Migration 0004 (users, sessions, config_changes, profile_versions, jobs,
rating_events.user_id, runs.report_json, issues.issue_json), the issue
snapshot writer and loader, the web module skeleton with layout and static
assets, axum-login/tower-sessions over a sqlx session store, password-auth
users with a CLI, the login throttle, the origin check, security headers,
and the public issue pages, archive, Atom feed and robots.txt.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NHyYupFdBiR4VfoUM7NjSM
DeepSeek's content filter rejects a whole request (400 "Content Exists
Risk") when one article trips it, which cost the other articles in the
batch their assessment and retried them every run. New curate/batch.rs
runs both stages through a bisecting runner: a rejected batch is split
until the offending article is isolated, that article is retried once on
the editor provider when it is a different one, and a still-rejected
article is recorded as a provider_rejected assessment row so it is not
retried for assessment_reuse_days. Cache reuse accepts rows from either
configured model. Each stage logs reused/requested/rejected counts, the
curation: line shows rejections when non-zero, explain prints them, and
the llm_assess span reports the deep-set size.
Implemented by a Claude agent from an orchestrator brief; verified
fmt/clippy(-W dead_code)/test green (354 lib tests).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A1rCLQeKBgnBo3oTgHuTMe
[llm] assigns the bulk and editor roles by name over a [providers.*]
registry (kind = openai | anthropic, per-provider model, effort, daily
ceiling and price table); DeepseekBackend becomes OpenAiCompatibleBackend
(reasoning_effort passthrough), AnthropicBackend builds from the same
ProviderConfig, meters and provider_costs are keyed by provider name.
Gemini 3.8 Flash is declared via Google's OpenAI-compatible endpoint so
switching the editor is one line (or DAILY_EPUB_LLM__EDITOR=gemini for an
A/B dry run). Stale [deepseek]/[anthropic] tables, the top-level
max_daily_usd and the old key env vars fail loudly.
daily-epub config check validates and prints the resolved roles, models,
key presence and paths without opening the database.
docs/runbooks/curation-v2-migration.md walks the server upgrade from v1.
Registry implemented by a Claude agent from an orchestrator brief;
verified fmt/clippy(-W dead_code)/test green.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A1rCLQeKBgnBo3oTgHuTMe
Dead code and stale v1 comments removed (clippy -W dead_code clean, the
three world.rs warnings fixed), the Brief chapter's TOC title renamed from
"From the Editor", features prune now also sweeps article_assessments and
generate runs the sweep once after publishing, the example config is
tested key-for-key against Config::default(), README commands match
--help, and docs/plans/2026-08-15-implementation-notes.md records the
Anthropic and Voyage facts, the new tables, the budget-day rule and the
lock.
Implemented by a Claude agent from docs/plans/curation-v2-briefs/step7.md.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A1rCLQeKBgnBo3oTgHuTMe
Behind-the-paper chapter (behind.xhtml, both editions) built from the run's
StageCounts and candidate_runs near misses; daily-epub stats [--days N];
StageCounts gains knn/feed gates and verdicts_in_prompt, timings split
into summaries + brief, the four-line §15.4 info block logged once per
run and printed by print_report; src/lock.rs flock guard on
<database_path>.lock for generate, profile rebuild, features backfill and
backfill-social; README updated for the new CLI, env vars and costs.
Implemented by a Claude agent from docs/plans/curation-v2-briefs/step6.md;
reviewed against plan §5, §15.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A1rCLQeKBgnBo3oTgHuTMe
assess.rs replaces score.rs (DEEP_INSTRUCTIONS, representative sample
with [BEGINNING]/[MIDDLE]/[END], facets, cached deep rows with --rescore
bypass), rank.rs adds the utility blend over present signals with gate
ramps and the leader-clustered shortlist (cap 2 → 3 → uncapped, protected
top-N, exploration reserve), editor.rs replaces select.rs with the §13
rendering and utility-ordered fallbacks. ScoredArticle is gone; Candidate
is the only flow type. deep_batch_size replaces score_batch_size.
Started by Codex (cut off by its usage limit mid-verification) and
finished by a Claude agent from docs/plans/curation-v2-briefs/step5.md;
reviewed against plan §12–§13.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A1rCLQeKBgnBo3oTgHuTMe
DeepSeek triage over every eligible article (plan §10) with cached
assessments in article_assessments, the union admission with quotas and
exploration slots (§11), hygiene moved to admit.rs with the churn rule
reading assessments, prefilter.rs reduced to hygiene and text heuristic,
prefilter_keep removed in favour of curation.ranking.deep_keep, the
scores table dropped (migration 0003), and --rescore on generate.
Implemented by Codex (gpt-5.4, high effort) from
docs/plans/curation-v2-briefs/step4.md; reviewed against plan §10–§11.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A1rCLQeKBgnBo3oTgHuTMe
Union merge of the Claude editor (step 2) and the embeddings/signals/telemetry
work (step 3): both config trees, per-provider LLM budgets plus the Voyage
meter, one startup provider log, runs.config_json with ranking and voyage
settings, and selected picks' why lines landing in candidate_runs.editor_why.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A1rCLQeKBgnBo3oTgHuTMe
- embedding.rs: EmbeddingBackend + VoyageBackend, batched bounded-concurrency
client with its own UsageMeter, f32 BLOB codec, article/interest embedding
cache keyed by model, dimension and sha256 of the embedded text.
- signals.rs: z-scored interest match, decayed rated-neighbour preference
with the knn gate, feed affinity with the feed gate, social, text heuristic
without social terms, mid-rank percentile normalizer, preliminary blend.
- telemetry.rs: candidate_runs writer with §7.5 signals_json, explain and
near-misses renderers, prune.
- [voyage] and the full [curation.ranking] config with validation.
- CLI: explain, features backfill|prune, generate --skip-embeddings.
- Pipeline: hygiene rows, embed and signals stages before the old prefilter;
same-date regeneration no longer excludes its own picks.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A1rCLQeKBgnBo3oTgHuTMe
- AnthropicBackend (Messages API, cached system block, output_config.effort,
server-side fallbacks, refusal surfaced as an error); Llms { bulk, editor }
with editor_or_bulk(); PriceTable-based UsageMeter per provider.
- [anthropic], [editorial], deepseek.max_concurrent_requests and
curation.max_article_count config; startup logs resolved providers.
- Budget day is the UTC date of started_at, preloaded from
runs.provider_costs_json; finish_run writes provider_costs_json and
config_json. Stage A batches run concurrently with per-batch budget checks.
- Editor prompt with one-line "why" per pick; no minimum lineup size;
--max-articles is a ceiling; top-up branch deleted; why stored on picks and
issue_articles.why and rendered in chapters and In this issue.
- Summaries on the editor client (3k-token input, concurrency 4, bulk then
excerpt fallback); "The Brief" replaces From the Editor; section intros gone.
- Colophon carries per-provider costs and models.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A1rCLQeKBgnBo3oTgHuTMe
- Migration 0002: rating_events, article_embeddings, interest_embeddings,
article_assessments, candidate_runs, runs.config_json/provider_costs_json,
issue_articles.why; copies ratings into rating_events and drops ratings and
feed_priors (scores stays until step 4).
- Vote is Loved | Good | NotForMe; legacy `up` links still verify as Loved.
- Footer offers Loved it / Good / Not for me; the confirmation page offers the
other two so a mis-tap can be corrected. handle_rating appends one event.
- db::current_ratings implements the latest-explicit-event rule with summaries
and facets; `ratings list|set|clear` CLI appends source='cli' events.
- data/profile.md replaces the hard-coded reader prose; the system prompt is
rebuilt every run in the §8.4 order with a recent-verdicts block.
- Weekly rebuild reads summaries, facets and notes; feed priors removed.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A1rCLQeKBgnBo3oTgHuTMe
The image fixes left two problems of navigation. Image handling was
spread across `extract.rs` (300 lines of normalization) and
`epub/images.rs` (download, re-encode, markup rewriting, plus generic
HTML helpers that comments, world and x4 were all reaching into a
module named "images" to borrow). And `epub/build.rs` had grown to
1,148 lines of code covering cover rendering, ten askama templates,
every chapter renderer and the zip assembly.
New homes:
- `src/html.rs` — markup helpers that do not care what the markup is
about: tag scanning, attribute parsing, entity decoding, escaping,
XHTML fixups, reading a fragment as text. Previously scattered between
`epub/images.rs` and `extract.rs`.
- `src/images/` — one module per stage of an article's images, in the
order they run: `normalize` (make `<img>` usable, pre-readability),
`refs` (what an article references), `fetch` + `encode` (download and
re-encode per edition), `embed` (point the markup at what shipped).
- `src/epub/{cover,chapters,build}.rs` — the cover, the chapter
renderers, and the ordering plus assembly that puts them together.
`epub/fixtures.rs` takes the shared test issue, which was a public
module wedged inside `build.rs`.
- `src/curate/profile/themes.rs` — a 260-line keyword table that sat in
the middle of the profile logic.
`curate::html_to_text` is renamed `prompt_text`: it is a different
function from `html::html_to_text` (collapses whitespace, no DOM, sized
for prompt budgets) and sharing a name with it was a trap.
Largest module drops from 1,148 code lines to 828, and no file mixes
two subjects. Behaviour is unchanged: 231 lib tests plus 25 integration
tests green, and the real-world audit over issues 1–3 still reports 214
images referenced, 214 shown, 0 placeholders, 0 orphaned assets.
`image_audit` gains `--epub-out DIR`, which writes a readable EPUB of
the audited articles so images can be checked on a device.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Across the first three issues, 145 of 208 referenced images reached the
page; 38 that had been downloaded, re-encoded and written into the EPUB
were never referenced by any chapter, and 69 became `[image: …]` lines.
Six independent causes, each verified against the real articles:
- **Entity-encoded URLs never matched their asset.** `prepare_body`
re-runs `ammonia::clean`, which writes `&` as `&`, and the tag
scanner that looks the asset up compared raw attribute text against a
URL a real parser had decoded. Every image with a query string lost.
`parse_attrs` now decodes entities.
- **Readability clobbers a working `src`.** Its lazy-image heuristic
copies any attribute containing `.jpg` over `src`, which on NPR meant
a `{width}` template (the CDN answers those with a grey square reading
"Image"), on Substack a JSON blob, on dfarq an entire `srcset` string.
Pages are now normalized before readability sees them: every `<img>`
is reduced to `src`/`alt`/`title` with the best candidate from the
lazy attributes, `src`, `srcset` and `<picture><source>`, so there is
nothing left for the heuristic to substitute. Candidates that cannot
resolve — braces, whitespace, quotes, `data:` — are rejected by shape
rather than by publisher.
- **Readability deletes `<button>` and its subtree**, taking lightbox
images with it and leaving the captions behind. Image-only wrappers
are unwrapped first.
- **SVG was undecodable**, so vector charts became placeholders. They
are rasterized with resvg, which the cover already depends on.
- **Relative URLs resolved against the pre-redirect URL**, 404ing every
image on an article reached through a shortener. `fetch_readable` now
reports where it landed.
- **The 12-image cap** silently truncated photo essays, and the images
past it were advertised as failures. Removed; the issue-wide byte
budget is the real backstop.
An image that still cannot be embedded is now dropped rather than
announced, unless its alt text is a real description — decorative rules,
spacers and dead links were generating most of the placeholder noise.
`examples/image_audit.rs` replays the pipeline over the articles of
published issues and reports what reaches the page. On issues 1–3:
208→214 images referenced, 145→214 shown, 69→0 placeholders, 38→0
orphaned assets. The three images still not embedded are a 14×14 favicon
and a 650×2 divider — correctly declined, and no longer announced.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
- World Briefing now starts with yesterday, preserves complete nested
event hierarchies and Wikipedia links, fetches deduplicated article
leads concurrently, and adds stable-ID summaries plus a daily overview.
- Enrichment runs after normal editorial work and degrades safely with
report warnings.
- Colophon facts now use X4-safe block paragraphs.
- Standard cover remains 1200×1600 RGB PNG.
- X4 cover is a 480×800 baseline RGB JPEG with consistent XHTML and OPF
declarations.
- Added CrossInk cache-clearing guidance to README.md.
- Worked around an epub-builder duplicate XML-ID defect discovered by
validation.
It turns out it is currently impossible for the Xteink X4 to download
XTC files from an OPDS server. So I switched the built-in OPDS server in
this binary (daily.hallada.net/opds) to serve the EPUB files instead of
the XTC files. The X4 is pretty capable of reading the X4 edition of the
EPUB that was optimized for it anyways, and I prefer the flexibility of
EPUB, so I might end up eventually deleting the XTC conversion. For now,
I kept the conversion step (in case I ever want to ever try manually
copying them to the device) and I keep a limited number of XTC issues on
the server since they are quite big in filesize. The BookOrbit
integration is now optional since the built-in OPDS server is able to
serve the EPUBs. In my installation, I serve the EPUBs through both. The
daily.hallada.net/opds server is just a little quicker to navigate and
download the EPUBs on the X4 since the BookOrbit OPDS requires diving
into a couple layers of folders before you get to the files.
Full implementation of a personalized daily newspaper delivered as an
EPUB.
Articles are pulled from a local self-hosted Miniflux instance, enriched
with comments, summarized and filtered by DeepSeek AI, and then
assembled into two EPUB editions: standard and optimized for the Xteink
X4 e-ink reader. Both are served by the local self-hosted BookOrbit OPDS
server in a separate library. Then the X4 edition is futher converted to
XTC format and served over a separate OPDS server hosted by the Rust
binary. Runs are tracked in a local SQLite database so runs are
idempotent per date.
Full documentation of the plan is in docs/plans and setup and install
instructions are in the README.md file.
thalladaandClaude Fable 5.1