* feat(images): route Cloudflare Workers AI via AI Gateway by default
Direct probe on 2026-04-21 confirmed the Cloudflare AI Gateway
route to Workers AI (`gateway.ai.cloudflare.com/.../workers-ai/@cf/...`)
is no longer 403ing — the upstream regression noted in v0.8.0's
ARCHITECTURE.md #16 and in thread image-pipeline seq 6 has been
resolved. Routing via the gateway gives us the same benefits we
already get for OpenRouter LLM calls: response caching, cost/latency
dashboard, rate limiting.
Changes:
- `Cloudflare_Image_Support::build_endpoint_url()` now prefers the
gateway URL derived from the existing `prautoblogger_ai_gateway_base_url`
option. Direct API remains the fallback.
- New private helper `build_gateway_workers_ai_url()` strips the
OpenRouter path suffix from the configured gateway URL and appends
`/workers-ai/{model}`. Returns empty string (fallback to direct) when:
no gateway URL configured, the new toggle is off, or the URL isn't a
valid `https://gateway.ai.cloudflare.com/...` shape.
- New admin toggle `prautoblogger_cf_image_via_gateway` (default on) in
the Images section. Operators can disable to force direct API without
clearing the LLM-path gateway URL.
- ARCHITECTURE.md #16 rewritten to reflect the v0.8.2+ gateway routing.
- CHANGELOG.md v0.8.2 entry under Changed.
Tests (new `CloudflareImageSupportTest`):
- Direct URL when gateway not configured.
- Gateway URL derived correctly when configured + toggle on.
- Toggle off forces direct URL even with gateway configured.
- Malformed gateway URL falls back to direct (defensive).
- `is_nsfw_error()` only matches error code 3030 (pinned separately so
a tolerance regression fails early).
Cost impact: near-zero. Gateway adds a few ms of latency and enables
response caching — could *reduce* cost on identical prompts within the
configurable cache TTL.
Uninstall: new option covered by existing prautoblogger_% LIKE sweep.
Implements Commit 1b (narrow AI-Gateway-migration scope) per thread
image-pipeline seq 7.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* chore(cost): seed new toggles in activator defaults; tighten URL stripper
QA review on cd45ae2 surfaced two things; one real, one hallucinated:
- **Real (fixed):** `prautoblogger_cf_image_via_gateway` and v0.8.1's
`prautoblogger_image_nsfw_retry` weren't seeded in
`Activator::set_default_options()`. Belt-and-braces — the `get_option()`
fallback at the read site already defaults to '1', so no user-visible
regression exists, but seeding the options follows the established
pattern for every other plugin setting and makes intent legible to
operators inspecting `wp_options` directly.
- **Hallucinated (rejected):** the reviewer claimed
`CloudflareImageSupportTest` methods 2–5 were empty stubs. Verified
— every test has real `$this->assert*` calls (10 total across the 5
cases). Ignoring. Flagging the false-positive to the CTO per the
qa-review quality-tracking protocol.
Also clarified the docblock + tightened the ternary inside
`build_gateway_workers_ai_url()` — explicit `strrpos` false-check
instead of the `?:` short-circuit, so a no-slash base (defensive edge)
falls through cleanly instead of returning a weirdly-sliced URL.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(schedule): interpret `prautoblogger_schedule_time` in site timezone
Per thread scheduled-gen-timezone seq 1: WordPress core calls
`date_default_timezone_set('UTC')` early in bootstrap, so the activator's
`strtotime("tomorrow HH:MM")` was resolving in UTC regardless of the
site's configured timezone. On peptiderepo.com (Asia/Singapore, UTC+8)
a `schedule_time` of `06:00` fired at 14:00 local.
Fold this into the existing v0.8.2 branch (PR #102) since both this fix
and the earlier CF-image-gateway change target the same version and
touch overlapping files (activator, settings-fields-extended).
Changes:
- `Activator::next_daily_generation_timestamp()` — new public helper
that reads `prautoblogger_schedule_time`, constructs a
`DateTimeImmutable('tomorrow', wp_timezone())`, applies the
configured hour/minute, returns a Unix timestamp. Defensive fallback
to the old UTC-interpretation behaviour on any DateTime parse error
so a bad tz string can't brick the schedule.
- `Activator::schedule_cron()` delegates to the helper instead of
calling `strtotime` directly.
- New `Activator::reschedule_daily_in_site_timezone_v082()` — one-shot
migration that clears the stale UTC-scheduled event and re-queues
using the helper. Gated by `prautoblogger_migrated_schedule_tz_v082`
option so it runs exactly once. Logs an INFO line comparing
before/after next_run for auditability.
- `class-prautoblogger.php::on_check_db_version()` invokes the
migration on every admin_init; the one-shot flag handles re-run
guarding (no DB_VERSION bump needed — the migration is independent
of the schema version).
- Settings description for "Generation Time" now says
"...in your site's configured timezone (Settings → General →
Timezone). Current site timezone: <wp_timezone_string()>" instead
of the ambiguous "server timezone" phrasing.
File-size compliance:
- Adding the timezone helpers pushed `class-activator.php` to 352
lines. Extracted the 104-line `create_tables()` SQL to a new
`class-schema-installer.php` (136 lines). Activator now 245 lines,
schema file fully dedicated to dbDelta.
Tests (new):
- `ActivatorScheduleTest` — SGT 06:00 → 22:00 UTC prev day; UTC 06:00
→ 06:00 UTC (no-op case); Asia/Kathmandu UTC+5:45 → 00:15 UTC
(fractional offset); malformed time input doesn't throw.
- `ActivatorMigrationTest` — pre-seeded UTC cron gets cleared and
rescheduled in SGT; second pass is idempotent.
CHANGELOG v0.8.2 entry now also covers the timezone fix under Fixed +
the description wording change under Changed.
Implements thread `scheduled-gen-timezone` seq 1 spec.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: peptiderepo <peptiderepo@users.noreply.github.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(image-pipeline): surface per-image failures in the event log
Two gaps made posts 650/657 appear to silently "lose" their featured
images on 2026-04-20:
1. The batch provider tagged each failed key with ['error' => ...] but
the ERROR log line it emitted didn't identify which key failed, so
operators couldn't tell image_a from image_b in the event log when
one of two parallel requests died.
2. Image_Pipeline::process_image_a/_b folded the per-key error into
$result['errors'] and returned silently, with no WARNING row. Ops
only learned about the miss when the outer catch in the caller fired
— and on the 2026-04-20 incident that outer catch didn't fire, so
nothing was logged at all except an unrelated pricing warning.
Fix:
- class-open-router-image-batch.php: collect_result() takes the request
key as its first argument and every Logger::error message reads
'OpenRouter batch image gen <class> for key "<key>": <detail>'.
Matches the shape of OpenRouter_Image_Support::log_retry.
- class-image-pipeline.php: process_image_a and process_image_b now
emit a Logger::warning('image_pipeline') row whenever $image_data is
missing or carries an error key, in addition to pushing to
$result['errors']. The warning fires whether or not the outer catch
runs.
Test: ImagePipelineTest::test_partial_batch_failure_emits_warning_log_row
mocks $wpdb, runs the pipeline with image_a erroring, and asserts at
least one warning/image_pipeline row was inserted.
No behavior change beyond logging; cost impact zero.
Part of the v0.8.0 image bundle (issues from thread image-mime-bug).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(images): retry Cloudflare NSFW-blocked prompts with a sanitized fallback
Cloudflare's Workers AI FLUX.1 schnell content filter sometimes rejects
peptide-comic prompts as NSFW (HTTP 400, errors[].code === 3030). On
2026-04-20 this produced a silent publish-without-featured-image for
post #650 ("Semaglutide vs Tirzepatide") — the LLM-rewritten scene
tripped the filter and we had no recovery path.
New behaviour: on a 3030 response, the Cloudflare provider throws a
typed PRAutoBlogger_Image_NSFW_Blocked exception. The batch wrapper
catches it and tags the per-slot result with
`error_type: nsfw_blocked`. Image_Pipeline hands the results to a new
PRAutoBlogger_Image_NSFW_Retry helper, which reruns the blocked slot
once with a sanitized fallback prompt (title + style suffix, no LLM
rewrite). On success the retry result replaces the error entry; on a
second block, a WARNING is logged and the pipeline publishes without
that image — same terminal behaviour we had before, just explicit.
Files:
- New providers/class-image-nsfw-blocked.php — typed exception.
- providers/class-cloudflare-image-support.php — is_nsfw_error() +
throw_if_nsfw() helpers.
- providers/class-cloudflare-image-provider.php — calls throw_if_nsfw
inside the 4xx branch; batch catches the typed exception and marks
the result with error_type.
- core/class-image-prompt-builder.php — new public build_fallback_prompt
shim so the retry layer can reach the rule-based fallback without
going through the LLM path.
- New core/class-image-nsfw-retry.php — single-shot retry helper.
- core/class-image-pipeline.php — invokes the retry between batch
completion and per-slot processing.
- admin/class-settings-fields-extended.php — new toggle
`prautoblogger_image_nsfw_retry` (default on) in the Images section.
Tests:
- CloudflareImageProviderTest — 3030 response throws NSFW exception;
batch tags the per-slot result with `error_type: nsfw_blocked`.
- ImagePipelineTest — retry succeeds when provider returns a good
result on the second call; retry is skipped when the setting is off.
Cost impact: worst-case an extra Cloudflare call per NSFW-blocked slot
(~$0.000834). Current incidence ~2% of generations; ~+$0.00002 per
generation amortized. Immaterial.
Uninstall: new option covered by the existing prautoblogger_% LIKE
sweep in uninstall.php.
Part of the v0.8.0 image bundle (issues from thread image-mime-bug).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(images): expose rewriter system prompt as admin setting
CEO feedback: every generated image is anchoring on the cartoon
peptide-with-arms-and-legs character because the hardcoded rewriter
system prompt both lists "anthropomorphized molecules" first in its
primer and ships a few-shot example that literally shows one. The ask
is to give CEO the lever without a code change.
Adds `prautoblogger_image_prompt_instructions` (textarea) in the
Images section. Default value is the existing REWRITER_SYSTEM_PROMPT
constant verbatim, so behaviour is unchanged on upgrade until someone
edits it. `Image_Prompt_Builder::rewrite_via_llm()` reads the option
at call time via a new `resolve_system_prompt()` helper and falls back
to the constant if the option is blank or whitespace-only, so a saved
empty value cannot brick image generation.
Const visibility flipped from `private` to `public` so the settings
field can reference it directly as the field default.
Tests:
- `test_rewrite_uses_setting_when_present` — option set, assert the
LLM's system message equals the option content.
- `test_rewrite_falls_back_to_default_when_setting_empty` — option
whitespace, assert the LLM's system message equals the constant.
Uninstall: new option covered by the existing prautoblogger_% LIKE
sweep in uninstall.php.
Cost impact: none by default. Future CEO edits will change prompt
length; current default is ~180 tokens, budget impact linear from
there.
Part of the v0.8.0 image bundle (thread image-mime-bug, message 03).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(images): collapse Image Provider + Image Model into a single dropdown
Until now, Image Provider and Image Model were two independent selects
in the admin UI. That let the CEO set Model=flux-1-schnell while
Provider was still openrouter — OpenRouter has no such model so every
image gen on posts 650/657 on 2026-04-20 returned HTTP 400, and the
article published without a featured image.
v0.8.0 removes the Image Provider field from the admin UI entirely.
The provider is now derived from the model's entry in the registry
every time the model is saved. A new
PRAutoBlogger_Image_Model_Registry::provider_for() helper centralises
the mapping, so admin save, the on-upgrade migration, and future
callers all read from one place.
On save, PRAutoBlogger_Admin_Page::sanitize_image_model() validates
the submitted slug against the registry:
- Known slug: persist it, and update `prautoblogger_image_provider` to
the paired provider as a side-effect.
- Unknown slug: add_settings_error with a clear message, and return
the previously-saved value so the site stays on a working model.
A one-shot v0.8.0 migration in
PRAutoBlogger::on_check_db_version() derives the provider from the
currently-saved model on first admin_init after the plugin is updated.
This auto-heals sites that were running mismatched pairs before the
upgrade, without requiring the admin to re-open settings.
`prautoblogger_image_provider` option key is preserved — all existing
reads in the pipeline, cost tracker, and validators continue to work
unchanged.
Tests (Services/ImageModelRegistryTest):
- Every registry entry has an id and a provider in {openrouter, cloudflare}.
- provider_for() returns the paired provider for a known slug.
- provider_for() returns '' for unknown/empty slugs.
Cost impact: none.
Migration idempotent via the `prautoblogger_migrated_image_provider_v080`
flag. Uninstall sweeps it along with every other plugin option.
Part of the v0.8.0 image bundle (thread image-mime-bug, Issue 1).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* chore(v0.8.0): version bump + CHANGELOG + ARCHITECTURE.md for image bundle
Bump plugin header and PRAUTOBLOGGER_VERSION 0.7.3 → 0.8.0. Minor bump
because Commit D removes the Image Provider dropdown from admin (a
UX-visible change), not a patch.
CHANGELOG:
- Changed: collapsed Image Provider + Model dropdowns; provider derived.
- Added: Image Prompt Instructions setting; Retry NSFW-Blocked Images
setting.
- Fixed: silent per-image failures in the event log.
ARCHITECTURE.md:
- Stage 6b (Image Pipeline) callout now describes provider derivation
and the NSFW-retry fallback path.
- wp_options table lists the four new/changed keys
(`prautoblogger_image_provider` as derived, new
`prautoblogger_image_prompt_instructions`,
`prautoblogger_image_nsfw_retry`, and the one-shot migration flag
`prautoblogger_migrated_image_provider_v080`).
DB schema unchanged — no PRAUTOBLOGGER_DB_VERSION bump needed.
Final commit of the v0.8.0 image bundle (thread image-mime-bug).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: peptiderepo <peptiderepo@users.noreply.github.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three fixes in one:
1. Duplicate image: prepend_caption_to_post() was embedding the full
image + caption as a <figure> block, but the theme already displays
the featured image. Now only the caption text is prepended — no
second copy of the image.
2. Hardcoded model names: removed all FLUX.1/FLUX.2 references from
model-agnostic code (pipeline, prompt builder, sideloader, OpenRouter
provider, interface, admin labels). The Cloudflare provider still
references FLUX since it specifically wraps that model family.
3. Test alignment: OpenRouterImagePricingTest was asserting against
removed models (flux-2-pro, seedream). Updated to match current
COST_PER_IMAGE entries (Gemini Flash, GPT-5 Image).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replaces Cloudflare Workers AI (FLUX.1 schnell) as the default image
provider with OpenRouter, which offers FLUX.2 Pro, Seedream 3.0,
Gemini Flash Image, and GPT Image 1 at comparable or better pricing.
- New provider class (3-file split: provider, pricing, support) reuses
the existing OpenRouter API key and Cloudflare AI Gateway
- Admin settings: new Image Provider dropdown (openrouter/cloudflare),
model picker shows all available image models from both providers
- Image pipeline auto-selects the configured provider at construction
- Cost logging now uses actual model name instead of hardcoded flux-1-schnell
- Default model changed to black-forest-labs/flux-2-pro ($0.03/image)
- Unit tests for pricing resolution and provider response parsing
Cost impact: FLUX.2 Pro at $0.03/image vs schnell at $0.0007/image —
~43x more expensive per image, but dramatically better quality. At 2
images/article × 1 article/day = ~$1.80/month. Seedream 3.0 ($0.003)
available as a budget alternative.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Both the provider and validator call wp_parse_url() for host
extraction. Tests were missing the Brain Monkey stub.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Extracted credential validation, request building, response parsing, and
configuration helpers from the 464-line OpenRouter provider into separate,
focused classes. This follows the pattern established by the Cloudflare
image provider and brings all files under the project's 300-line limit.
Files created:
- class-open-router-validator.php (152 lines): Handles credential validation
with HTTP probes and diagnostic messaging. Delegates from the provider.
- class-open-router-config.php (84 lines): Centralizes admin config resolution
for API base URL and cache TTL settings. Used by both provider and validator.
- class-open-router-request-builder.php (79 lines): Encapsulates request header
building and cURL auth filter registration for hosting environments that
strip Authorization headers.
- class-open-router-response-parser.php (87 lines): Handles response parsing,
error classification, and retry-decision logic.
- Tests: OpenRouterValidatorTest.php with comprehensive credential checks.
Provider shrinks from 464 to 278 lines. All classes are under 300 lines.
Dependencies remain clean: validator and config are independent, request-builder
and response-parser have no circular imports.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Implements image-provider interface + Cloudflare Workers AI adapter with pricing and validator. Independent QA verdict: APPROVE on 52fd11e (qa-reviews/prautoblogger/2026-04-15-52fd11e2ce97dd9e659df5cc87ec1382a668bec1.md). All 4 CTO blockers resolved (PHPUnit retry mocks, XSS hardening on error strings, PHPCS on new files, 300-line cap via support-class extraction). PHPCS backlog (586 errors across 38 pre-existing files) remains on continue-on-error; Workstream C material.
- bootstrap: add PRAUTOBLOGGER_MAX_RETRIES, API_TIMEOUT, RETRY_BASE_DELAY constants
- BaseTestCase: add wp_json_encode + home_url stubs, reset Logger singleton in tearDown
- CostTrackerTest: use openai/gpt-4o (in MODEL_PRICING), fix log_api_call args,
allow multiple wpdb inserts (Logger may also insert event_log rows)
- OpenRouterPricingTest: add wpdb mock with prefix for Logger access on unknown models
- OpenRouterProviderTest: test send_chat_completion throws RuntimeException when
no API key is configured (matches actual guard clause behavior)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- ContentAnalyzerTest: pass Cost_Tracker as 2nd constructor arg
- IdeaScorerTest: stub get_option for topic_exclusions
- CostTrackerTest: stub get_option for monthly_budget_usd
- OpenRouterPricingTest: stub get_transient with cached model data
- OpenRouterProviderTest: stub get_option for API key + get_transient
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
All 14 test files updated to use correct constructor patterns (array-based
value objects, dependency injection for ContentAnalyzer, etc.) and verified
method names against source.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
