* feat(pipeline): rotating refresh + patch/apply pipeline, drop lists/
Replace check_paid / check_alive / update_reviews / update_status: each
crawled all 2,600+ pages sequentially, overran its job timeout, saved
only at the end and has lost every run since late June (unreported:
timeouts end as "cancelled").
- scripts/refresh.py: daily slice of ~1/7 of the catalog (pending
strikes first, then least recently checked), one GET per game for
alive / paid / rating / status (+ release_date, thumbnail, updated_at
backfill). Removal needs the same 404/410 or paid strike on two
different UTC days. Paced requests, capped Retry-After back-off, stop
after 3x 429, checkpoints and an in-script deadline. --full replaces
force_update.py (rotating batches, editable fields preserved).
- Scanners emit a URL-keyed patch (patch.py); apply_patch.py applies it
to the latest main and runs validate.py; bash/commit_push.sh fetches,
re-applies and retries pushes, so concurrent writers never clobber
each other.
- update_info.py: canonical URLs (canonical.py), skip games in the
deleted log, keep transient failures queued (max 3 runs, tracked in
scripts/state/), stamp added_at, never drop links queued meanwhile.
- data_store.py: games stay in their chunk, re-slice only past
thresholds, write only changed bytes; index/count_history change only
with data (no more no-op commits).
- validate.py (schema, enums, canonical/unique URLs, index consistency)
+ normalize one invalid safe_virus ("y" -> "Yes").
- tests/: 85 network-free pytest cases, HTML fixtures (from sample/),
shared URL vectors and golden chunk/serialization fixtures for the
future Worker port.
- Workflows: refresh.yml (scan -> apply -> notify), update.yml (runs on
queue change), force_update.yml (refresh.yml full mode),
notify-failure.yml (Discord on failure/cancel/rate limit),
python-ci.yml, webapp-ci.yml, deploy-status.yml; draft-first release
with one create-release job and tauri-action v1, no caches in release
builds; SHA-pinned actions, permissions: {} defaults, reusable
workflows pinned with explicit secrets, injection fixes; dependabot.yml.
- Remove lists/, generate_md, generate_table, log_deleted,
deleted_games.txt, the Telegram bot-ingest path and old bash wrappers;
update README (EN/VI), CONTRIBUTING, templates, CLAUDE.md, CHANGELOG
and mark the retired Telegram path in the policies.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(pipeline): review findings on the refresh / ingest rewrite
Removals
- Mass-change guard: a run whose removals or new strikes exceed
max(10, 10%) of the games it checked withholds all its removals and
alerts (mass_change). Withheld games are marked withheld_at and stay
withheld in later (even single-URL or rate-limited) runs until a healthy
check clears them or a run is dispatched with allow_mass_removal.
- Checkpoints never carry removals, so a scan that dies before its guard
cannot delete games; apply_patch re-checks the per-run limit and caps
any patch at max(25, 5%) of the catalog (ALLOW_MASS_REMOVAL=1 overrides).
- Strikes mature by time, not calendar day: a second sighting must come
>= 20 h after the first (strike_at), so two checks around midnight UTC no
longer remove a game.
- Overlapping scans can't undo each other: refresh_state merges keep the
newest check and the newest full stamp in either apply order, and a
removal is skipped when a newer check has landed since.
- A corrupt count_history.json fails the apply instead of being replaced
by a one-row history.
Ingest
- A second spelling of a queued game no longer drops the first copy
before it reaches a final outcome.
- A dispatched url is processed first and kept in temp_link.json
(queue_add) until it finishes, so a retry, deadline or rate-limit stop
keeps it.
- Failures count as an attempt only >= 6 h apart and a 429 counts as none,
so a burst of queued runs can't burn a link's three attempts.
- Non-string queue entries are cleaned on every apply instead of crashing
it; {"url": ...} objects are unwrapped.
CI
- queue: max on the ingest, refresh-apply and force-update concurrency
groups: the default queue replaced (and silently dropped) pending runs.
- Seed scripts/state/*.json and mkdir -p before git add (the pathspec
failed on an empty patch until the directory existed on main).
- Cancelling a refresh / force-update / desktop release really stops it
(!cancelled() instead of always()).
- deploy-status.yml becomes an hourly check of main's latest finished
"Workers Builds: free-games-itchio-list" run (bash/cf_build_status.sh):
GitHub never starts check_run workflows for GITHUB_TOKEN data commits.
- The reusable release notifiers still take their script from
poli0981/.github@main; the caller comments now say so.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(pipeline): close the remaining mass-removal and monitor gaps
From the second review round of the pipeline fixes:
- A strike spike now marks every strike of that run withheld_at (the
checkpoint of a run that dies mid-spike too), so a later short,
rate-limited or single-URL run can't remove them without an
allow_mass_removal dispatch.
- Withheld games take at most half of a run's budget (all of it for the
allow_mass_removal run that confirms them), and a withheld full-mode
game moves its `full` stamp, so neither rotation stalls on them.
- refresh.yml runs one scan -> apply pipeline at a time (workflow-level
concurrency refresh-pipeline, queue: max), so an older removal can't be
applied after a newer healthy check of the same game.
- deploy-status.yml alerts once per failed Cloudflare build, keyed on its
check-run id and marked only after the Discord post, instead of by age
(which could drop an alert after a missed run, or repeat it); the
monitor walks main's first-parent history so preview builds of merged
branch commits are never mistaken for the site's build.
- CLAUDE.md: merge PRs with squash or a merge commit, never by fast-forward.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>