Chronicles

Last updated: March 23, 2026

Chronicles are not session summaries. Summaries capture individual sessions: what happened in this one conversation. Chronicles synthesize across sessions into longer narratives: "what happened this week on the mesh coordination work." Builder's logs, generated from session data.

A two-agent pipeline produces them. The chronicle-analyzer runs hourly, reads session summaries, and produces a JSON analysis: which sessions belong to which journey, what the arc is, what's significant. The chronicle-scribe watches the analysis directory and generates prose. Markdown output, organized by month, landing in workspace/users/{you}/public/chronicles/.

About 120 chronicle entries from the bosun era, plus 234 older ones from before this setup existed. Eight months of work in there. What I did, why decisions got made, what went wrong, what worked.

I keep finding uses. Standup briefings. "What did I actually ship this week?" without reconstructing it from git logs and memory. Decision audit trails. "Why did we switch from channels to sync.Cond and then back?" Progress tracking across long-running projects. Handoff documentation for threads of work that span weeks.

The public chronicles feed the builder's log. Writing about what you're building while you're building it, except the agents do the writing.

Chronicles are part of the loop: sessions produce JSONL, the daemon summarizes them, summaries feed chronicles, chronicles become searchable knowledge in session history. Each layer adds more structure. Raw logs → structured summaries → narrative chronicles. The further up the chain, the more useful for humans. The further down, the more useful for agents.

Building the chronicle system was itself a chronicle-worthy story. The first implementation dumped everything into one expensive Sonnet call: read all session files, analyze, synthesize. $0.15 per chronicle. The fix: delegate the reading to Haiku, parallelize the writing across multiple Haiku scribes, use Sonnet only for orchestration. A single documentation fix unlocked the proper architecture, and costs dropped 12x.

The realization hit when I calculated token usage. The first chronicle consumed ~50K tokens, mostly reading session files into the expensive Sonnet model. The fix was obvious in hindsight: delegate file reading to a cheaper model. Not all work should be done by the most capable model.

Chronicle: Building the Chronicle System, Jan 2026