HappyPet — Session Handoff (2026-07-20)

Ground truth as of commit 8d46e73 on main. Working tree clean except two long-standing untracked files (CLAUDE.md, GENERATION_RESULT.json) — leave them, they are pre-existing and gitignored-in-spirit. Test suite: ./.venv/Scripts/python.exe -m pytest test_pipeline.py -q → 125 passed, 0 failed (the old 2 Windows-cp1252 failures were fixed this session).

Full review + phased plan (the “F1–F17” item IDs referenced below) lives in docs/superpowers/specs/2026-07-20-happypet-autonomy-review-and-plan.md. Auto-memory index: ~/.claude/projects/C--Users-derek-MAEVE-HappyPet/memory/happypet-autonomy-plan.md.

1. Mission

HappyPet is a Jekyll affiliate blog (happypetproductreviews.com) with a Python + GitHub Actions content pipeline (generate → publish → deploy → pin). The goal is to run the content pipeline autonomously. This session did a multi-model code review, shipped the autonomy-readiness fixes, and then chased the real blocker: the generator’s output can’t clear a deliberately strict anti-AI reviewer, so every article gets held instead of published. The pipeline plumbing works end-to-end; the open problem is generation quality vs. the review bar.

2. Current State

Working and verified (all merged to main this session, 13 PRs):

• Pipeline is hardened for unattended runs: crash-safe atomic products.json writes (json_io.py), push-retry on all main-writers, orphan-draft prevention, drained-queue signal, Chewy 429 backoff, workflow guards. A CI test-gate (.github/workflows/test.yml) runs the suite on every PR — green on Ubuntu.
• Publish blocker F1 fixed + proven on real data: the gate accepts amazon.com/dp/...?tag= links (was amzn.to-only); the current 23-entry queue is now publishable.
• Generator is Claude Sonnet (anthropic/claude-sonnet-4.5) via OpenRouter (Gemini demoted to fallback). Slug verified live (run used it, no fallback).
• Prompts de-sabotaged: generation prompt no longer models em dashes / first-person; split into a system prompt (GENERATOR_SYSTEM_PROMPT, rules in <writing_rules>) + XML user task; rewrite prompt re-asserts hard rules and no longer tells the model to “add a human moment” (which caused first-person + fabrication). Temp 0.7→0.5.

The blocker — two supervised live runs HELD (did not publish):

• Preflight secrets check: PASS (Gemini, OpenRouter paid tier, Sheets/Gmail/IFTTT/Impact/Facebook all present; PA-API blank as expected).
• Run 1 (Gemini, before swap): human_voice=2, em_dashes=12, first-person, fabricated quotes → held.
• Run 2 (Claude Sonnet, before the prompt PRs #63–#65): human_voice=3, warmth=3, em_dashes=8→12; the rewrites made it worse (fabricated stats, “My Lab Buster” first-person). That rewrite-degradation is fixed in #64, but no run has been done since the prompt PRs.

Exact next action: Implement the deterministic em-dash strip (F3-adjacent). No LLM — Gemini or Claude — emits exactly 0 em dashes, and the reviewer rule is “any em dash = FAIL.” Strip — (U+2014 → hyphen/comma) from the article body right before the review gate, mirroring the existing scrub_banned_phrases. This is a ~10-line, fully-testable change and is the single fix most likely to flip “held” → “passed.” Then do one supervised confirmation run (gh workflow run generate.yml -f force_cap=1, ~2¢) to see where human_voice/warmth land — that’s the only remaining unknown, and it drives the Open Question about the ≥4 bar.

3. Decisions Made (and Why)

• Decision: Deliver a plan + fixes, keep the generate.yml cron HELD. Alternatives: flip live now. Reason: Director wants a supervised, verified path; going autonomous is his explicit call. Reversibility: load-bearing — do NOT enable the cron without an explicit “go.”
• Decision: Generator = Claude Sonnet via OpenRouter. Alternatives considered & rejected: the Max subscription (interactive-only, cannot authenticate a CI pipeline); “free” Claude (OpenRouter’s free models are open-weight, not Claude); a direct Anthropic API key (violates the standing “Claude only via OpenRouter, no direct key” rule). Reason: reuses the reviewer’s proven plumbing, stays in-convention, ~pennies/article. Reversibility: model is a one-line constant (OR_GEN_MODEL_CLAUDE).
• Decision: Em dashes must be handled by deterministic code, not prompting. Reason: empirically, both Gemini (12) and Claude (8–12) produce them despite explicit prohibition. Reversibility: easy; it’s a scrub step.
• Decision: Rules moved to a system message + XML tags (system/user split). Reason: Anthropic guidance — Claude obeys system-role, XML-delimited rules far better than a flat user blob. Reversibility: contained to make_prompt/call_generator.
• Decision: Self-merge each green, CI-passing PR to drive to done rather than queueing them for review. Reason: Director said “I need a working project, not a project car we keep piecing together.” Reversibility: N/A (process).

4. Architecture & Key Files

• generate_posts.py — the generator. Key symbols this session: GENERATOR_SYSTEM_PROMPT (system-role persona + rules), make_prompt() (returns XML <task>/<featured_product>/<structure> user content), call_generator(prompt, api_key, system="") (Tier 1 Claude Sonnet via OpenRouter with system message + temp 0.5, Tier 2 Gemini fallback), make_rewrite_prompt() (now re-asserts hard rules), evaluate_scorecard() (deterministic pass/fail — thresholds + em-dash-in-body check; this is where the em-dash strip should feed from / near), validate_output() (publish gate; accepts amzn.to OR amazon.com/dp; AFFILIATE_LINK_RE), validate_product() (rejects NEEDS_ASIN/NEEDS_IMAGE), persist_generated_article() (pin-queue-then-draft, orphan-safe). The api_key/groq_key params threaded through are dead (vestigial) — safe to remove later (F13).
• json_io.py — NEW. atomic_write_json() + read_json()/CorruptJSONError. Used by all 4 products.json writers.
• test_pipeline.py — 125 tests. New classes this session: TestProductValidation, TestScorecardEvaluation, TestProductsJsonAffiliateContract (loads real products.json), TestJsonIO, TestArticlePersistence, TestChewyApiRetry, TestGeneratorModel, TestGenerationPromptHygiene, TestGeneratorSystemPrompt, TestRewritePromptGuardrails.
• .github/workflows/ — test.yml (NEW, PR gate; installs pytest), plus generate/publish/pin/deploy hardened (push-retry, drained-queue, deploy single-fire). generate.yml cron still commented out (HELD).
• docs/superpowers/specs/2026-07-20-happypet-autonomy-review-and-plan.md — the full F1–F17 review/plan.
• chewy_lookup.py — _impact_get() now retries 429/503.

5. Gotchas & Hard-Won Knowledge

• No LLM emits zero em dashes. Prompting reduces but never eliminates them. The “any em dash = FAIL” rule therefore requires a deterministic strip. This is the crux of why runs keep holding.
• The prompt was sabotaging itself: it contained 9 em dashes while forbidding them, and its “good opening” examples were all first-person while forbidding first-person. Fixed (#63/#65), but the lesson: check that a prompt doesn’t model the behavior it bans.
• The rewrite loop degraded good drafts. make_rewrite_prompt told the model to “add a concrete human moment” for warmth → it invented first-person anecdotes + fake stats. Fixed (#64). If rewrites regress again, re-check that prompt.
• human_voice ≥ 4 may be unwinnable by automated generation. A 4 means “genuinely non-AI-sounding”; Claude Sonnet reliably produces 3 (“competent but generic”). Even relaxing to ≥3 wouldn’t pass a 2. This is an unresolved product decision (see Open Questions).
• A hand-written Claude prototype is NOT representative of Claude-Sonnet-via-API. An earlier free prototype (written by Opus in-session, 0 em dashes, human_voice≈4) was too optimistic; the real API run scored lower and had em dashes. Trust live runs, not in-session prototypes.
• The Max subscription cannot power the CI pipeline (interactive auth only). External API key required.
• CI gate needs pytest explicitly — requirements.txt is runtime-only; test.yml installs pytest separately.
• Model slug anthropic/claude-sonnet-4.5 is live-verified on this OpenRouter account (a real run used it). Bump the constant for a newer Sonnet if desired.

6. Conventions In Play

• Branch + PR for everything, never direct to main. Branch naming claude/happypet-recovery-N-<slug> — next N is 35. Data-only changes use refill/<date>.
• TDD for code: write the failing test, watch it fail, implement, watch it pass. Prompt changes are guarded by content-assertion tests (e.g. “prompt contains ≤1 em dash”).
• Self-merge green PRs this session (gh pr merge --merge --delete-branch) after the suite passes; CI gate also runs on the PR.
• Conventional commits (feat(...), fix(...), ci(...)). Commit messages end with Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>; PR bodies end with the Claude Code footer.
• Model constraints: reviewer = Claude Haiku 4.5 via OpenRouter; generator = Claude Sonnet 4.5 via OpenRouter; no direct Anthropic key.

7. Open Questions

1. After the em-dash strip + a confirmation run, does human_voice clear the bar? If it lands at 3, the Director must choose: (a) lower human_voice/warmth pass threshold 4→3 (keep fabrication/first-person as hard holds), (b) try Opus as generator (more cost, uncertain), or (c) human-in-the-loop on held articles. Do not lower the quality bar without an explicit yes — it changes the site’s content standard.
2. When to spend on confirmation runs? Each is ~2¢ and, if it passes, publishes one LIVE article + Pinterest pin. Director controls timing.
3. Go-live “go” for the generate.yml cron — unchanged, still his call, after a clean supervised run.
4. Mark CI — Tests / pytest a required status check in main branch protection (UI action, can’t be done from code).

8. Do Not Touch

• generate.yml’s commented-out schedule: block — the load-bearing hold. No cron without an explicit “go.”
• Reviewer pass thresholds (REVIEW_SCORE_MINIMUMS, human_voice≥4 etc.) — do not relax without the Director’s decision (Open Q1).
• PA-API — parked; refill.yml auto-trigger stays HELD until keys land. Amazon scrape path in refill_products.py is deliberately dead.
• Chewy COVERAGE_AUTO_ACCEPT/SCORE_* thresholds and _first_brand_token — empirically settled; don’t retune casually (F10 tokenization tidy is deferred and needs real-case regression).
• The direct-Anthropic-key rule — stay on OpenRouter for Claude.

Deferred backlog (optional, non-blocking): F9 auth-failure alerting, F10 Chewy tokenization consistency, F13 remove dead groq_key/api_key params, F15 IFTTT-key shift-left validation. (F12 already covered by pin.yml’s existing failure-alert step.)

9. Resume Command

Read HANDOFF.md. Then implement the deterministic em-dash strip: strip — (U+2014) from the article body immediately before the review gate in generate_posts.py (TDD — add a test that content with em dashes comes back with zero, mirroring scrub_banned_phrases). Run ./.venv/Scripts/python.exe -m pytest test_pipeline.py -q (expect all green), then branch claude/happypet-recovery-35-em-dash-strip, PR, and merge once CI is green. Do not enable the generate.yml cron, lower the reviewer thresholds, or trigger a live generate.yml run without asking the Director first (a passing run publishes a live article + Pinterest pin). After merging, ask the Director whether to spend ~2¢ on a confirmation run.