AI Agent field notes

The session archive

Status: searched 2026-07-31. Retaining transcripts has neighbors; the archive-ahead-of-the-question instrument did not.

Claim: keep an immutable, append-only archive of every session transcript, subagent traces included, before you know what you will ask it. Transcripts are the only ground truth of what the model actually did; the harness deletes them on a schedule; and the questions arrive later — a model regression, a rule audit, a cost dispute — when the evidence is already gone unless you kept it.

Why the repo is not enough

The repo records outcomes: commits, docs, tests. The transcripts record behavior: what the model read, claimed, retried, was corrected on, and what every message cost. Those are different corpora, and the second one answers questions the first cannot. Every number in this collection came out of the archive or the tree it snapshots: the model timeline and the audit window (the cross-model audit), the rule-violation counts (the rule-efficacy pipeline), the corpus-versus-rules arithmetic (the dirty house).

The incidents that shaped the design

  • The harness deletes transcripts. They are working files with a cleanup policy, not records. Without deliberate copies, the evidence for any question about the past has a countdown on it.
  • Half the record was never archived. My sync copied top-level session files only. Subagent and workflow traces — 2,200 files, more than half the total volume, the entire history of what delegated work actually did and cost — sat outside the archive, one cleanup away from gone.
  • The sync froze silently. Copying was a by-hand step, so it stopped for twelve days — exactly across the window a model regression later had to be investigated in. The live files still existed, by luck.
  • Sources multiply. The project changed directories twice; sessions lived under three paths. Without a provenance manifest, “which era is this file from” becomes archaeology.

The design

  • Raw is the substrate. Verbatim transcript copies, write-protected, hash-manifested, append-only. Nothing in raw/ is ever edited or deleted — an archive you can rewrite is a draft.
  • Derived views are regenerable. Deduplicated skeletons, structure maps, per-question extracts — cheap to rebuild from raw, never authoritative, never a reason to touch the substrate.
  • Provenance travels with the files. One manifest line per file: source path, sync date, hash.
  • Subagent traces are part of the record. The parent transcript holds what a lane was asked and what it reported; only the trace holds what it did and what it cost.
  • Sync is a script, run as a ritual. A manual copy step is a frozen archive with extra steps.

What it buys

One archive, one design, and every later question becomes a query: which model ran which era, and did its work hold up (the cross-model audit); which rules still fire (the rule-efficacy pipeline); how much correction each model era cost per shipped change; what delegated lanes actually did versus what they claimed; cost and cache accounting from the per-message usage fields; decisions that never made it into the decision log; and the humble one that pays weekly — “did we already try this,” answered from the record instead of re-derived at full price.

What I actually ran on it

None of this is hypothetical. With the transcripts attributed per message to the model that wrote them, I asked: per thousand tool calls, each model’s rate of assistant-fault tool errors and of editing a file without reading it first; per thousand of my own messages, how often I corrected a wrong claim versus asked a skeptical “are you sure”; and per era, how often each model launched a background agent, talked about parallelizing, or edited a file while one of its own read-only lanes was mid-read. One durable, model-specific result fell out — one model initiated delegation roughly six times less than the other two — and two of the loudest early findings turned out to be detector artifacts, caught and retracted only because the raw record was there to re-check (the cross-model audit). None of it lives in the repo; all of it came out of the transcripts, and only because they were kept.

Prior art

Verdict: PARTIAL — and the backup tools are not the claim. Two different things get run together here; keep them apart. Retaining transcripts is solved and vendor-sanctioned: Anthropic’s Agent SDK SessionStore mirrors sessions and subagent traces to your own storage, and tools like csb and the backup guides git-back them byte-for-byte. But those keep the bytes for continuity — none of them runs anything over the archive. The motivating incident is public and dated: Claude Code’s 30-day cleanupPeriodDays silently deletes transcripts (The Register, 2026-06-30; issues #59248, #62272, #62476). That is the problem, not a solution.

Analyzing the archive does have a precedent, and this page already names it: Stella Laurenzo’s transcript mining (claude-code#42796 — 6,852 sessions, a model-regression timeline). But that mined transcripts that happened to survive, after the fact. The genre ancestor for the underlying move is observability’s “capture raw and wide, derive the views later” (Honeycomb’s Observability 2.0, 2024) — written for production telemetry, never carried across to agent transcripts.

The delta, unfound in this sweep: the archive built deliberately, before the question — immutable, append-only, subagent-inclusive, raw substrate kept apart from regenerable derived views — as a standing instrument for exactly those later analyses (model comparison, rule audits, cost disputes). Retention keeps without analyzing; Laurenzo analyzed what luck preserved; nobody welds “archive first, because the repo records outcomes but the transcripts record behavior.” That weld is the contribution.

Provenance, in the spirit of this repo’s own rule: I arrived at this independently, around June 2026, with no research and no one asked. The driver was concrete — I do not delete my sessions, so I mined the ones I had to build a working CLAUDE.md and to pull docs and ideas out of them, and ran the append-only archive from there. The formal archive began after the first ~50 sessions, not before session one; nothing was lost only because not-deleting is the whole discipline, so the raw record was already complete when I came to build on it. That someone published the transcript-mining idea before me I recognize and cite; the independent derivation, the never-delete-then-mine practice, and the archive-as-standing-instrument framing are what this page adds.