HappyPet Pipeline Audit Findings

Date: 2026-05-19 Auditor: Maeve (mobile session) Method: Full file reads — generate_posts.py (50KB truncated), post_pins.py, push_pins_to_sheets.py, generate.yml, publish.yml, pin.yml + skills: vibe-code-auditor, python-pro, bash-pro, maeve-github-devops, engineering-craft, jekyll-affiliate, avoid-ai-writing, beautiful-prose

CRITICAL

C1 — Fact-check truncates to 1500 chars (P3)

File: generate_posts.py, fact_check_alternatives() Code: content_fc = content[:1500] if len(content) > 1500 else content Impact: 7000-char articles fact-checked at 21%. Fabrications in sections 2-7 are never caught. Both primary and fallback paths receive the truncated prompt. Fix: Check full article. Chunk at 3500-char overlapping windows if model context requires it. Never truncate below 60% of article length.

C2 — Generator exits 0 on total provider failure (P4)

File: generate_posts.py main(), generate.yml Impact: GHA reports conclusion: success even when 0 articles were generated. Every “silent failure” run you’ve ever seen is this. Fix: Write GENERATION_RESULT.json at script end with articles_generated, articles_held, articles_failed. GHA step reads it and exit 1 if both generated and held are 0.

C3 — Fact-check routes OpenRouter model to Groq endpoint (NEW — not in problem map)

File: generate_posts.py, fact_check_alternatives() Code: OR_FACTCHECK_MODEL = "openai/gpt-oss-20b:free" sent to GROQ_URL with groq_key auth. Impact: Groq API rejects the unknown model. Primary fact-check fails on every roundup article. Gemini fallback catches it silently. No error surfaces in logs beyond “Fact-check primary failed.” Fix: Route OR_FACTCHECK_MODEL to OPENROUTER_URL with OPENROUTER_API_KEY. Groq models (if re-enabled) go to GROQ_URL. Never mix endpoint and model families.

C4 — GHA action versions do not exist

File: generate.yml, publish.yml, pin.yml Code: actions/checkout@v6, actions/setup-python@v6, actions/upload-artifact@v7 Impact: These versions don’t exist. Latest stable: checkout@v4, setup-python@v5, upload-artifact@v4. GHA resolves unknown versions unpredictably — may pin to latest or fail. Fix: Pin all actions to verified current versions.

HIGH

H1 — Reviewer rubric catches 6 of 21 AI patterns (P2)

File: generate_posts.py, make_review_prompt() Impact: Significance inflation, copula avoidance, superficial -ing analyses, synonym cycling, inline-header bullet lists, rule-of-three overuse, title case headings all pass review. em_dash_count is logged but does not gate pass/fail. Fix: Full replacement of make_review_prompt() built from avoid-ai-writing 43-entry catalog and beautiful-prose style contract. Em dash count > 0 must gate fail.

H2 — P8 utcnow() deprecation

File: post_pins.py: fired_sentinel.write_text(_dt.datetime.utcnow().isoformat()) Impact: DeprecationWarning in Python 3.12, will raise in future versions. Fix: Global replace across all scripts: datetime.datetime.utcnow() → datetime.datetime.now(datetime.timezone.utc). Confirm import datetime (not from datetime import datetime) or adjust accordingly.

H3 — No S2→S3 automatic trigger (P7, queue item #3)

File: publish.yml, pin.yml Impact: pin.yml is workflow_dispatch-only. publish.yml writes .pending-slugs but nothing triggers pin.yml automatically. Pins never fire without manual intervention. This is structural. Fix: Add workflow_run trigger to pin.yml on Stage 2 — Publish Articles completion, or implement PENDING_DRAFTS.json model where S2 writes the file and S3 reads it.

H4 — Rewrite attempt 2 references Groq (blocked from GHA)

File: generate_posts.py, review_and_rewrite(), attempt-2 branch Code: Logs “REWRITING via Groq llama-3.3-70b” then bails if GROQ_API_KEY missing or CF-blocked. Impact: Attempt-2 rewrites always fail silently on GHA. Articles that fail review twice are held without a second rewrite attempt ever executing. Fix: Replace Groq rewrite path with Gemini direct API (already exists for attempt-1) or OpenRouter fallback.

H5 — P5 pin image URL semantics not enforced by named contract

File: post_pins.py ensure_cache_bust(), push_pins_to_sheets.py line ~73, generate_posts.py build_pin_image_url() Impact: The rule (IFTTT = bare URL, queue/sheets = ?v= URL) is correct but enforced by developer memory and a late-append patch in push_pins_to_sheets.py. Wrong URL type in wrong context is one refactor away from breaking. Fix: Two named functions: build_pin_image_url_for_queue(slug) → URL with ?v=YYYYMMDD, build_pin_image_url_for_ifttt(slug) → bare URL. All callers use only these. Delete the late-append patch in push_pins_to_sheets.py.

MEDIUM

M1 — No output contracts between stages (P1)

File: generate_posts.py main(), all stage functions Impact: review_and_rewrite() returns (content, False, flags) on failure; caller logs and continues. No exception propagates to GHA. A stage that fails silently produces a held article with no exit signal. Fix: Add validate_output(stage, result) to each stage. Raises named exception on failure. GHA step captures non-zero exit.

M2 — FB Queue dedup does not survive workspace recreation (P6)

File: push_pins_to_sheets.py Impact: sent/ is gitignored. On fresh GHA workspace, sent_dir.glob('*.json') is empty. Re-running push_pins_to_sheets.py appends duplicate FB Queue rows. Fix: Either un-gitignore sent/ or maintain a committed processed_slugs.json as the dedup record. The .fired sentinel pattern (already committed) is the correct model — apply same pattern here.

M3 — Rewrite prompt make_prompt() uses first-person plural

File: generate_posts.py, make_prompt() Code: Write in first person plural ("we tested", "we found", "we noticed"). Impact: Reviewer explicitly fails on first-person voice. Generator is instructed to use it. Reviewer then flags it. This is a contradiction baked into the prompts. Fix: Remove the first-person plural instruction from make_prompt(). Replace with second-person or third-person instruction consistent with reviewer rubric.

M4 — pipeline scripts reach into Brain via raw SQLite

File: push_pins_to_sheets.py, retire_from_products() Impact: Not a GHA risk (no concurrent access), but a coupling smell. If Brain schema changes, pipeline scripts break silently. Fix: Extract Brain writes to a dedicated brain_writer.py helper. Pipeline scripts call the helper, not raw SQLite.

M5 — P9 stale site reference location

File: push_pins_to_sheets.py Impact: Needs grep to confirm exact line. Logged as pending verification.

LOW

L1 — shutil import at top of post_pins.py is misplaced

File: post_pins.py line 1: import shutil before the module docstring Impact: Cosmetic. No runtime effect. Violates PEP 8 (imports after docstring).

L2 — check_url_live() re-imports urllib.request inside function

File: post_pins.py Impact: Redundant import. Module already imports urllib at top level.

NEW FINDINGS (not in problem map)

1. C3 — Fact-check routes OR model to Groq endpoint (breaking every roundup fact-check)
2. C4 — GHA action versions @v6/@v7 don’t exist
3. M3 — Generator instructs first-person plural; reviewer fails first-person. Direct contradiction.
4. H4 — Rewrite attempt-2 Groq path always fails on GHA (CF-blocked, was never replaced)

CORRECTED UNDERSTANDING OF P1 (step order)

The queue item #1 (“post_pins before push_pins”) describes an older publish.yml architecture. Current publish.yml does NOT call either script. Both are called from pin.yml in the correct order (post_pins → push_pins). The step order bug as described no longer exists. The real bug is H3: no automatic trigger from S2 to S3.

FIX PRIORITY ORDER (revised from handover)

Per P1-P10 from handover + new findings, adjusted for dependencies:

1. P8 — utcnow (10 min, zero risk)
2. C4 — GHA action versions (5 min, zero risk)
3. P9 — stale site reference (grep + one-line)
4. C3 — Fact-check endpoint routing (one function, high impact)
5. P3 — Fact-check truncation (one function)
6. M3 — First-person contradiction in make_prompt()
7. P5 — Pin image URL named functions
8. P6 — FB Queue dedup (processed_slugs.json or un-gitignore sent/)
9. P4 — GENERATION_RESULT.json + GHA exit gate
10. H3 — S2→S3 workflow_run trigger
11. H4 — Rewrite attempt-2 Groq replacement
12. P2 — Reviewer prompt rebuild (largest change)
13. P1 — Output contracts / validate_output()
14. P10 — Chewy validation script (new)