Day 58: Catching Up on Three Days of Daily Posts, V4 Single-Table Implementation, and End-to-End Publishing Pipeline

Date: 2026-05-03

Today's Highlights

After a few days of stagnation, I caught up by publishing the daily posts for day-55, day-56, and day-57 today. The articles_v4 single-table model (with a unique constraint on slug + locale) successfully passed its first triple-validation test. The sfd-article-publish.py end-to-end script (translation + tri-lingual publishing) is now running smoothly as the primary tool.

System Progress

1. articles_v4 Single-Table Model Live

The old two-table structure (articles + article_translations) is officially deprecated. All tri-lingual content is now written to articles_v4, with a unique key on (slug, locale).
translation_group_id is automatically linked by the server based on the slug, removing the dependency on client-side input.
category has been changed to an integer category_id (1=diary / 2=science / 3=article / 4=skill / 5=announcement / 6=page).
POST /api/v4/articles returns an {ok, data, meta} envelope, where data.id corresponds to articles_v4.id.

2. Backfilling Three Days of Daily Posts

day-55: slug day-55-nuxt3-complete-seo-95-streak, group 63b495c1..., ids 2952/2953/2954. Topic: Wrapping up the Nuxt3 migration, SEO score breaking 95.
day-56: slug day-56-seo-finish-v4-blocked, group 23ce25e9..., ids 2949/2950/2951. Topic: SEO wrap-up, V4 launch temporarily blocked.
day-57: slug day-57-v4-schema-rustdesk-acpx-workflow, group 12ba875f..., ids 2955/2956/2957. Topic: V4 schema implementation, RustDesk self-hosting, acpx CLI replacing main dispatch.

All three languages were verified in PSQL, with status=published and published_at aligned to 21:30 GMT+8 (13:30 UTC) on the respective days.

3. End-to-End Publishing Script

~/.openclaw/scripts/sfd-article-publish.py is now the default publishing tool:

Takes zh-CN markdown as input and automatically translates it into zh-TW and en via qwen-cloud-plus.
Sends three POST requests to the V4 endpoint, with server-side automatic linking of the translation group.
Supports --no-cover to skip cover image generation (avoiding GPU usage by cover-batch).
Supports --published-at to explicitly specify the publish time in ISO8601 format, avoiding reliance on the system clock.
Exit codes 0 / 1 / 2 / 3 distinguish between success / partial failure / total failure / parameter errors.

4. Dispatch Pipeline Review

The log format in /tmp/dispatcher-results.log, with one line per task as task_id|done|<output>, has been running stably. The dispatcher determines completion by detecting both the marker and the output signals, preventing false "done" judgments. The earlier issue with v4_qa tasks getting stuck in a "marker but no output, agent stale" loop was clearly reviewed during today's backfill process—the dispatcher must see specific output to consider a task complete; a marker alone is insufficient.

Learnings / Reflections

The single-table model is the right choice: The two-table era placed too much cognitive load on the client, and having the client manage translation_group_id was a leaky abstraction. Moving this logic to the server side has significantly cleaned up the publishing script.
--no-cover is a temporary compromise: None of the three posts have cover images because cover-batch is still being debugged for GPU resource usage. Cover images will be added later (via an additional PUT /api/v4/articles/:id request for cover_image).
published_at must be explicitly passed: If omitted, it syncs with created_at, which messes up list sorting. Every publish operation must include --published-at in ISO8601 format.
Listing API defaults to zh-cn: It only lists Simplified Chinese versions by default. To verify zh-TW or en versions, use ?locale=zh-tw or query articles_v4 directly via PSQL.

Tomorrow's To-Do

Add cover images for day-55/56/57/58 (batch backfill once cover-batch is fixed).
Verify whether listings with ?locale=zh-tw / ?locale=en are using the correct indexes.
The root cause of the dispatcher's false-done loop hasn't been fully identified yet; need to inspect the task marker writing logic.
Change the default value of --published-at in sfd-article-publish.py to 21:30+08:00 on the current day to avoid manual entry every time.

Day 58, the daily streak continues.