Spoiler-safe session recall for singleplayer games.
A personal tool for resuming singleplayer games after long breaks. Reads save files where it can, lets you annotate where it can't, and produces a re-immersion briefing that respects how far you've actually progressed — no plot reveals you didn't ask for.
Visual walkthrough of v1 against Sekiro: progression-DAG maps, Ship-Log-style journals, the structural spoiler filter in action across three player-state checkpoints, and a mocked synthesis briefing. The headline shot:
| Player view — masked | Authoring view — reveal |
|---|---|
 |
 |
14 observed, 5 frontier (? (type) stubs), 30 gated nodes invisible |
All 49 nodes — endings, late-game bosses, branching paths |
Same code, same data, one boolean. The model that synthesizes a briefing literally never receives the gated nodes — the filter happens at prompt-build time, not by asking the model nicely.
- Two knowledge stores. Game context (wiki-sourced MD files) ships with the app. Player state (save-derived observations) stays on your machine.
- Providers turn signals into normalized observations: save-file parsing (preferred), screenshot/OCR (later), manual notes.
- Synthesis is a single Anthropic API call — no agentic loops at recall time, because any tool that fetches external content is a spoiler vector.
- Progression is a DAG per game. Each piece of game context declares
which nodes gate it; the synthesis prompt is filtered by
gates ⊆ observedbefore the model ever sees it. The spoiler boundary is enforced structurally, not by prompt instructions. - CLI-first, on-demand. No daemon.
Full design in DESIGN.md. Per-file demo guide in demo/README.md.
v1 framework complete; one human-loop task remains. Save-format crypto,
synthesis layer, structural filter, CLI, and visualization all working. The
remaining v1 task is field-offset discovery in Sekiro .sl2 slot bodies —
which bytes correspond to sen / prayer beads / gourd seeds / etc. — done
by save-diff with the developer at the controls.
- v1 — Sekiro. Save parser, game-context files, synthesis, CLI end-to-end. Developer's own playthrough is the test set.
- v2 — Subnautica. Blind-playthrough test of the spoiler model under real stakes.
- v3+ — Generalize. Candidate: Cyberpunk 2077.
Personal tool. I play several singleplayer games in parallel and lose context during work/life breaks. Logging manually doesn't survive contact with my own discipline; reading the save file does.