deadreckon
receipts · 12

Evidence, not a transcript

Each accepted run promotes to a review packet you read like a pull request: what changed, why, which call touched which file, and what it cost.

maps to as-built §14§25

read it like a pull request · an artifact, not a scrollback
You come back to evidence, not a chat log On the left, what an ordinary agent leaves behind: a patch and a chat transcript you have to scroll and take on faith. On the right, what deadreckon promotes every accepted run into: a review packet you read like a pull request. It contains RUN-NARRATIVE.md for what happened, RUN-DECISIONS.md for decisions and tradeoffs, RUN-AS-BUILT.md for the resulting shape, provenance.jsonl linking every changed file back to the turn, tool call, and model that produced it, and manifest.json recording what was built and from what. ORDINARY AGENT a patch + a chat transcript scroll it · take it on faith promote DEADRECKON · THE REVIEW PACKET RUN-NARRATIVE.md what happened, in prose RUN-DECISIONS.md decisions, tradeoffs, open questions RUN-AS-BUILT.md the subsystem shape afterward provenance.jsonl every file ↔ turn ↔ tool ↔ model manifest.json what was built, and from what
Every changed file links back to the turn, tool call, and model that produced it. Auditable on disk, not scrolled back in a chat window.

Most agentic tools leave you with a patch and a transcript, a long chat you scroll back through and largely take on faith. deadreckon promotes every accepted run into something you can actually review, the way you would review a pull request.

The review packet is plain files in the library:

  • RUN-NARRATIVE.md: what happened, in readable prose.
  • RUN-DECISIONS.md: the decision ledger of choices, deviations, tradeoffs, and open questions.
  • RUN-AS-BUILT.md: the shape of the subsystem after the run.
  • provenance.jsonl: every changed file linked back to the turn, tool call, and model that produced it.
  • manifest.json: what was built, and from what.

Because provenance ties each file to the exact call that touched it, you are never guessing why a line changed. The whole run is auditable on disk, and attach --view narrative shows the same story live while the run is still going.

source