A Continuity Layer for LLM Collaboration Storytime dropped on Hacker News this week as an attempt to solve one of the most annoying problems with AI coding assistants: context loss. The plugin, installable via `claude install-plugin ~/workspace/storytime`, acts as a continuity system that maintains session state across compactions, naps, shifts, and full sessions. Instead of starting each conversation cold, Storytime stages "remembrance" documents before `/compact` commands—essentially wakeup docs that get loaded as the first action post-compaction. It's a clever workaround for the context window ceiling that's been haunting LLM harness users since these tools went mainstream.

Structured Persona Workflows The core workflow runs through six phases: Survey, Assemble, Icebreaker, Breakout, Converge, and Review. Each phase collapses when empty, so not every run uses every gear. When you invoke `/storytime` with a problem like "our public API has no rate limiting," the system surveys your codebase, finds the Express middleware chain lacks throttling, then assembles a team of domain-expert personas to investigate and produce a plan. These aren't generic skill triggers—they're non-human codenamed lenses (anchor, lattice, kestrel) that function as persistent experts with specific viewpoints on your actual code.

The Consolidation Loop The real magic is the consolidation loop: Context → consolidation events → document structure → continuity. Storytime operates at six scales—phase, commit, nap, shift, session, and compact—with `/storytime-remember` staging remembrance before each compaction event. Decisions get append-only numbering per topic (like RATE-001), pinned to commits and stored in per-topic threads rather than scattered across files. Cross-topic references use `Callout-> / Callout<-` sigil lines—forward is authoritative, reverse is lint-cached. This creates an auditable trail of reasoning that survives context window resets.

Personas as Domain Lenses The default persona cohort runs OWNER × OPERATOR × CRITIC² with non-human codenames. One driver leads each leg while supporters stay silent unless their input would be both useful and non-distortive—preventing the classic "too many cooks" problem where every agent chimes in uselessly. The `/storytime-cohort` skill lets you hire, fire, bench, promote, or evolve personas mid-session. Want to bring in a performance specialist for a specific investigation? That's a personnel call Storytime takes seriously. @role operates as a lens directive rather than a simple skill trigger—these personas have genuine perspective on your codebase.

Decisions, Threads, and Dreams The Decisons & Callouts system creates numbered, append-only records (RATE-001 style) pinned to commits. Sessions become chapters in Episodes & Threads, with `_thread.md` serving as the continuity ledger storing decisions, consolidation logs, and open questions. You can park mid-session and resume later with full context intact. Then there's "Dreams"—optional, ancillary, disableable per-commit side artifacts capturing hunches, noticed-but-not-said observations, and daydream design material. Not on the critical path, just accumulated hunch value over time for when you need to remember why you thought a refactor was a good idea at 2 AM.

Skills Inventory Storytime ships with 19 skills across three categories: Core Workflow (`/storytime`, `/storytime-survey`, `/storytime-breakout`, `/storytime-converge`, `/storytime-buildout`, `/storytime-retro`), Team & QA (`/storytime-cohort`, `/storytime-qa`, `/storytime-echo`, `/storytime-pr-qa`), and Continuity & Control (`/storytime-remember`, `/storytime-lint`, `/storytime-bootstrap`, `/storytime-consolidate`, `/storytime-absorb`, `/storytime-export`, `/storytime-status`, `/storytime-undo`). Use cases span feature design, architecture decisions, bug investigation, refactoring, onboarding, and post-implementation retrospectives.

Key Takeaways

Storytime solves context window loss through staged remembrance documents before each compaction cycle. Structured persona teams with non-human codenames bring domain expertise without hallucination-prone genericism. Append-only numbered decisions (RATE-001 style) create an auditable trail of architectural reasoning. The Dreams feature captures hunches and side observations off the critical path—accumulated insight for later retrieval.

The Bottom Line This is what LLM harness tooling needed—a proper continuity layer that treats context as something worth preserving rather than something you lose to entropy. The consolidation loop concept alone could influence how we think about AI-assisted development long-term, turning one-off coding sessions into persistent architectural narratives. Worth watching closely as it matures.