Skip to content

process: the orchestrator holds the review bound, the packet is finished before the verdict, progress is evidence - #22

Open
paveq wants to merge 1 commit into
mainfrom
process/retro-review-bound-packet-freeze-and-held-facts
Open

paveq wants to merge 1 commit into
mainfrom
process/retro-review-bound-packet-freeze-and-held-facts

Conversation

@paveq

@paveq paveq commented Sep 16, 2026

Copy link
Copy Markdown
Contributor

Why

A delivery-session retrospective (2026-09-16) found four habits that the skills state only where the wrong reader sees them, or not at all. None is a new process rule; each line applies a rule PROCESS.md already holds.

What changes

  • rdd-deliver step 4 — the loop read "repeat until TODO" with no round bound. The bound lived only in rdd-cold-review and PROCESS.md §Planning and readiness: the reviewer's text, not the orchestrator's. The session ran a third cold-review round on its own call. The loop now carries the bound and the stop: after a second FAIL, hand what is known to the human.
  • rdd-plan step 4 — a packet stated a 5-minute stale bound read from a constant's definition; the agent path applied 10 minutes and the 5-minute constant was unused. A stated value is now read at the call site that applies it, not at its definition (the same rule rdd-audit gives for quoted expressions).
  • rdd-plan step 7 and rdd-build step 9 — a bookkeeping push (a code citation, a built-at note) after the cold-review verdict moved the packet fingerprint and voided the review, twice. The packet is finished before the verdict; which edits move the fingerprint is the store's decision, not the editor's intent. Build progress is recorded as evidence, never as an edit to the reviewed packet.
  • rdd-start session discipline — the human had to ask "what is not yet in the store?" to learn two items were being held back. A pass report names what the session holds that the store does not.

Not changed

PROCESS.md, rdd-cold-review (its round bound was already right), file-state/.

🤖 Generated with Claude Code

…hed before the verdict, progress is evidence, and a report names what the store does not yet hold

A delivery-session retrospective (2026-09-16) found four habits the
skills stated only where the wrong reader would see them, or not at all:

- rdd-deliver step 4 read "repeat until TODO" with no round bound, while
  the bound lived only in rdd-cold-review and PROCESS.md — the reviewer's
  text, not the orchestrator's. The session ran a third round on its own
  call. The loop now carries the bound and the stop.
- rdd-plan step 4 asked that claims be verified by reading the code, but a
  packet still stated a stale-bound constant from its definition while the
  agent path applied a different value. A stated value is read at the
  call site that applies it.
- rdd-plan step 7 and rdd-build step 9: a bookkeeping push after the
  cold-review verdict moved the packet fingerprint twice and voided the
  review. The packet is finished before the verdict, and build progress
  is recorded as evidence, never as an edit to the reviewed packet.
- rdd-start discipline: the human had to ask "what is not yet in the
  store?" to learn that two items were being held back. A pass report
  now names what the session holds that the store does not.

No PROCESS.md rule changes; each line applies a rule that already exists.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant