Work log: Phantasmagoria — March 29, 2026

What shipped today

The session focused on two themes: hardening the test suite and implementing batch tech deduplication across the entire generation pipeline.

Test coverage expansion. The test count jumped from 385 to 451 — a net gain of 66 tests across 5 new or expanded test files. The first wave covered shared generator utilities (parse_json_response, check_pure_upside_options, validate_output) and config/rolling functions (_deep_merge_dicts, roll_reward_scope, roll_tone, roll_site_weirdness, roll_int_range). The second wave added tests for create_release.py — the release scaffolding orchestrator — with mocked AI calls for generate_release_yaml. A scout scan found 5 situation events missing the required situation_type field, which was causing the integration test to fail; those were fixed across both dark_discoveries and veiled_directive releases.

Tech deduplication pipeline. The long-standing PRODUCT.md known issue about narrow tech variety in batch generation has been addressed end-to-end. The work decomposed into three pieces: (1) an extract_used_techs() utility that scans event YAML for give_technology, add_research_option, and add_tech_progress effects; (2) tech repetition detection in validate_batch_diversity() that warns when the same specific tech appears in multiple events; and (3) a “technologies already awarded” section injected into all 4 generators’ AI prompts to steer the model toward unused alternatives. This is prompt-level guidance, not hard enforcement — the AI can still repeat techs if it has good reason — but the combination of validation warnings and prompt context should meaningfully improve variety.

Infrastructure cleanup. Dependencies were bumped (anthropic>=0.80.0, openai>=2.20.0), silent error swallowing in generate_site.py and generate_situation.py was fixed (bare except: continue now logs warnings), and .scout.yml was added for the /scout skill.

Completed

• #250 — Silent error swallowing in generate_site.py and generate_situation.py
• #251 — Fix failing integration test: dark_discoveries missing situation_type
• #252 — Add unit tests for generator modules (decomposed into #255-#257)
• #253 — Implement tech deduplication in batch generation context (decomposed into #258-#260)
• #254 — Update dependency versions (anthropic, openai)
• #255 — Test shared generator utilities
• #256 — Test generator config/rolling functions
• #257 — Test create_release.py
• #258 — Extract used techs from event data for batch dedup
• #259 — Add tech repetition detection to validate_batch_diversity
• #260 — Add used-techs context to batch generation prompts

Release progress

• v1.5: 18/18 closed (complete)
• v2: 5/5 closed (complete)

Carry-over

• Stage 3/4 split plan exists but hasn’t been executed (stellaris_mod_linter → validator rename, check_c3_to_ship split)
• Forge World (forgeworld.app) PRODUCT.md spec was drafted earlier today but no code yet

Risks

No new risks identified. The tech dedup is prompt-level only — if AI ignores the guidance, duplicates still pass through. The validator warns but doesn’t block.

Flags and watch-outs

• The data/releases/celestial_equinox/ directory has backup files from a Stage 1 reset (events_pre_stage1_reset_*, SUMMARY_pre_stage1_reset_*) that are untracked. These are intentional backups but may confuse git status.
• All 4 generators have copy-pasted build_existing_events_context and _build_tech_context functions. A future refactor to extract these to a shared module would reduce maintenance burden.

Next session

1. Execute the Stage 3/4 split plan — the plan file at .claude/plans/shimmering-skipping-oasis.md has the full spec. Rename linter → validator, split check_c3_to_ship into check_c3_to_c4 + check_c4_to_ship, update docs.
2. Run /scout to find new work — the queue is empty after clearing 11 issues today.
3. Consider the two backlog items — #179 (un-gate situation generation) and #222 (early/mid/late game timing) are the only open issues. Both are feature work that Paul may want to prioritize.