Accordion

Most agent context tooling asks you to trust a summary. Accordion asks you to look at the context first: one square per block, colored by type, with token counts, so you can see exactly what is eating your budget before anything gets folded. That alone is worth the install for anyone who has stared at a session and wondered where the tokens went. The fold model is the part I'd actually defend. A folded block collapses to a tagged digest carrying a six-character base36 code derived from the block's id — the same block yields the same code across sessions — and the agent can call unfold() to bring it back or recall() to return the full block once as a tool result without changing the view. Nothing is deleted. Folding is substitution, and it is reversible, which is more than you can say for a compaction pass. The Conductor is where it gets interesting, and where I'd stay cautious. You choose a strategy, from FIFO to thermocline, which uses attention from a 500M-parameter model to score each block's relevance and compress the least relevant first under a hard token budget. On the vendor's SlopCodeBench subset — DeepSeek V4 Pro, 100k-token budget, one run each — thermocline hits 5/6 checkpoints against naive compaction's 2/6. One run each is a thin proof; treat the 2.5x as directional, not settled. Pick Accordion when you run pi, your sessions run long enough to hit the budget, and you want a manual fold/unfold lever plus an optional automatic editor you can switch off. Pass if you're standardized on Claude Code or Codex today — those harnesses are roadmap items, not shipped ones — or if you want a hosted, team-shareable context layer. This is local, MIT, and single-player. One caveat worth knowing before you rely on it: only Accordion writes the fold tags. A digest you