FE-883: Cook artifact lifecycle — git-merge slice composition by kostandinang · Pull Request #230 · hashintel/brunch

kostandinang · 2026-06-18T09:52:00Z

Stack Context

Part of the FE-883 cook artifact-lifecycle work. Stacked on #227 (FE-881, cook agent skills); GC follow-up is #231.

What changed

Wires the already-built run-artifact.ts git-merge composer (previously dead code) into the live brownfield cook path, replacing the file-copy union that silently resolved overlapping edits as last-slice-wins.

Composer is correct under dependency-seeding — slice commits record their dependency commits as parents, so a dep-seeded file folds as an edit, not an add/add false conflict.
Brownfield promotion calls harvestCookRun — a fold conflict is now a fatal run outcome, not a buried event field. The user's checkout is never touched.
Brownfield verify-epic composes via the same fold in a detached worktree, so the tree verified equals the tree shipped; a real conflict fails the epic instead of verifying a partial tree. Greenfield keeps the file-copy union (no common ancestor for a 3-way merge).

Effect

Two slices editing different hunks of the same pre-existing file now both survive; a true conflict halts fail-closed instead of silently shipping last-writer-wins.

Verification

Full orchestrator suite green; check + build pass. (One client build-boundary test fails only in the dev worktree due to build-plugin resolution — it passes on base and in CI.)

Follow-ups

FE-883: Cook run worktree/branch GC #231 (stacked) — worktree/branch GC; retires the now-dead promoteBrownfieldRun.
Multi-slice end-to-end runCook test is the remaining coverage gap (merge logic is unit-tested; the 1-slice brownfield-smoke exercises the engine path). Tracked in memory/CARDS.md.

kostandinang · 2026-06-18T09:52:25Z

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

FE-883: Cook artifact lifecycle — git-merge slice composition #230 👈 (View in Graphite)
FE-881: Cook agent loads the target repo's sandbox-scoped skills #227
FE-883: Cook operational hardening — runtime stability, agent skills, git-merge composition, run GC #224 : 1 other dependent PR (#232 )
FE-864: Brownfield feature delivery from spec — detect, classify, probe, oracle, promote, serve #212 : 2 other dependent PRs (#213 , #223 )
FE-841: Pi agent foundations — in-process SDK, toolchain profiles #194 : 2 other dependent PRs (#198 , #211 )
main

This stack of pull requests is managed by Graphite. Learn more about stacking.

cursor · 2026-06-18T10:38:04Z

PR Summary

Medium Risk
Changes brownfield artifact composition and epic verification semantics; mitigated by fail-closed conflicts, plumbing-only git operations (I135-K), and expanded unit tests, but multi-slice end-to-end engine coverage is still a noted gap.

Overview
Brownfield cook now composes slice output with a git merge-tree fold instead of copying trees with last-slice-wins. The existing run-artifact.ts composer is hooked into promotion (harvestCookRun in cook-cli) and verify-epic (materializeEpicVerifyTree in net-compiler), so the tree verified matches the tree shipped. Greenfield still uses the file-copy union in mergeSlicesIntoEpicSandbox.

Slice commits now record dependency commits as parents (via commit-tree), fixing false add/add conflicts when a dependent slice extends dep-seeded files. commitSliceWorktree is idempotent so promotion reuses commits from an earlier verify pass. Real fold conflicts fail closed (CLI exit / epic passed:false) instead of being buried in merge events.

CLI finish output moves to cookFinishLines; presenter phase detection accepts ✓ cook → promoted. SPEC I124-K and planning docs describe the plan.mode composer fork. promoteBrownfieldRun is no longer used on the live brownfield path (removal deferred per CARDS 1d).

^{Reviewed by Cursor Bugbot for commit 4cc6d23. Bugbot is set up for automated code reviews on this repo. Configure here.}

cursor

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 6b3aaba. Configure here.}

The git-merge run-artifact composer (871ef08) was left unwired pending "a live-run check of the dependency-seed interaction". That check, encoded as a test, fails: a dependent slice that extends a dep-seeded file (slice B has A's output copied in by seedSliceSandboxFromDeps, then edits it) raises a spurious merge-tree conflict, because every slice branch is rooted at the run base with no inter-slice ancestry — so the dep-seeded file looks like an add/add. Fix: commit each slice (in dependency order, which harvestCookRun already uses) recording its completed-dependency commits as additional parents. The fold's merge-base then becomes the dependency, so a dep-seeded file reads as an edit the dependent slice evolved — clean fold, dependent's version wins — while genuine sibling conflicts still surface fail-closed. commitSliceWorktree switches from `git commit` to commit-tree + update-ref to carry the parents. Also corrects the happy-path harvest test, which declared B as depending on A but never seeded A's output into B — impossible under real dep-seeding; made them the disjoint siblings the test name claims. Composer correctness only; wiring it into the live promotion/verify path (replacing the file-copy union) is the remaining Slice 1 work. 🍳 Built with brunch Co-Authored-By: Opus 4.8 <noreply@anthropic.com>

Re-record the scope against the true FE-864 seam: the git-merge composer (run-artifact.ts, 871ef08) already exists, so FE-883 is wiring it in, not building it. CARDS.md tracks Slice 1 sub-steps (1a done; 1b wire, 1c verify-epic consistency design fork, 1d delete file-copy path) + Slice 2 GC. PLAN.md gets the frontier definition + a Parallel/Low-conflict sequencing entry. 🍳 Built with brunch Co-Authored-By: Opus 4.8 <noreply@anthropic.com>

The headline correctness win over the file-copy union, pinned: two independent slices editing different lines of the same pre-existing file both survive the merge-tree fold, where a whole-file last-slice-wins copy would drop one. 🍳 Built with brunch Co-Authored-By: Opus 4.8 <noreply@anthropic.com>

…mechanism) 1c decision: verify-epic must run tests against the same merged tree promotion ships, not a file-copy union that diverges on same-file edits. This builds the load-bearing mechanism the unknown was flagged on (checking a merge-tree fold out as a worktree): - Factor foldToCommit out of foldSliceBranches (fold N slice commits onto a base, fail-closed, write no refs) so callers can either publish (promotion) or check out (verify). - materializeFoldedWorktree: fold onto a base + `git worktree add --detach` the result at destDir, rework-safe (removes any prior worktree first). Caller relinks shareable gitignored entries (node_modules) since the fold tree carries only tracked content. Test proves different-hunk edits to one file both land on disk in the checked-out verify worktree (the file-copy union would drop one) and that re-materializing is idempotent. Mechanism only; the net-compiler verify-epic + cook-cli promotion wiring is the remaining 1b/1c integration. 🍳 Built with brunch Co-Authored-By: Opus 4.8 <noreply@anthropic.com>

… plan 🍳 Built with brunch Co-Authored-By: Opus 4.8 <noreply@anthropic.com>

Brownfield cook now composes by the merge-tree fold end-to-end, replacing the file-copy union that silently last-slice-wins on same-file edits: - net-compiler verify-epic: brownfield builds __epic__/<epicId>/ via materializeEpicVerifyTree (commit slices in dep order → fold → checkout a detached worktree → relink node_modules), so verify-epic tests the SAME tree promotion ships (1c). A real fold conflict fails the epic (emits a passed:false epic-verified report → fail sibling) instead of verifying a partial tree. Greenfield keeps the file-copy union (no common ancestor; spike decision). - cook-cli promotion: brownfield calls harvestCookRun instead of promotionSourceDir + promoteBrownfieldRun; fold conflicts are a fatal run outcome (recordCookExitStatus(false)), not a buried event field. I135-K holds (all plumbing; the user's checkout is never touched). Shared plumbing: commitSlicesInDependencyOrder factored out of harvestCookRun and reused by verify; commitSliceWorktree made idempotent so the promotion pass reuses the commits a prior verify pass made (else add -A finds nothing and drops every slice). Exported resolveEpicSandboxDir + SHAREABLE_TOP_LEVEL_ENTRIES. Tests: verify→promote reuse, different-hunk 3-way, fold materialization on disk. Full orchestrator suite green (672). Remaining: 1d delete now-unused promoteBrownfieldRun + the stale epic-sandbox-merge.ts TODO; SPEC I124-K amendment. 🍳 Built with brunch Co-Authored-By: Opus 4.8 <noreply@anthropic.com>

- epic-sandbox-merge.ts: replace the now-resolved cook-artifact-lifecycle TODO with a note that brownfield folds (run-artifact.ts, wired) while mergeSlicesIntoEpicSandbox is the greenfield composer by design. - SPEC I124-K: amended to record the plan.mode fork — brownfield git merge-tree fold (verify == ship, fail-closed) vs greenfield file-copy union. - CARDS.md: 1b/1c done; 1d (retire dead promoteBrownfieldRun, blocked on the landCookBranch fixture rewrite) + a multi-slice end-to-end engine test remain. 🍳 Built with brunch Co-Authored-By: Opus 4.8 <noreply@anthropic.com>

When a run promotes, replace the ad-hoc `✓ promoted → …` one-liner with a clean finish block: the navigable directory, the branch + short commit, and copy-paste `next` commands tailored to how the run landed (greenfield cd-in, brownfield merge-when-ready, --land outcomes). Centralized as a pure, golden-tested `cookFinishLines` builder alongside the other cook-report strings; halt/conflict/error paths keep their existing inline lines. The brigade phase tracker keyed on the literal `✓ promoted` prefix to light plate; broaden it to match the `promoted` token after the ✓ so the new `✓ cook → promoted` phrasing still advances the footer.

kostandinang mentioned this pull request Jun 18, 2026

FE-881: Cook agent loads the target repo's sandbox-scoped skills #227

Closed

kostandinang changed the title ~~FE-883: fix run-artifact fold false-conflict on dependency-seeded files~~ FE-883: Cook artifact lifecycle — git-merge slice composition Jun 18, 2026

kostandinang mentioned this pull request Jun 18, 2026

FE-883: Cook run worktree/branch GC #231

Closed

kostandinang marked this pull request as ready for review June 18, 2026 10:37

cursor Bot reviewed Jun 18, 2026

View reviewed changes

Comment thread src/orchestrator/src/run-artifact.ts

kostandinang force-pushed the ka/fe-883-orchestrator-improvements branch from b902c7a to 3890b32 Compare June 18, 2026 10:42

kostandinang changed the base branch from ka/fe-864-pi-timeout-600s to ka/fe-881-cook-agent-skills June 18, 2026 10:42

kostandinang force-pushed the ka/fe-881-cook-agent-skills branch from fbd12b2 to c830824 Compare June 18, 2026 11:43

kostandinang force-pushed the ka/fe-883-orchestrator-improvements branch from 3890b32 to 6b3aaba Compare June 18, 2026 11:43

cursor Bot reviewed Jun 18, 2026

View reviewed changes

Comment thread src/orchestrator/src/cook-cli.ts

kostandinang force-pushed the ka/fe-881-cook-agent-skills branch from c830824 to a811396 Compare June 18, 2026 11:46

kostandinang force-pushed the ka/fe-883-orchestrator-improvements branch from 6b3aaba to 4df8c49 Compare June 18, 2026 11:46

kostandinang requested a review from lunelson June 18, 2026 12:59

kostandinang self-assigned this Jun 18, 2026

kostandinang force-pushed the ka/fe-883-orchestrator-improvements branch from d9937b3 to c361f84 Compare June 18, 2026 13:07

kostandinang and others added 5 commits June 18, 2026 14:15

FE-883: record 1c decision (verify against folded tree) + integration…

01fd4cd

… plan 🍳 Built with brunch Co-Authored-By: Opus 4.8 <noreply@anthropic.com>

kostandinang and others added 3 commits June 18, 2026 14:15

kostandinang mentioned this pull request Jun 18, 2026

FE-884: Recover from cook epic-verification failure instead of halting #232

Open

kostandinang force-pushed the ka/fe-881-cook-agent-skills branch from a811396 to a29e185 Compare June 18, 2026 16:14

kostandinang force-pushed the ka/fe-883-orchestrator-improvements branch from c361f84 to 4cc6d23 Compare June 18, 2026 16:14

kostandinang closed this Jun 18, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

FE-883: Cook artifact lifecycle — git-merge slice composition#230

FE-883: Cook artifact lifecycle — git-merge slice composition#230
kostandinang wants to merge 8 commits into
ka/fe-881-cook-agent-skillsfrom
ka/fe-883-orchestrator-improvements

kostandinang commented Jun 18, 2026 •

edited

Loading

Uh oh!

kostandinang commented Jun 18, 2026 •

edited

Loading

Uh oh!

cursor Bot commented Jun 18, 2026 •

edited

Loading

Uh oh!

Uh oh!

cursor Bot left a comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

kostandinang commented Jun 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Stack Context

What changed

Effect

Verification

Follow-ups

Uh oh!

kostandinang commented Jun 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

cursor Bot commented Jun 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

PR Summary

Uh oh!

Uh oh!

cursor Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

kostandinang commented Jun 18, 2026 •

edited

Loading

kostandinang commented Jun 18, 2026 •

edited

Loading

cursor Bot commented Jun 18, 2026 •

edited

Loading