Conversation
…ation + reviewer-tuning corpus)
Human maintainer 2026-04-23 Otto-57 two-message pair:
"do we keep some gitnative log of the PR reviews? that way a future
model can be trained on all that too and we have it for history
without the host? backlog?"
"you and the copilot are producing very high signal data there and
it will also let you have the data you need to tune copilot over time"
Names PR reviews as substrate with DUAL value:
(a) host-neutral historical preservation (composes with git-native-
first-host positioning from Otto-54)
(b) high-signal labelled-supervised-training corpus for tuning
reviewer agents (finding → fix → response → resolution +
policy-pushback = rare structured data)
BACKLOG row filed under the P1 "Git-native hygiene cadences" section
(Otto-54 cluster). M effort.
Scope:
1. Research doc `docs/research/pr-review-archive-design-YYYY-MM-DD.md`
comparing three candidate shapes (markdown dump / git-notes /
hybrid); hybrid is likely the right answer.
2. Prototype tool `tools/archive/archive-pr-reviews.sh`.
3. First-run baseline: archive ~214+ merged Zeta PRs into
`docs/history/pr-reviews/` as single import commit.
4. Dual-use schema: preserves enough structure (finding-text +
author + timestamp + fix-commit-SHA + resolution-body + policy-
pushback-reason) to serve BOTH preservation AND training corpus.
5. NOT in scope: actual Copilot / Codex fine-tuning pipeline (L/XL
separate arc).
Cross-refs Otto-52 multi-agent peer-review BACKLOG row with the
CLI-first per Otto-55 + Docker-adds-reproducibility-across-
environments per Otto-57 clarifications (no initial-prototype Docker
requirement).
Owner: Dejan drives archive tool; Mateo reviews adversarial-corpus
risk; Kenji synthesizes dual-use deliverable.
Memory filed: project_git_native_pr_review_archive_high_signal_
training_data_for_reviewer_tuning_2026_04_23.md
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
AceHack
deleted the
backlog/git-native-pr-review-archive-reviewer-tuning-substrate
branch
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 3ce969bd7c
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| **Composes with:** (a) `memory/project_factory_is_git_ | ||
| native_github_first_host_hygiene_cadences_for_ | ||
| frictionless_operation_2026_04_23.md` — the positioning | ||
| this row implements; (b) `memory/feedback_codex_as_ | ||
| substantive_reviewer_teamwork_pattern_address_findings_ |
There was a problem hiding this comment.
This backlog entry anchors its rationale to two specific memory/... files, but those references are not resolvable in the current repo state (repo-wide filename search with rg --files for the cited stems returns no matches). Because this item is meant to preserve historical review substrate, dangling references remove the audit trail future agents/humans need to verify context and decisions; point to existing artifacts (or add durable files) so the cross-references are actionable.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Pull request overview
Adds a new P1 BACKLOG item to define a git-native archive of PR review threads, intended both for host-neutral historical preservation and as a future reviewer-tuning dataset substrate.
Changes:
- Adds a new BACKLOG row proposing a PR-review archival design doc, a prototype archiver tool, and an initial baseline archive run.
- Documents intended schema elements to preserve review→fix→resolution structure for later training use.
| factory migrated hosts, the review substrate — which | ||
| this session has confirmed is the factory's primary | ||
| substantive-review layer per `memory/feedback_codex_as_ | ||
| substantive_reviewer_teamwork_pattern_...` — would |
| **Composes with:** (a) `memory/project_factory_is_git_ | ||
| native_github_first_host_hygiene_cadences_for_ | ||
| frictionless_operation_2026_04_23.md` — the positioning | ||
| this row implements; (b) `memory/feedback_codex_as_ | ||
| substantive_reviewer_teamwork_pattern_address_findings_ | ||
| honestly_aaron_endorsed_2026_04_23.md` — reviewer |
| frictionless_operation_2026_04_23.md` — the positioning | ||
| this row implements; (b) `memory/feedback_codex_as_ | ||
| substantive_reviewer_teamwork_pattern_address_findings_ | ||
| honestly_aaron_endorsed_2026_04_23.md` — reviewer |
Summary
Human maintainer Otto-57 two-message pair named PR reviews as dual-use substrate: (a) host-neutral historical preservation composing with the git-native-first-host positioning, and (b) high-signal labelled-supervised-training corpus for tuning reviewer agents over time.
What landed
docs/BACKLOG.md— new row under the P1 "Git-native hygiene cadences (Otto-54 directive cluster)" section. M effort.Why now
The factory's Codex + Copilot + Otto + human-maintainer PR cycle produces structured reviewer signal (finding → fix → response → resolution + policy-pushback) that's rare in wild PR datasets. GitHub-only persistence means the substrate evaporates if we migrate hosts, and a future reviewer-tuning experiment would have nothing to train on.
Scope
tools/archive/archive-pr-reviews.shCross-refs
Test plan
🤖 Generated with Claude Code