#015 — deeply nested JSON raised RecursionError (a RuntimeError, not
ValueError) and became an unhandled 500 on two paths:
- json-kind content validation now catches RecursionError -> 400.
- Added a custom DRF EXCEPTION_HANDLER
(artifacts/exception_handlers.py, wired via REST_FRAMEWORK) mapping
otherwise-unhandled RecursionError -> 400, covering the parser-level
crash on deeply nested request bodies (metadata etc.) too.
#016 — wrap_csv_viewer handed Grid.js rows without reconciling width to the
header, so over-long rows silently dropped cells. Now normalizes every row
to the max width: long rows widen the columns (auto-named "Column N"),
short rows pad with "". No cell is dropped.
Both verified live; 4 regression tests added. 106 tests pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
markdown/deck/html_fragment rendered author-influenced content verbatim,
so <script>/<img onerror>/etc. executed on the Daybook origin — self-XSS
against the owner's session, and (post-sharing) unauthenticated via /s/.
Two layers, both in artifacts/views.py:
- Create-time allowlist sanitization via nh3 (maintained ammonia/Rust
sanitizer, not deprecated bleach). Keeps formatting/tables/links/images
(incl. relative /assets URLs) and codehilite markup; drops script, on*
handlers, style, unknown tags, and javascript:/data: URLs. html_full is
deliberately NOT sanitized (documented verbatim escape hatch).
- Serve-time CSP (script-src 'none') for the sanitized kinds, on both
/a/{id} and the unauthenticated /s/{token} route. Being a response
header, it also neutralizes scripts in artifacts stored before this fix.
Not applied to html_full or json/csv (json/csv data is already escaped;
their viewers need their own JS).
Verified live across all kinds; 7 regression tests added. 102 tests pass.
Not addressed (noted in the issue): CSV formula injection,
CSRF_COOKIE_HTTPONLY (moot with XSS closed), nonce CSP for json/csv/full.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Uploading a .xlsx file no longer becomes a generic downloadable binary -
it's read server-side with openpyxl and converted to a csv artifact,
rendering in the same interactive grid as a native CSV post. Detected by
content type or .xlsx extension (some clients send generic/incorrect
content types for multipart uploads). Only the active sheet converts;
formula cells use whatever cached result the spreadsheet software last
saved (openpyxl doesn't evaluate formulas itself). Corrupt/unreadable
files return 400.
Since the existing posting script already forwards any non-text file via
multipart with no special-casing, `daybook_post.py balance_sheet.xlsx`
already works with zero script changes - verified live.
Two real bugs found and fixed while building this, both now covered by
regression tests:
- {**a_querydict, ...} silently wrapped every value (including title) in
a list - Python's ** spread on a dict subclass bypasses QueryDict's
__getitem__ override and exposes its internal one-list-per-key storage
directly. Fixed by building the delegated dict explicitly via .get().
- Pre-setting None for absent optional fields (e.g. source_prompt) defeated
create_text_artifact's own data.get(field, default) calls, since .get()
only falls through to its default when the key is *missing*, not when
the value is None - caused a NOT NULL constraint failure. Fixed by only
setting keys that were actually provided.
- csv.writer's default "\r\n" terminator was getting doubled by
Path.write_text()'s own newline translation on Windows, producing a
blank line between every row. Fixed with lineterminator="\n".
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Any artifact can now be made publicly viewable via a share link, no login
required - opt-in per artifact, toggled from the feed UI next to Keep.
- Artifact.share_token: nullable, unique, high-entropy (secrets.token_urlsafe(32)).
NULL = not shared; SQL treats each NULL as distinct so uniqueness still
holds for the many not-shared rows.
- New public route GET /s/<token> (shared_artifact_view), completely
unauthenticated, serving the same rendered/raw content as the owner
route via a shared serve_artifact_file() helper.
- toggle_artifact_share: generates/clears the token: mirrors
toggle_artifact_keep's pattern.
- Fixed a real gap this surfaced: embedded images in a shared artifact's
content would otherwise break for an anonymous visitor, since asset_view
was owner-only. It now also allows access when the asset's parent
artifact is currently shared - unattached (orphaned) assets stay private
regardless.
- serialize_artifact now reports is_shared/share_url so the API reflects
sharing status (read-only for now; toggling stays UI-only).
- Feed UI: Share button (mirrors Keep's toggle interaction), a copyable
share-link box that appears when shared, and a "shared" badge on the
artifact row - distinct blue accent from Keep's green so both can be
visible on the same row without confusion.
Verified live end-to-end (toggle on/off, public access, fake-token 404,
old-link-dies-on-unshare, embedded-asset access before/after sharing)
before writing the 12 automated tests in tests/views/test_sharing.py.
Also fixed two unrelated stale spots noticed in docs/api.md while adding
the Sharing section: the example response was missing is_kept/size_bytes,
and the raw-content table didn't list json/csv (added earlier this
session but never documented there).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
TEXT_KINDS and TEXT_SUFFIXES now include json/csv, with matching
sniff_text_kind branches so both explicit --kind and positional auto-mode
work. Added validate_payload's JSON shape check mirroring the server.
Updated SKILL.md/docs/llm_posting_helper.md with real usage examples,
replacing the stale --as-markdown workaround guidance (kept only for
genuinely unstructured code, which still has no dedicated kind). Also
fixed an unrelated stale doc line found in the same table: html_full's
validation was documented as requiring <!doctype AND <html, but only
<html is actually required.
Verified via dry-run (explicit --kind, auto-mode, invalid-JSON rejection)
and one real end-to-end post through the script against the live dev
instance.
Added tests/scripts/test_daybook_post.py - 17 tests, plain unittest, zero
new dependencies, loaded via importlib since the script isn't a package.
Closes the test-coverage question the issue raised.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
These accepted any JSON type on both POST and PATCH, silently mangled
into Python repr() strings at save time (a dict became the literal text
"{'a': 1}", unparseable as JSON) - and the immediate response echoed the
raw pre-save value, while a later GET returned the mangled one, an
internally inconsistent API.
Added validate_string_fields(), rejecting non-string values with 400
before they reach save(), matching #005's approach for tags. Verified live
and with new tests.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Gap 1 from planning/2026-09-02-artifact_kind_gaps_for_frequent_agent_output.md.
Chose new first-class kinds over riding on html_fragment, resolving that
doc's open design question in favor of filtering support, raw-source
preservation (like markdown's .md sidecar), and server-owned validation.
- json: validated with json.loads at post time (400 on invalid JSON),
rendered with renderjson (collapsible tree), vendored locally.
- csv: parsed server-side with the stdlib csv module into columns/rows and
handed to Grid.js as plain data - no client-side CSV-parsing plugin
needed, keeping normalization server-owned. Vendored locally (MIT).
- Both preserve their original text as a sidecar file, generalizing what
was previously markdown-only handling (storage.py's SIDECAR_SUFFIXES) so
disk-size accounting, eviction cleanup, and ?raw=1 all cover json/csv too.
- Raw JSON is embedded in a <script type="application/json"> block with
`</script` occurrences escaped (safe because JSON's \/ escape round-trips
correctly) - verified with a live </script>-injection payload against the
actual running server, not just unit tests.
New migration (0012) for the KIND_CHOICES change - SQLite doesn't enforce
choices at the schema level, but Django's migration state tracks it.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Enables codehilite on the markdown renderer (Pygments, 'friendly' style,
vendored as static/themes/pygments.css) so fenced code blocks get real
syntax highlighting with zero client-side JS - matching the "themes are
CSS-only" stance in framework_artifact_rendering.md. Verified live against
the running dev instance: a fenced python block renders with highlighted
keyword/string/function spans.
This was Gap 2 from planning/2026-09-02-artifact_kind_gaps_for_frequent_agent_output.md,
picked as the first fix since it needed no new kind or design decision -
just enabling an existing markdown extension.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- #006 (high, data loss): make_room_for() no longer deletes unkept
artifacts on a doomed request - checks whether eviction can possibly
succeed before deleting anything. Also fixed a related latent bug found
while testing: the function returned evicted Artifact objects after
calling .delete() on them, so their .id was always None (Django clears
pk on delete); now returns ids captured beforehand.
- #003: added server-side field-length validation (source/title/
source_name/source_id/theme) on POST and PATCH, matching the model's
own max_length limits, which were previously unenforced on the ingest
path entirely.
- #004: ported daybook_post.py's content-shape validation
(html_fragment/html_full/deck/markdown) to the server, closing the gap
for any client that doesn't go through the helper script.
- #005: PATCH now rejects wrong-typed `tags` (previously silently wiped
existing tags) and rejects immutable fields (content/source/kind/day)
with 400 instead of silently no-op'ing them.
- #002: deleted the misleading skills/daybook-posting/.env.example (a
zero-dependency script that never reads .env files) and clarified
SKILL.md/docs/llm_posting_helper.md that these must be real
process/shell env vars.
Adds tests/unit/ for storage-layer tests. 46 tests passing (was 25).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Peter needed a second user for throwaway testing without going through
Django admin each time. Superusers now get a "Users" tab in Settings
(hidden for regular users) with a create-user form and a list of existing
accounts. New users are regular (non-staff, non-superuser) by default, with
an optional "Grant admin access" checkbox; creating one immediately issues
their API token and displays it once, since that's the credential an agent
account needs and a human account can otherwise regenerate for themselves
later from their own Settings -> Account.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Daybook: a self-hosted, multi-user Django + DRF app where agents, scripts,
or people publish artifacts and each logged-in user browses their own
time-ordered, label-filterable feed.
This squashes prior local development history (kept privately on a separate
branch) into a single clean starting point for the public repository.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
