse surface unifies the curated marketplaces
(admin-managed git mirrors) and the community Flea Market behind
three tabs — Curated / Flea / My Stack — with per-tab category
filter, search across both sources with scope checkboxes, and
numeric pagination, all driven by URL query state. Plugin detail
at /marketplace/curated/<slug>/<plugin> and /marketplace/flea/<id>;
nested skill / agent detail at /marketplace/curated/<slug>/<plugin>/
{skill,agent}/<name> and the flea-side single-page detail.
Model B opt-in: an RBAC grant on a curated plugin is now only
*eligibility*. The user must click "Add to my stack" for it to
enter their served Claude Code marketplace. Composition flips
from (rbac ∖ opt_outs) ∪ store_installs to
(rbac ∩ subscriptions) ∪ store_installs. The legacy
user_plugin_optouts table is renamed user_curated_subscriptions
(schema v27) — same table shape, inverted semantic, repository
methods become subscribe / unsubscribe / is_subscribed.
UX vocabulary: Install → Add to my stack, Installed → In your
stack, card "Installed" badge → "In stack" (amber pill), tab
"My Subscriptions" → "My Stack". Bridges the two-step model
(server-side bookmark vs. on-laptop install) the previous label
hid. Click triggers an inline post-add hint panel under the
description with the agnes refresh-marketplace recipe + Copy
chip, dismissible per-browser via localStorage.
Per-tab info blocks above the filter row:
- Curated: trust signal — "Each plugin here has a named curator
accountable for it." (blue accent + See-all-curators link)
- Flea: open-shelf signal — "Anyone in the company can upload
here." (purple accent + Tips-for-sharing link)
- My Stack: personal-shelf orientation — "Your AI stack —
everything you've added." (slate accent, no link)
Tabs carry per-tab Heroicons (shield-check / building-storefront
/ rectangle-stack) tinted to match each tab's accent; flips white
when the tab is active for contrast.
Hero illustration anchored to the right of the blue hero panel
(absolute, 47% wide, behind the search row content). Hidden
under 900px viewport.
Action-row CTAs realigned to publication intent: curated
"How to add new content" → "Submit a plugin" (links to the
guide page); flea button removed since +Upload sits next to it.
Empty-state CTAs match. /marketplace/guide/{curated,flea}
routes now host publication-flow guide pages with placeholder
ledes — full copy to be authored separately.
Categories: Heroicons-based icons mapped per category in
src/category_icons.py (zero new dependencies; SVG path strings
inlined). Marketplace cards, filter pills, and detail pages
read from the same source.
API endpoints under /api/marketplace:
- GET /items per-tab listing (curated / flea / my)
- GET /categories per-tab non-zero counts
- GET /curated/{slug}/{plugin} plugin detail
- POST/DELETE /curated/{slug}/{plugin}/install subscribe toggle
- GET /curated/{slug}/{plugin}/{skill,agent}/{name} inner item
The tab=my branch reads directly from
user_curated_subscriptions ∪ user_store_installs (not
resolve_user_marketplace, which bundles flea skills/agents into
a single store-bundle synthetic entry useful for serving the
Claude Code marketplace ZIP/git but wrong for browsing where
each item should appear as its own card).
Detail pages: plugin detail surfaces inner skills/agents as
clickable nested cards; commands/hooks/MCPs render as plain
name lists. Skill/agent detail mirrors the plugin layout with
kind-tinted accents (skill = green, agent = purple), Description
+ Details sidebar, Files + Docs sections, and the "How to call
it" copy-able invocation chip showing /<plugin>:<inner-name>
exactly as Claude Code namespaces it post-install. Curated
nested has no install button — links back to the parent plugin.
Navbar: standalone "My AI Stack" relabelled "My Stack" and
points at /marketplace?tab=my; "Store" link removed (Store
flow is reachable via the Flea Market tab's +Upload button).
The standalone /my-ai-stack and /store routes still work for
old bookmarks.
Tests cover the new browse / categories / install / RBAC paths
under tests/test_marketplace_api.py; existing marketplace and
store tests updated for Model B (explicit subscribe in fixtures).
Schema bumped v26 → v27 with idempotent migration that wipes
existing user_plugin_optouts rows on flip and adds
marketplace_plugins.created_at with registered_at backfill.
* Fix v28 migration + post-rebase test fallout
v28 ALTER TABLE marketplace_plugins ADD COLUMN created_at conflicted with
_SYSTEM_SCHEMA's earlier CREATE that already includes the column on fresh
installs (test fixtures starting at any pre-v28 version trip on it).
Switch to ADD COLUMN IF NOT EXISTS — same idiom as the upstream v27
Keboola sync-strategy migration on the same ladder.
Two test patches needed after the rebase bumped SCHEMA_VERSION 27 → 28:
- test_keboola_v27_migration.py: test_schema_version_constant_is_27 was
pinning ==27. Loosened to >=27 (the test's purpose is to verify the
v27 Keboola migration, not to pin the current SCHEMA_VERSION).
- test_setup_page_unified.py: was monkeypatching resolve_allowed_plugins
but compute_default_agent_prompt now reads from resolve_user_marketplace
(Model B-aware). Stub the right function so the test exercises the
v28 served-set path.
* Harden curated skill/agent inner endpoints against path traversal
`_read_inner`, the `skill_dir` walk in `curated_skill_detail`, and the
`agent_path.stat` in `curated_agent_detail` joined URL path-params onto
`plugin_root` without verifying the resolved candidate stayed inside it.
Starlette's `[^/]+` on `{skill_name}` / `{agent_name}` blocks the direct
URL exploit (encoded `/` 404s before the handler), but a curator-planted
symlink inside a curated marketplace's git mirror could still dereference
outside the plugin tree on read.
Adds `_safe_join(plugin_root, *parts)` doing
`Path.resolve(strict=True)` + `relative_to(plugin_root.resolve())`, used
by all three call sites so the boundary is enforced once and consistently.
Tests cover the helper directly (normal path resolves, escaping `..`
returns None, escaping symlink returns None, missing file returns None)
plus an end-to-end check that the symlink case actually 404s on the
HTTP endpoint. Symlink tests skip on Windows where symlink creation
needs elevated permissions; they run on Linux CI.
---------
Co-authored-by: Minas Arustamyan <arustamyan.minas@gmail.com>
## Summary
Two minimum-viable fixes after today's 0.44.0 → 0.47.3 release train and the production 30-user launch. Devil's advocate review of a 3-PR / 7-item plan cut scope to these 2 — the rest is deferred to a separate "operate-first, instrument-second" backlog item.
### B2 — Docker session_collector log skip
`services/session_collector` was logging `Collection complete: 0 users, 0 files copied` + `WARNING: Group 'data-ops' not found, using default group` every 10 minutes in the Docker layout (where `/home/*/user/sessions/` doesn't exist). New env var `AGNES_SKIP_LEGACY_COLLECTOR=1` set by default in `docker-compose.yml` short-circuits the collector pass.
The bare-VM deployment path (where /home/* IS populated by Claude Code) leaves the env var unset and continues to scan normally — including the data-ops warning, which is load-bearing for catching missing-group mis-deploys.
### O2 — FIFO check in `_check_session_pipeline`
The existing check compares `MAX(processed_at)` to newest jsonl mtime — catches "detector hasn't run lately" but blind to "old file was skipped while newer ones were processed". New code finds the oldest FS jsonl that's NOT in `session_extraction_state.session_file` and flags if its mtime is older than `SESSION_PIPELINE_STUCK_FILE_GRACE_SECONDS` (default 4× the existing grace = 2h).
Severity intentionally starts at `info` so we can collect prod data on false-positive rate before tightening to `warning`. The aggregator already treats `info` as non-promoting (see the severity vocabulary docstring at the top of `app/api/health.py`), so the headline `status` stays at `healthy` even when this fires — the operator sees the entry in the per-check breakdown but no spurious `degraded` overall.
## Test plan
- [x] `pytest tests/test_session_collector.py` — 17 tests pass (existing 9 + new 8 covering env-set/unset, truthy variants, falsy non-skip).
- [x] `pytest tests/test_health_session_pipeline.py` — 8 tests pass (existing 4 + new 4 FIFO tests covering stuck-file, under-threshold, all-processed, env-override).
<!-- devin-review-badge-begin -->
---
<a href="https://app.devin.ai/review/keboola/agnes-the-ai-analyst/pull/229" target="_blank">
<picture>
<source media="(prefers-color-scheme: dark)" srcset="https://static.devin.ai/assets/gh-open-in-devin-review-dark.svg?v=1">
<img src="https://static.devin.ai/assets/gh-open-in-devin-review-light.svg?v=1" alt="Open in Devin Review">
</picture>
</a>
<!-- devin-review-badge-end -->
## Summary
`agnes self-upgrade` without `--force` previously short-circuited on the local 24h `update_check.json` cache. After a server-side version bump within that window, the explicit command exited silently as a no-op — empirically observed today when prod 0.47.1 → 0.47.2 didn't propagate.
Fix: always invalidate the cache in `_resolve_info`. The cache still gates the implicit warning loop in the root callback (correctly — that runs on every `agnes <anything>` and can't hammer `/cli/latest`).
## Test plan
- [x] New `test_self_upgrade_bypasses_24h_cache_without_force` — stale cache claims current; mocked server reports newer; assert UpdateInfo carries the newer version, not the cached one.
- [x] Existing self-upgrade tests pass (including `--force` semantics — force is now downstream-only, behavior preserved).
## Summary
Smoke-testing the just-shipped 0.47.1 against production exposed two regressions:
1. `agnes query --remote "SELECT FROM unit_economics WHERE bad_col=1"` returned `Table "unit_economics" must be qualified` (the OLD error) instead of `Unrecognized name: bad_col` (the #218 fix's intended behavior).
2. `agnes query "DESCRIBE unit_economics"` showed only DuckDB's misleading `Did you mean order_economics?` with no Agnes hint paragraph (the #219 fix is missing).
Root cause: PR #217's squash merge (`506a378c`) carried stale snapshots of `app/api/query.py` and `cli/commands/query.py` from before #218 and #219 merged. The rebase-and-merge auto-merged those files cleanly (no conflict markers) but the result silently reverted both fixes.
Restore the two changes verbatim. Tests for both fixes already on main and continue to pass against the restored code.
## Test plan
- [x] `pytest tests/test_api_query_guardrail.py tests/test_cli_query.py` — clean
- [x] Manual repro against prod after deploy: both flows now surface the intended diagnostic.
<!-- devin-review-badge-begin -->
---
<a href="https://app.devin.ai/review/keboola/agnes-the-ai-analyst/pull/225" target="_blank">
<picture>
<source media="(prefers-color-scheme: dark)" srcset="https://static.devin.ai/assets/gh-open-in-devin-review-dark.svg?v=1">
<img src="https://static.devin.ai/assets/gh-open-in-devin-review-light.svg?v=1" alt="Open in Devin Review">
</picture>
</a>
<!-- devin-review-badge-end -->
## Summary
Brings the Keboola connector to feature parity with the legacy internal data-analyst's per-table sync strategies. Closes the four documented gaps from the spec branch (`zs/keboola-connector-specs`):
- **Typed parquet** in the legacy SDK extraction path — column types from Keboola Storage metadata (provider cascade `user > ai-metadata-enrichment > keboola.snowflake-transformation`) survive the CSV → parquet roundtrip; invalid date strings (`'0000-00-00'`) and invalid numeric strings (`'Non-Manager'`) become NULL while keeping the column's typed schema. Pre-fix everything was VARCHAR.
- **Incremental sync** via Storage API `changedSince` — opt-in per table; pulls only delta rows, merges into the existing parquet by `primary_key` (drop_duplicates with keep='last'). Cuts daily extraction from O(full table) to O(delta).
- **Partitioned sync** — flat per-partition layout `data/<table>/<key>.parquet` (e.g. `2026_05.parquet`), per-affected-partition merge for daily updates, chunked initial load with 1-day overlap and 2-empty-chunk stop heuristic.
- **`where_filters`** — server-side row filter with date placeholders (`{{today}}`, `{{last_3_months}}`, `{{start_of_3_months_ago}}`, etc.) resolved at sync time. Force the SDK path; reject `incremental + where_filters` combination at API layer (changedSince already filters temporally).
## Architecture
- **Schema migration v25 → v26**: 7 new columns on `table_registry`. Existing `sync_strategy` column reused (pre-v26 it was inert catalog metadata; post-v26 the extractor dispatches off it).
- **Per-table dispatcher** in `extractor.run()` routes to one of `_extract_via_extension` (full_refresh + extension), `_extract_via_legacy` (full_refresh + filters or extension fallback), `extract_incremental`, or `extract_partitioned`.
- **API conflict policy**: `incremental + where_filters` → 422; `partitioned + query_mode='remote'` → 422; `partitioned ⇒ partition_by required`.
- **Admin UI**: third "Direct extract (Storage API)" radio in the Keboola Register / Edit modals, alongside existing "Whole table (extension)" and "Custom SQL". When selected, exposes a v26 sync-strategy panel with conditional fields per strategy.
## Test plan
- [x] **Unit + module** — 134 v26 tests covering migration, repo, parquet_io, where_filters, incremental (compute_changed_since + merge_parquet + extract_incremental E2E), partitioned (key derivation + merge_partition + chunked windows + extract_partitioned E2E), extractor dispatcher, admin API validators, PUT field clearing, registry-shape → dispatcher bridge
- [x] **HTML form structure** — all v26 inputs + visibility classes + JS payload fields verified in rendered template
- [x] **Real Keboola roundtrip** — registered a small test table as `sync_strategy='incremental'` against a test Storage project, triggered two syncs:
- Sync 1: `changedSince=None` → full pull → 9 rows typed parquet
- Sync 2: `changedSince=last_sync - 1d window` → 9 delta rows merged with 9 existing → 9 after dedup on primary_key (PK merge confirmed)
- [x] **Browser UX** — agent-browser session against a local uvicorn: login → admin/tables → register modal → switch radios → verify field visibility per strategy → submit → edit existing row → switch to Direct/Incremental → save → confirm DB persistence
- [x] **Regression** — no regressions in the broader 3252-test suite (3 pre-v26 tests updated for the deprecation-marker removal + schema-version bump; 2 pre-existing environment-sensitive test failures unrelated to this change)
## Bugs caught + fixed during E2E
The browser + real-Keboola roundtrip exposed four bugs the unit tests missed:
1. **JS visibility race** — two competing `forEach` loops set `display=''` then `display='none'` on form elements sharing `kb-strategy-incremental kb-strategy-partitioned` classes (window_days + max_history_days are reused across strategies). Fix: single-pass selector with class-based visibility resolver.
2. **PUT cannot clear field** — pre-v26 `updates = {k: v ... if v is not None}` collapsed "omitted from body" and "sent as null" into the same case, so admin couldn't switch a partitioned row back to full_refresh and have stale `partition_by` clear. Fix: `model_dump(exclude_unset=True)`.
3. **Subprocess DB lock conflict** — `_read_last_sync` reopened `system.duckdb` while the parent server held the write lock (subprocess contract at `app/api/sync.py:_run_sync` line 260). Fix: parent injects `__last_sync__` into table_config before subprocess spawn.
4. **Wrong KBC table_id** — `extract_incremental` / `extract_partitioned` built the Storage API table_id from the registry row's slugified `id` (`circle_inc`) instead of `bucket.source_table` (`in.c-finance.circle`), producing 404s. Fix: prefer `bucket+source_table`; fall back to `id` only when bucket empty.
## Operator notes
- Existing tables stay on `full_refresh` after migration; admins opt individual tables in via `agnes admin register-table --sync-strategy ...`, the Keboola Edit modal, or `POST/PUT /api/admin/registry`.
- `merge_parquet` and `merge_partition` use `pd.concat + drop_duplicates`, loading both existing and delta into pandas RAM. For tables in the multi-million-row range this may OOM — switch to `partitioned` strategy for those (per-partition merge keeps memory bounded). Documented in `### Internal` of the changelog entry.
- Date placeholders are resolved at **sync time**, not register time — a typo'd `{{lasst_week}}` is accepted at register and surfaces only when the next sync runs. By design (rolling windows need late-binding).
## Spec source
The four corresponding plans on the `zs/keboola-connector-specs` branch under `docs/superpowers/plans/2026-05-07-0[1-4]-*.md` capture the design rationale and link back to internal repo references for each subsystem.
<!-- devin-review-badge-begin -->
---
<a href="https://app.devin.ai/review/keboola/agnes-the-ai-analyst/pull/217" target="_blank">
<picture>
<source media="(prefers-color-scheme: dark)" srcset="https://static.devin.ai/assets/gh-open-in-devin-review-dark.svg?v=1">
<img src="https://static.devin.ai/assets/gh-open-in-devin-review-light.svg?v=1" alt="Open in Devin Review">
</picture>
</a>
<!-- devin-review-badge-end -->
## Summary
- Catalog enrichment for `query_mode='remote'` rows: `rows`, `size_bytes`, `partition_by`, `clustered_by` per table (BQ + Keboola providers).
- `/api/v2/schema/{id}` cache miss: 2 BQ jobs → 1 (-50%) via shared `fetch_bq_columns_full`.
- All four catalog/schema/sample/metadata caches flush on registry change; single-row re-warm scheduled.
- Automatic cache warmup at server startup (bounded concurrency, opt-out via `AGNES_SKIP_CACHE_WARMUP=1`).
- SSE-driven freshness toolbar on `/admin/tables` with progress bar, log, and per-row badge.
- New admin doc `docs/admin/query-modes.md` — single source of truth on `local` / `remote` / `materialized` choice.
Closes#155.
Closes#156.
## Test plan
- [x] 65+ targeted tests pass across 11 new test modules + 3 modified ones.
- [x] No DB migration; no wire-break; `MIN_COMPAT_CLI_VERSION` unchanged.
- [ ] Reviewer: register a remote BQ table via `/admin/tables`, observe the toolbar populates within ~2 s and the per-row badge transitions warming → fresh.
- [ ] Reviewer: trigger `Re-warm all`, verify SSE log scrolls and `cacheWarmupBar` progresses.
- [ ] Reviewer: edit a registered row's bucket, verify `agnes schema <id>` returns updated columns immediately (no 1-hour staleness).
- [ ] Reviewer: confirm `agnes admin register-table --query-mode remote` prints the new IAM-smoke-check hint.
## Notable design decisions
- BigQuery `INFORMATION_SCHEMA.TABLE_STORAGE` is the only valid scope for size+rows (verified live 2026-05-07; dataset-scoped doesn't exist). Region resolved from `instance.yaml.data_source.bigquery.location` → `bq.client().get_dataset(...)` → fall back to legacy `__TABLES__`.
- VIEW handling: TABLE_STORAGE returns no rows for views, fall through to `__TABLES__` (also empty) → `TableMetadata(rows=None, size_bytes=None, partition_by=..., clustered_by=...)`. Null size signals analyst Claude to apply existing CLAUDE.md guidance.
- `size_bytes` is `active_logical_bytes + long_term_logical_bytes` — full BQ scan reads both; reporting only active undercounts aged partitioned tables.
- Source-agnostic provider seam: per-source `connectors/<source>/metadata.py:fetch(MetadataRequest)`; dispatcher in `app/api/v2_catalog.py:_metadata_provider_for` lazily imports per source_type so a Keboola-only deployment doesn't pay the BQ-extension import cost.
- Warmup non-blocking: FastAPI `lifespan` schedules `asyncio.create_task(_warm_catalog_caches_bg)` before `yield`. Per-row failures isolated.
## Out of scope
- Profile / column histograms / dimension cardinality for remote tables (separate issue).
- Onboarding nudge ("you have 0 remote tables, consider registering some BQ ones") — separate UX call.
- Provider plug-in registration via entry-points (the dispatch table is a hardcoded if-tree today; one line per future source).
## Release
Bumps `pyproject.toml` 0.46.1 → 0.47.0 (main shipped 0.46.0 + 0.46.1 during this PR — see commit `d98976ec`). New CHANGELOG section under `## [0.47.0] — 2026-05-07`.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
<!-- devin-review-badge-begin -->
---
<a href="https://app.devin.ai/review/keboola/agnes-the-ai-analyst/pull/223" target="_blank">
<picture>
<source media="(prefers-color-scheme: dark)" srcset="https://static.devin.ai/assets/gh-open-in-devin-review-dark.svg?v=1">
<img src="https://static.devin.ai/assets/gh-open-in-devin-review-light.svg?v=1" alt="Open in Devin Review">
</picture>
</a>
<!-- devin-review-badge-end -->
## Summary
Two bugs in `agnes describe` surfaced from a real analyst session following the CLAUDE.md agent-rails discovery workflow. Together they break `agnes describe` end-to-end for any analyst (or analyst-AI) who follows the documented form.
### A) CLI parsing
`agnes describe TABLE -n 5` failed with `Missing argument 'TABLE_ID'`. Root cause: the command was registered as a `Typer.Typer` subcommand group via `app.add_typer(describe_app, name="describe")` + `@describe_app.callback(invoke_without_command=True)`, and that pattern mis-parses positional + short-int option in some orderings. Same pattern in `cli/commands/schema.py` works only because schema has no INTEGER short option. Fix: switch to flat `@app.command("describe")`.
### B) Server NaN
`/api/v2/sample/<id>` (called by `agnes describe`) returned HTTP 500 with `ValueError: Out of range float values are not JSON compliant: nan` whenever a row contained NaN. Fix: sanitize NaN/±inf to None before JSON serialization.
## Test plan
- [x] `pytest tests/test_cli_describe*.py` — added regression tests pinning `-n` parsing on either side of the positional.
- [x] `pytest tests/test_api_v2_sample*.py` — added regression test for NaN row → JSON `null` (not 500).
<!-- devin-review-badge-begin -->
---
<a href="https://app.devin.ai/review/keboola/agnes-the-ai-analyst/pull/224" target="_blank">
<picture>
<source media="(prefers-color-scheme: dark)" srcset="https://static.devin.ai/assets/gh-open-in-devin-review-dark.svg?v=1">
<img src="https://static.devin.ai/assets/gh-open-in-devin-review-light.svg?v=1" alt="Open in Devin Review">
</picture>
</a>
<!-- devin-review-badge-end -->
## Summary
`claude -p` (headless mode) gives SessionEnd hook subprocesses ~1 second before SIGTERM, regardless of work in progress. `agnes push` for a typical workspace takes 5-30s. The current synchronous SessionEnd hook (`agnes push --quiet 2>/dev/null || true`) was therefore being killed mid-first-upload — `|| true` masks the SIGTERM as exit 0, so this regression was invisible until I traced it via a wrapper script and Claude's `~/.claude/debug/<sid>.txt` log.
Fix: wrap SessionEnd push in `bash -c "( nohup agnes push --quiet </dev/null >/dev/null 2>&1 & ) ; true"`. The subshell exits immediately, orphaning the upload child to init so it survives the hook subprocess kill. Same `bash -c` pattern as the existing `refresh-marketplace` SessionStart entry (for Windows compatibility).
End-to-end verified against production: claude exited in 5s, detached child completed the upload, file `491e3a23-...jsonl` landed on the server within 30s with mtime 14:30 UTC.
## Test plan
- [x] `pytest tests/test_lib_hooks.py` — added `test_session_end_push_is_detached` regression test asserting `nohup`, `&`, `</dev/null` are all present.
- [x] `pytest tests/test_setup_hooks_template.py` — assertions loosened from `==` to `in` where necessary.
- [x] Verified end-to-end against production with the detached wrapper before opening this PR (manual probe).
<!-- devin-review-badge-begin -->
---
<a href="https://app.devin.ai/review/keboola/agnes-the-ai-analyst/pull/222" target="_blank">
<picture>
<source media="(prefers-color-scheme: dark)" srcset="https://static.devin.ai/assets/gh-open-in-devin-review-dark.svg?v=1">
<img src="https://static.devin.ai/assets/gh-open-in-devin-review-light.svg?v=1" alt="Open in Devin Review">
</picture>
</a>
<!-- devin-review-badge-end -->
## Summary
Verified against production: `claude -p` headless mode doesn't fire SessionEnd hooks (proven via `--output-format stream-json --include-hook-events`: zero `SessionEnd` events), so any session JSONLs from `-p` invocations stay orphaned locally and never reach the server. Fix: add `agnes push --quiet` as a third SessionStart entry — symmetric self-heal alongside the existing `agnes pull` entry. Existing workspaces pick this up on their next `agnes init` via the marker-based migration already in `cli/lib/hooks.py`.
Separately: a colleague's fresh install showed `agnes diagnose` warning "uploads are not being processed", which led them to suspect their `agnes push` was broken. The warning is actually about the LLM-based `verification-detector` backlog (uploads themselves were arriving fine — confirmed by 23+3 JSONLs landed on the server while the warning was firing). Reword the warning to "verification-detector backlog" + add `last_processed` to the diagnose dict so operators don't have to grep logs to confirm.
## Test plan
- [x] `pytest tests/test_lib_hooks.py` — updated count + added `agnes push in SessionStart` assertion.
- [x] `pytest tests/test_setup_hooks_template.py` — updated.
- [x] `pytest tests/test_clean_install_integration.py` — updated.
- [x] `pytest tests/test_health_session_pipeline.py` — updated warning text + asserted `last_processed` field.
<!-- devin-review-badge-begin -->
---
<a href="https://app.devin.ai/review/keboola/agnes-the-ai-analyst/pull/220" target="_blank">
<picture>
<source media="(prefers-color-scheme: dark)" srcset="https://static.devin.ai/assets/gh-open-in-devin-review-dark.svg?v=1">
<img src="https://static.devin.ai/assets/gh-open-in-devin-review-light.svg?v=1" alt="Open in Devin Review">
</picture>
</a>
<!-- devin-review-badge-end -->
## Summary
`agnes query "DESCRIBE unit_economics"` (where `unit_economics` is `query_mode='remote'`) previously returned DuckDB's nearest-name suggestion (`Did you mean "order_economics"`?), sending users down the wrong path. Now appends a friendly hint about remote tables.
Reproduced from a real analyst session — colleague spent ~30s diagnosing what was actually "this is a remote table, not materialized locally".
## Test plan
- [x] New test: `_query_local("DESCRIBE unit_economics", ...)` against an empty local DuckDB triggers the new hint, original DuckDB error still echoed.
- [x] Negative test: a syntax-error query does NOT trigger the hint (regex only matches "Table with name X does not exist").
- [x] `pytest tests/test_cli_query*.py` clean.
<!-- devin-review-badge-begin -->
---
<a href="https://app.devin.ai/review/keboola/agnes-the-ai-analyst/pull/219" target="_blank">
<picture>
<source media="(prefers-color-scheme: dark)" srcset="https://static.devin.ai/assets/gh-open-in-devin-review-dark.svg?v=1">
<img src="https://static.devin.ai/assets/gh-open-in-devin-review-light.svg?v=1" alt="Open in Devin Review">
</picture>
</a>
<!-- devin-review-badge-end -->
## Summary
When `agnes query --remote` references a column that doesn't exist on the FROM table, users were seeing `Table "<id>" must be qualified with a dataset` instead of the actually-useful `Unrecognized name: <column>` from BigQuery. Surface the first-attempt diagnostic now; keep the second-attempt context as `underlying_original`.
Reproduced against production:
```
$ agnes query --remote "SELECT COUNT(*) FROM unit_economics WHERE authorize_date = DATE '2025-05-06'"
Error: remote_estimate_failed (HTTP 400)
message: Could not estimate scan size for this query.
underlying: 400 ... Table "unit_economics" must be qualified with a dataset.
```
(`unit_economics` has `authorize_timestamp`, not `authorize_date`.)
## Test plan
- [x] New `test_remote_estimate_failed_surfaces_first_error_when_attempts_differ` asserts the first-attempt message wins, second-attempt is preserved as `underlying_original`, hint points to `agnes schema`.
- [x] Existing `test_guardrail_returns_400_remote_estimate_failed_on_double_parse_error` still passes (both attempts mocked to identical error).
- [x] `pytest tests/test_api_query_guardrail.py` clean.
<!-- devin-review-badge-begin -->
---
<a href="https://app.devin.ai/review/keboola/agnes-the-ai-analyst/pull/218" target="_blank">
<picture>
<source media="(prefers-color-scheme: dark)" srcset="https://static.devin.ai/assets/gh-open-in-devin-review-dark.svg?v=1">
<img src="https://static.devin.ai/assets/gh-open-in-devin-review-light.svg?v=1" alt="Open in Devin Review">
</picture>
</a>
<!-- devin-review-badge-end -->
Cuts the [Unreleased] section into [0.46.0] in CHANGELOG.md and
bumps pyproject.toml. The user-visible content was already on main
via PR #190 (commit 28430ced); this is the release-cut commit that
should have been the last commit on that PR — splitting it out so
the operator-facing release artifact (tag + GitHub Release) lines up
with what's already deployed at :stable.
* fix: cutover regressions + parallel Keboola legacy fallback
Bundled fixes from a fresh-deploy run on a Keboola Storage backend with
the block-shared-snowflake-access feature flag — DuckDB Keboola
extension's per-table scan can't access bucket schemas, so the legacy
kbcstorage Storage-API client is the only working path.
CUTOVER REGRESSIONS
- agnes pull hash mismatch on every Keboola local-mode table —
src/orchestrator.py:_update_sync_state stored md5(mtime+size)[:12]
while the CLI compares against full 32-char content MD5. Now stores
the same content MD5 the materialized SQL path already used.
- Trailing-slash sanitization in connectors/keboola/access.py and
extractor.py — DuckDB Keboola extension's ATTACH fails when the URL
ends in / (canonical form).
- src/profiler.py:TableInfo.description becomes optional — two call
sites instantiated without it, crashing the profiler pass.
- scripts/ops/agnes-auto-upgrade.sh: chown on UID change — older images
ran as root, current runs as agnes (uid 999). Reads target uid:gid
from /etc/passwd inside the new image and chowns ${STATE_DIR},
/data/extracts, /data/analytics when the digest moves.
- POST /api/sync/trigger is now singleton per process — two
near-simultaneous trigger calls each forked an extractor subprocess,
fought for extract.duckdb's file lock, starved uvicorn, flipped the
container to unhealthy. Trigger now returns 409
(sync_already_in_progress) when held; _run_sync acquires non-blocking.
PARALLEL LEGACY FALLBACK
- Process pool fan-out for the _extract_via_legacy queue (default 8
workers, override via AGNES_KEBOOLA_PARALLELISM). Process pool, not
thread pool, because connectors/keboola/client.py:export_table does
os.chdir(temp_dir) — process-global, so threads raced and slice files
landed in the wrong directory ("[Errno 2] No such file or directory:
'<job_id>.csv_X_Y_Z.csv'").
- Extractor subprocess timeout 1800s -> 3600s (configurable via
AGNES_EXTRACTOR_TIMEOUT_SEC). 28+ tables × multi-minute Keboola export
jobs need the headroom on telemetry-class projects.
- Process group cleanup on timeout — Popen(start_new_session=True) puts
the extractor in its own group. On timeout the parent SIGTERMs the
group (10s grace) then SIGKILLs stragglers. Without this, the pool
workers were reparented to PID 1 and continued holding open Keboola
Storage export jobs. Inline extractor script also installs a SIGTERM
-> sys.exit(143) handler so the with ProcessPoolExecutor(...) block
__exit__ runs cleanly.
Tests: existing tests that patched subprocess.run updated to patch
subprocess.Popen with a _FakePopen stand-in (same exit-code-injection
contract). Two tests that exercised the parallel path forced
AGNES_KEBOOLA_PARALLELISM=1 to keep mocks alive (mocks don't ride into
ProcessPoolExecutor subprocesses).
Squashed onto current main (was 7 commits + multi-commit CHANGELOG +
agnes-auto-upgrade.sh conflicts; squash avoids per-commit conflict
resolution against main's flat-mount STATE_DIR refactor and 0.38.0
release cut).
* feat(keboola): Storage API direct extract path; drop extension data path
The DuckDB Keboola extension's COPY routes through Keboola QueryService,
which is unreliable on linked-bucket projects (extension v0.1.6 fixes
that case but isn't yet in the community CDN, and pre-fix any project
with the block-shared-snowflake-access feature flag couldn't see bucket
schemas at all). Move the extract path off the extension entirely and
talk to the Storage API directly via signed-URL download — works on any
project, regardless of extension state.
connectors/keboola/storage_api.py (NEW)
Lightweight client built on requests.Session. Three endpoints:
- POST /v2/storage/tables/{id}/export-async (kicks off job)
- GET /v2/storage/jobs/{id} (poll until done)
- GET /v2/storage/files/{id}?federationToken=1 (signed URL detail)
- GET <signed_url> (download bytes)
Supports sliced exports (manifest + per-slice signed URLs) and gzipped
payloads. ExportFilter dataclass mirrors the Keboola filter spec
(whereFilters / columns / changedSince / limit) and handles JSON
round-trip with the registry's source_query column. Token redaction
in error messages. Bounded exponential backoff on job polling.
No cloud-SDK dependency on the data path; thread-safe.
connectors/keboola/extractor.py
- materialize_query() rewritten: takes bucket/source_table/source_query
(JSON filter spec), exports via KeboolaStorageClient, converts CSV
to parquet via DuckDB, atomic os.replace. Same return shape so
sync.py downstream code stays uniform with the BQ branch.
- _extract_via_legacy() also moved to Storage API direct (kept the
name for caller compatibility with _legacy_worker / the parallel
batch extractor). Per-call temp directories — no os.chdir, threads
don't race.
app/api/sync.py
_run_materialized_pass for source_type='keboola' rows now constructs a
KeboolaStorageClient (replaces KeboolaAccess) and passes
bucket/source_table/source_query to materialize_query. Reuses one
client across rows for HTTP keep-alive. Sources keboola URL from env
too (KEBOOLA_STACK_URL) when instance.yaml doesn't have stack_url
configured.
cli/commands/admin.py
discover-and-register defaults Keboola rows to query_mode='materialized'
(NULL source_query = full table), matching the v26 migration's
unification of the local/materialized split for Keboola. BigQuery and
Jira keep their per-source defaults.
src/db.py
Schema bump 25 → 26. Migration: UPDATE table_registry SET
query_mode='materialized' WHERE source_type='keboola' AND
query_mode='local'. NULL source_query on those rows means "full table
export" — same effective behavior the local mode provided, but now
via Storage API instead of the extension.
pyproject.toml
kbcstorage dep stays (admin-side bucket/table list still uses the
SDK in app/api/admin.py / connectors/keboola/client.py); only the
data path is migrated off the SDK. Comment updated to reflect the
new boundary.
tests
- test_keboola_storage_api.py (NEW, 19 tests): ExportFilter parsing,
HTTP client (token redaction, retry logic, polling), download_file
(single, gzipped, sliced), end-to-end export_table_to_csv.
- test_keboola_materialize.py rewritten: mocks KeboolaStorageClient
instead of FakeAccess; same atomic-write + zero-rows + unsafe-id
contracts.
- test_sync_trigger_keboola_materialized.py: registry rows now carry
bucket+source_table+JSON-shape source_query.
114+ Keboola-impacted tests green locally.
* test: schema version assertion bumped to 26 alongside the keboola query_mode migration
* fix(keboola): cutover hot-patches surfaced on agnes-dev
Five small fixes that were applied as in-container hot-patches during
agnes-dev cutover and need to be on the source-of-truth image so a fresh
upgrade does not undo them.
- app/api/sync.py: auto-discover gate considers the WHOLE registry (any
source, any mode), not just rows where source matches and query_mode
is local. After the v25→v26 keboola materialized migration an
instance can have 30 materialized rows and zero local rows; the
previous gate kept re-firing _discover_and_register_tables every
scheduler tick, creating duplicate auto-discovered rows with the
wrong bucket prefix every time.
- app/api/admin.py: _discover_and_register_tables reassembles the
bucket as <stage>.<bucket-id> (e.g. in.c-finance) instead of
dropping the stage prefix; default query_mode for keboola is now
materialized (the v26 contract); validator allows NULL source_query
for keboola materialized rows (full-table export via Storage API
export-async, no SQL needed).
- cli/commands/admin.py: register-table mirrors the server validator
(NULL source_query allowed for source_type=keboola); --bucket help
text generalized to cover both BQ dataset and Keboola bucket id.
- connectors/keboola/extractor.py: max_line_size=64 MiB on
read_csv_auto so embedded JSON / SQL cells (kbc_component_configuration
in particular) do not trip the default 2 MiB ceiling.
- connectors/keboola/storage_api.py: GCP backend support — when the
Storage API returns a manifest whose slice URLs are gs://
references with a gcsCredentials block, rewrite to the JSON REST
download endpoint and authenticate with the issued OAuth bearer
token; redact tokens in any surfaced error string.
* test: align with new keboola materialized + auto-discover-gate contracts
- test_admin_keboola_materialized: rename
test_register_keboola_materialized_rejects_missing_source_query →
test_register_keboola_materialized_accepts_missing_source_query.
v25→v26 introduced 'keboola materialized with NULL source_query
means full-table export via Storage API export-async' as the
default registration shape; the rejection case is no longer the
contract.
- test_sync_filter: add list_all() to _StubRegistry. The auto-discover
gate in _run_sync now keys off the WHOLE registry (not just local
rows) so materialized-only Keboola instances do not re-trigger
discovery on every tick.
* feat(keboola): native parquet export — skip CSV roundtrip
Storage API export-async accepts fileType={csv,parquet}. Switching the
materialized sync to parquet eliminates the CSV → DuckDB COPY → parquet
roundtrip that pinned a single uvicorn worker over 4 GiB on multi-GB
tables (read_csv with all_varchar + max_line_size=64MB has to
materialize the whole CSV in memory before COPY can stream out a
parquet). Snowflake UNLOAD on Keboola's side already produces typed,
self-contained parquet files; the extractor downloads them and renames
into place.
Two cases:
- **Single-file** export (small table): file_info.url points at one
signed URL; download_file streams chunks straight to .parquet.tmp
and we're done. No DuckDB.
- **Sliced** export (Snowflake UNLOAD respects MAX_FILE_SIZE — 16 MiB
default — so anything larger arrives as N parquet slices): each
slice is a complete parquet file with its own footer; naive concat
would corrupt them. download_file_slices keeps the slices as
separate files in a tempdir, then DuckDB COPY (SELECT * FROM
read_parquet([slice0, slice1, ...])) merges them into one
consolidated parquet. DuckDB streams row groups during this — peak
memory bounded to one row group (~1 MiB) regardless of source size.
The legacy CSV path stays as the explicit opt-in via source_query=
'{"file_type":"csv"}' for projects whose backend can't UNLOAD
parquet (none known today; cheap escape hatch). Backward-compat alias
KeboolaStorageClient.export_table_to_csv kept.
Also fixes a latent bug in download_file's gzip detection: previous
heuristic flagged any unencrypted file as gzipped, which would have
corrupted parquet downloads at gunzip time. Name-suffix-only now.
* fix: tempdir leak cleanup, every 0m schedule, /sync/trigger body shapes
Three small self-contained fixes uncovered during agnes-dev cutover.
- connectors/keboola/extractor.py: tempfile.TemporaryDirectory now uses
ignore_cleanup_errors=True so a worker death mid-write doesn't leave
multi-GiB stale slice trees on the boot disk. (12 GiB seen after a
disk-full crash where TemporaryDirectory's own cleanup also raised
and got swallowed.)
- src/scheduler.py: is_valid_schedule accepts 'every 0m' (interval=0
= always due). Force-resync of an errored row no longer requires
waiting out the default 'every 1h' interval — admin can flip the
schedule, trigger, then flip back.
- app/api/sync.py: POST /api/sync/trigger accepts both ['table_id']
(legacy bare-array body) and {'tables': ['table_id']} (matches the
response payload shape, more discoverable for clients building
requests by hand). Malformed bodies return 422 with a structured
detail; null/missing means 'sync everything' as before.
Tests cover: tempdir cleanup on raise (sliced parquet path),
is_valid_schedule + is_table_due 'every 0m' acceptance, and trigger
body parametrized matrix (8 valid shapes + 6 rejection cases).
* fix: targeted-trigger filter in materialized pass + auto-upgrade defer
Two operational gaps observed during agnes-dev cutover, in the same
sync-routing area.
- _run_materialized_pass now takes a 'tables' arg and skips rows not in
the target set with reason='not_in_target'. POST /api/sync/trigger
with a body of tables previously only scoped the legacy extractor
subprocess — the materialized pass kept iterating every due
materialized row, so an admin asking to re-sync kbc_job re-ran
every other due materialized row alongside it. Match on registry id
OR name (admins commonly pass either form). tables=None preserves
the no-filter behavior.
- New GET /api/sync/status (public, no auth) returns {locked: bool}
off _sync_lock.locked(). agnes-auto-upgrade.sh probes this before
docker compose up -d and exits 0 with a 'deferred recreate' log
line if a sync is in flight — the next 5-min cron tick retries.
Pre-fix, an auto-upgrade triggered mid-sync would recreate the
uvicorn worker and kill the in-flight extractor / Snowflake-UNLOAD
download (observed when kbc_job's first 7-day retry got SIGKILLed).
Connection failures in the probe fall through to the upgrade —
being stuck on a wedged image is worse than interrupting a
hypothetical sync.
* fix: auto-discover protects admin overrides + surfaces drift
Two real-world incidents on agnes-dev drove this:
1. kbc_job was registered manually with the correct
(in.c-kbc_telemetry, kbc_job) coordinates. A naive auto-discover
re-run would have inserted a SECOND kbc_job row at the slugified
id 'in_c-keboola-storage_kbc_job' (where Keboola's discovery
places it) — and that row's Storage API export-async 404s.
2. An earlier auto-discover bug stripped the stage prefix from
bucket ids ('c-finance' instead of 'in.c-finance'), inserting
137 rows whose syncs all failed.
Fix:
- _discover_and_register_tables now builds a plan first
(_build_keboola_discovery_plan) classifying each discovered table
into one of new / existing_match / existing_drift / invalid, then
executes only the 'new' bucket. Drift rows are reported with both
sides of the disagreement plus drift_kind:
- same_id_diff_coords: registry has the same id but different
bucket / source_table (admin migrated coords inline).
- name_collision: discovery's slugified id differs from any
registry id, but the discovered .name matches an existing row's
.name (case-insensitive). Catches the kbc_job case.
- Bucket detection now prefers the API's authoritative bucket_id
field (separate field on the Keboola tables.list response,
normalised by KeboolaClient.discover_all_tables). Falls back to
id-string parsing only when bucket_id is missing (older fallback
path inside discover_all_tables).
- Endpoint POST /api/admin/discover-and-register?dry_run=true
returns the plan without writing — would_register, drift,
invalid lists. Lets an operator audit before merging discovery
with a registry that has admin overrides.
Removed 'every 0m' from test_register_request_rejects_malformed_sync_schedule
— the runtime started accepting it in the previous commit (force-resync
override) and the validator follows suit.
* feat(keboola): AGNES_TEMP_DIR routes tempfiles off overlayfs /tmp
The container's /tmp lives on the boot disk's overlayfs (29 GiB on
agnes-dev, shared with /var). Snowflake UNLOAD of a wide table writes
slices into per-call /tmp tempdirs that fill multi-GiB / many-slice
exports long before the dedicated data disk fills. agnes-dev hit
100% boot-disk while the 20 GiB data disk had 15 GiB free.
connectors.keboola.storage_api.get_temp_root() reads AGNES_TEMP_DIR;
mkdirs the target on first use; unset / empty / unwritable falls
back to None (system tempdir, OSS-pre-fix behaviour). Both
materialize_query (parquet path) and _extract_via_legacy (CSV
fallback) and the sliced-CSV concat path in storage_api use the
helper now.
docker-compose.yml defaults AGNES_TEMP_DIR=/data/tmp on app, scheduler,
and extract services. The data volume is the dedicated disk in
production layouts and a plain docker volume in single-disk
dev/laptop setups — same blast radius as the previous /tmp default
on the latter, no regression.
Operator-and-analyst quality bundle: a security fix for the optional
Telegram bot, two CLI gaps closed, and three rounds of UX polish on
`agnes diagnose` and `agnes pull` so non-TTY consumers (CI runners,
Claude Code SessionStart hooks, sub-agent watchdogs) get readable,
actionable signal.
- Pairing-code RNG: random.choices -> secrets.choice (CSPRNG).
- Telegram script runner: refuse out-of-shape usernames before sudo -u.
CLAUDE.md.bak.<ISO-timestamp> before regenerating.
- agnes admin unregister-table <id> -> DELETE /api/admin/registry/{id}
- agnes admin update-table <id> --field=value ... -> PUT /api/admin/registry/{id}
response but never promotes the headline. BQ billing-equals-data check
downgraded warning -> info.
default (5 s / 1 MiB vs 30 s / 10%) so sub-agent watchdogs don't kill
the pull as a hung process. New env knobs:
AGNES_PULL_PROGRESS_INTERVAL_{SECONDS,BYTES}.
--include-schema (or ?include=schema) to opt back in.
Tests: 120 passed across the touched modules, including new tests for
each fix. Pre-existing failures on main (DB migration v1->v9, binary
rename) are unrelated and not introduced here.
The startup script runs on every boot but the metadata_startup_script
field is in lifecycle.ignore_changes — so a TF apply that changed
image_tag does NOT reach a long-lived VM until someone explicitly
recreates it. Meanwhile, operators commonly hand-edit /opt/agnes/.env
to pin a specific image (custom branch builds, staged rollouts).
Pre-fix, every boot rewrote .env from the baked-in template and
clobbered the operator's choice — concretely, a stop+start triggered
by a machine_type change would reset AGNES_TAG to whatever was in
the template at first provision, regardless of the operator's
intervening edit.
Now the script reads the existing .env (when present) for AGNES_TAG
and AGNES_TEMP_DIR; when those keys are set, the existing values
win over the template-computed ones. Logged on stdout when AGNES_TAG
disagrees with $IMAGE_TAG so an operator audit-trails the boot.
Fresh provisions are unchanged (no .env yet → template values land).
To force a TF-driven reset on an existing VM: rm /opt/agnes/.env
and reboot. Cut as infra-v1.8.0 — additive, downstream consumers
opt in by bumping the module ref.
The 'Add to group' dropdown on /admin/users/{id} silently filtered out
every Google-Workspace-managed group (rightly — the API would 409 on
POST). On deployments where Admin and Everyone are both Workspace-mapped
via AGNES_GROUP_{ADMIN,EVERYONE}_EMAIL and no custom Agnes groups exist
yet (FoundryAI prod + dev today), the picker showed only the literal
'— Pick a group —' option with the 'Add' button disabled. Operator had
no indication that they needed to create a custom group first.
Three states surface a hint below the picker now:
- user is already in every group (literally nothing left)
- every remaining group is Google-Workspace-managed (link to
/admin/groups + admin.google.com explainer)
- no groups exist at all
The skip-google-managed logic stays — POST would still 409 on those
rows, this just stops the empty-state from being a silent dead end.
Hook JSON path uses /reload-plugins (no restart needed); manual-mode
echo path was still telling the operator to /exit + restart. Both now
say /reload-plugins.
Tests renamed to *_reload_hint_* to match the new wording.
Claude Code's `/reload-plugins` slash command picks up newly installed
plugins into the running session without forcing the user to /exit and
restart Claude Code. The hook JSON `systemMessage` and `additionalContext`
both now point at it.
Tests updated to pin the new hint shape.
Without `--scope project`, `claude plugin update <name>@agnes` operated
at user scope (the default) instead of updating the project-scoped
install — so version bumps in the served manifest never propagated to
the workspace, even though `claude plugin install` correctly used
`--scope project` for the missing-plugin path.
Mirrors the install line in the same function. Any change refresh-
marketplace makes to a plugin must now stay in project scope —
consistent with the SessionStart hook firing per-workspace.
Adds end-to-end flow for installing and keeping the per-user filtered
Claude Code marketplace in sync with the user's Agnes stack
(admin RBAC grants \ MyAIStack opt-outs U /store installs).
Setup (one-liner in install prompt step 5):
`agnes refresh-marketplace --bootstrap` clones the per-user marketplace
bare repo to ~/.agnes/marketplace, strips PAT from the cloned origin
URL, registers the local path with Claude Code, and installs every
plugin in the served manifest at --scope project. Replaces a 15-line
inline shell sequence that tripped Claude Code's agent-driven `rm -rf`
permission gate.
Auto-refresh (SessionStart hook installed by `agnes init`):
`agnes refresh-marketplace --quiet` runs every Claude Code session,
fetches+resets the clone (server rebuilds as orphan commits, so
pull --ff-only is impossible), and version-aware reconciles:
- missing in workspace -> claude plugin install <name>@agnes --scope project
- version differs -> claude plugin update <name>@agnes
- matches -> skip
Don't auto-uninstall plugins that disappeared from the manifest --
a transient empty manifest from the server would wipe the stack.
Hook output: when --quiet AND something actually changed, emits Claude
Code hook JSON on stdout -- `systemMessage` (transient toast) and
`hookSpecificOutput.additionalContext` (model-side system reminder),
both carrying the change summary plus a "/exit + restart Claude Code"
instruction (Claude only scans plugins at session start).
Windows hook compatibility: the refresh-marketplace hook command is
wrapped in `bash -c "..."` because Claude Code on Windows runs hook
commands directly without invoking a shell, so `2>/dev/null || true`
would otherwise be passed as literal argv tokens.
Cross-cutting:
- cli/lib/marketplace.py: shared CLONE_DIR + MARKETPLACE_NAME constants.
- cli/lib/hooks.py: SessionStart now has two independent entries
(pull + refresh-marketplace) so a failure in one doesn't suppress
the other; legacy `da sync` and prior single-pull layouts upgrade
cleanly on re-init.
- PAT injection on every git fetch via per-invocation credential
helper (token in \$AGNES_TOKEN env, never in argv or .git/config).
- Pre-snapshot of installed plugins captured BEFORE
`claude plugin marketplace update` so silent auto-applied version
bumps still fire notifications.
- scripts/dev/agnes-client-reset.sh: cleans ~/.claude/plugins/marketplaces/agnes,
~/.claude/plugins/cache/agnes, drops uv build cache, documents
workspace-scoped residue that can't be enumerated from the script.
- app/web/setup_instructions.py: legacy AGNES_DEBUG_AUTH path also
uses clone (direct HTTPS marketplace add is broken end-to-end on
every Claude Code distribution -- stores response as single file,
plugin source paths then 404).
28 new tests (test_cli_refresh_marketplace.py) + extended hook + setup
template tests cover bootstrap, fetch+reset ordering, version-aware
reconcile, project-path filtering, hook JSON shape, and the bash-c
Windows wrapper invariant.
Add allow_stopping_for_update=true on google_compute_instance.vm. Without
it, a TF change to machine_type triggers ForceNew (destroy + recreate);
with it, the provider stops + mutates + restarts the VM in place, which
is what an operator resizing a running deployment expects. Tag as
infra-v1.7.0; consumers opt in by bumping the module ref.
See CHANGELOG.md for the full entry. (Bumped from 0.42.0 to 0.43.0 since
0.42.0 was taken by PR #208's backtick-rewriter fix during this branch's
review cycle.)
Single hook entry: 'agnes self-upgrade --quiet ... || true; agnes pull
--quiet ... || true'. Shell semicolon guarantees ordering across every
Claude Code version (no reliance on undocumented multi-hook execution
semantics); each segment's || true preserves the original property
that an upgrade failure does not abort the pull.
Reuses cli.update_check.check() for the version probe — extended with
bypass_disabled=True so explicit user-typed self-upgrade is not silenced
by AGNES_NO_UPDATE_CHECK (which is for the implicit warning loop).
Install path: uv tool install --force when uv is on PATH; otherwise
curl + pip via sys.executable (NOT system python3, NOT --user — both
would land outside the agnes venv and silently no-op the upgrade).
Smoke test execs the binary at the install-resolved path (uv tool dir
joined with agnes-the-ai-analyst/bin/agnes, or sys.executable's sibling
agnes for pip) — never via shutil.which, which can resolve a stale shadow
on PATH and produce a false-positive smoke pass on the OLD version. Smoke
also asserts --version output contains info.latest via PEP 440 Version()
equality (so 0.40.0 does not falsely match 0.40.10).
On smoke fail: rollback to last_known_good.json (written only after a
previous run's smoke passed). Rollback rc is captured and surfaced on
stderr if it also fails. First-ever upgrade or unrecoverable rollback
prints the canonical bootstrap recovery: curl -fsSL <server>/cli/install.sh | bash.
AGNES_SELF_UPGRADE_IN_PROGRESS=1 is set for the duration of the run
and propagated to the smoke-test subprocess. Layer B's _check_version_headers
honors the sentinel and skips the < min hard-stop, so an in-flight
upgrade can never sys.exit(2) itself.
--force invalidates the update_check cache BEFORE probing. --force +
offline = exit 1 with explicit stderr (without --force, offline is silent).
--quiet suppresses progress output but never gags failure stderr.
Every API response is inspected via httpx event_hooks. When the server
reports X-Agnes-Min-Version > local, CLI prints a remediation message
and exits 2. Latest-version drift continues to be handled by the
update_check warning loop — no double-warning on every API call.
Adds X-Agnes-Latest-Version and X-Agnes-Min-Version headers to every
/api/* response. CLI consumes these to hard-stop on incompatible drift.
MIN_COMPAT_CLI_VERSION ships at 0.0.0 — no enforcement until a deliberate
wire-protocol break bumps it.
Also dedupes app version logic: app/main.py:_app_version() helper deleted,
replaced by app/version.py:APP_VERSION as the single source of truth.
test_app_version.py rewritten to target app.version.
`_rewrite_user_sql_for_bigquery_query` does its own bare-name detection
(mirroring the non-RBAC parts of `_bq_guardrail_inputs`). The backtick
masking from #201 was applied to `_bq_guardrail_inputs` and the
forbidden-table loop, but missed this third site — so a registered
local-mode table name appearing as the table segment of a
user-supplied full backtick path (e.g. ``\`prj.ds.orders\`` matching
registered local ``orders``) tripped the cross-source guard and
forced every backtick-path query into the 50-100× slower
ATTACH-catalog fallback.
Mask once at the top of the function, route both the BQ-name
detection (line ~830) and the cross-source check (line ~867) through
the masked copy. New regression test
`test_local_name_inside_backtick_path_does_not_trip_cross_source`
proves the wrapper now wraps when it should.
Adds Pass 3 to `_bq_guardrail_inputs` that scans user SQL for full
backtick paths `<project>.<dataset>.<table>` and gates them
identically to the `bq."<dataset>"."<table>"` pass:
- Project must match the configured BigQuery data project
(`get_bq_access().projects.data`). Mismatch → HTTP 403
`bq_path_cross_project`.
- Path must point at a registered row. Unregistered → HTTP 403
`bq_path_not_registered`.
- Non-admin caller must hold a grant on the registered row's id.
Missing grant → HTTP 403 `bq_path_access_denied`.
Pre-fix, full backtick paths bypassed Agnes RBAC entirely — only the
service account scope limited reach. Post-fix the boundary matches
what `agnes catalog`-driven flows already enforce. Admin still
bypasses the per-id grant check but cannot bypass registration or
project match.
Pass 3 also seeds `dry_run_set` for resolved registered paths so the
cost-cap dry-run runs against the same physical table the user named
— composing cleanly with the Layer 2 fail-fast fallback.
minasarustamyan
