A mobile-readable visual brief distilled from the PKM archaeology package. It keeps the core lesson visible: the tools did useful work, but durable progress came from source-of-truth discipline, review closure, and bounded agent orchestration.
The source package inventories everything and deep-reads the documents that explain the pattern. This page shows the decision surface, not the private transcript bodies.
PKM Agent verification, transcript index counts, Wiki surfaces, postmortems, and key operating docs were represented in a single JSONL ledger. That ledger prevents the old failure mode: agents rediscovering the same terrain and smoothing uncertainty away.
Planning, PRD, implementation, GitHub/CI, multi-agent, Markdown, waste, success, and review workflows all have at least five ledger examples.
This public-by-link page excludes raw transcript bodies, tarballs, logs, private ledger rows, and local file bodies. It is for review on the go, not publication.
The report treats the failure as an operating-model issue: orchestration sprawl, context rehydration, and missing stop protocols.
The corpus reads like a founder learning the mechanics of software development by forcing real artifacts through agents, repos, PRs, and docs.
The evolution is not just product progress. It is a shift from "ask the model" to "make the repo, Wiki, issue, PR, dashboard, and ledger carry the truth."
Ambiguous ideas become acceptance behavior, non-goals, and operating assumptions that agents can execute.
Plans, source ledgers, handoffs, and decisions start carrying state across sessions instead of chat memory.
PRs, review comments, CI, and deploy previews create the real "done or not done" surface.
Claude Code and Codex work better when one implements and one reviews, not when both coordinate independently.
Actionable comments are fixed, replied to, resolved, and reflected in dashboard truth after merge.
The tools produced real work. The expensive part was making them repeatedly rebuild context, coordinate through overlapping substrates, and act when they should have stopped.
When the prompt carries the database, each new agent must reload the world. When three orchestration layers all think they own the repo, the system creates motion without a single scoreboard.
This is corpus-derived, not a permanent product claim. Recheck current docs before making public guidance.
The playbook is the practical output: how to brief agents, bound work, inspect outputs, and decide when to stop.
Objective: [One concrete outcome.] Repo / folder: [Exact path.] Current source of truth: [Plan, PRD, issue, PR, dashboard, or Wiki page.] Allowed scope: [Files/folders the agent may read and edit.] Validation: [Commands, browser checks, data samples, or review steps.] Stop condition: [When to stop and hand back.] If blocked: Return BLOCKED with exact missing input, evidence, and next action.
This Cloudflare copy is a visual review layer over the private package. It intentionally leaves out raw transcripts, private file bodies, and detailed ledger rows. Treat it as a portable founder review surface, then make promotion/sanitization decisions separately.