[Defilan]

What

maybeOpenPullRequest now resolves the coder's PR description before opening the Workload's PR on a reviewer GO, and a new githubpr.DescriptionBody renderer posts that description as-is instead of scaffolding it on top of the target repository's raw PR template.

• coderPRBody(ctx, task) finds the code task for the same branch: same foreman.llmkube.dev/workload label, Spec.Kind == issue-fix, Spec.Payload.Branch equal to the review task's, newest CreationTimestamp on a tie; decodes Status.Result.Raw through the extra.modelExtra envelope and returns a non-empty prBody.
• Body precedence at the open seam: coder prBody → reviewer prBody → the diff-grounded reviewer summary (today's path). The record gains extra.prBodySource (coder / reviewer / summary).
• DescriptionBody(desc, issue, workload): trimmed description, Fixes #<issue> appended ONLY when the description does not already contain #<issue>, always ending with the _Opened by foreman on review GO (workload %s)._ line. No template prepended.
• PRBody is byte-for-byte unchanged and still renders the summary fallback with the repo template ([FEATURE] Foreman-opened PRs should honour the target repository's PR template #1541). groundPRSummary and maybeRefreshPRBody are untouched.

Why

Fixes #1768. The PR opened on review GO was built from the reviewer's result, so the coder's complete description sat unread in status.result.extra.modelExtra.prBody and every Foreman PR shipped the unfilled template with unchecked boxes — out of compliance with the CONTRIBUTING.md AI-disclosure policy — plus a Fixes # that would auto-close issues the coder deliberately referenced with Refs #N. The coder is the only stage that knows what the change does and the only stage prompted to write a PR description, so its prose must win.

How

maybeOpenPullRequest calls coderPRBody before openPullRequest. The winning body travels as the summaryBody argument plus, for the coder, a COPY of r.Extra carrying it under prBody — r.Extra is persisted as the review task's record, so writing the coder's body into it would claim the reviewer authored it. The lookup is best-effort and silent by design: a nil e.Client, a list error, an empty workload label, no matching code task, a nil Result or malformed JSON all yield "" and keep today's grounded-summary behaviour.

In openPullRequest the renderer now depends on the body's provenance: a prBody from either author goes through DescriptionBody; the summary fallback still goes through PRBody(FindTemplate(workspace), …).

Tests: five new cases in executor_pr_test.go (coder body wins; sibling-branch coder task does not; nil client degrades; reviewer prBody still wins with no code task; newest coder task wins a fix-cycle tie; a workspace holding .github/pull_request_template.md contributes no template text to a prBody body) and TestDescriptionBody / TestDescriptionBody_NeverCarriesTemplateText in template_test.go. The existing PRBody and TestOpenPullRequest_PrBodyPreferred expectations are unchanged; TestMaybeOpenPullRequest_BodyCarriesReviewSummary gained a prBodySource == "summary" assertion. Mutation-checked: forcing coderPRBody to return "" fails TestMaybeOpenPullRequest_CoderPRBodyPreferred ("body must be the coder's description…", "prBodySource = summary, want "coder"").

Checklist

• Tests added/updated
• make test passes locally — run for the touched packages (go test ./pkg/foreman/agent/... ./pkg/foreman/agent/githubpr/ -count=1, all ok) plus go build ./...; the full envtest suite runs in the clean-room gate
• make lint passes locally — gofmt clean, go vet clean, and bin/golangci-lint run reports 0 issues on both touched packages after the over-long-line fix
• Commit messages follow conventional commits
• All commits are signed off (git commit -s) per DCO
• AI assistance (if any) is disclosed above, per CONTRIBUTING.md — this PR was produced by Foreman's coder agent under operator supervision; a human is accountable for it
• Documentation updated (if user-facing change) — no doc change: this fix restores the documented intent (the coder's prBody is the PR description), and no file under docs/ describes the PR-body composition path

Revisions

• r1 (wl-1768-revise-r1, 32 turns, same coder): findings from the adversarial review. The issue-reference check in DescriptionBody was a bare substring match (#1768 satisfied issue 176, #17680 satisfied 1768). It is now digit-bounded, with three table cases, and two tests cover the coder-task fallbacks (no prBody key, malformed result). The rebase revision rewrote the branch into a single commit, so the commit message is r1's; this PR's title carries the change's subject.
• bad2378f (test-only, human commit): drops the unparam-flagged parameter from the codeTaskOnBranch helper so make lint passes. The r1 gate failed on exactly that finding and nothing else.

Provenance

Assisted-by: Foreman coder agent (qwen38fn-coder, Qwen3.8-Flash-Next NVFP4 on two DGX Sparks, workload wl-1768-coder-prbody, 92 turns) for the change and the description above; the r1 revision applied the human review's findings. Adversarial review, three hand mutations (each caught), the local make test gate on the first commit, the boundary probes, and the lint commit were done by Chris with Claude Code (session https://claude.ai/code/session_01XzCYKzbnCnoYWZmdVHmwqg). A human is accountable for every commit. The full make test envtest suite passed locally on the final head bad2378f (38 packages ok), and make lint is clean on it.

This body was replaced by hand after the PR opened: the operator serving this repository still runs the pre-fix code, so the body it generated was the very defect #1768 describes (raw template, unchecked boxes, unfilled Fixes #).

[codecov]

Codecov Report

❌ Patch coverage is 91.42857% with 6 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
pkg/foreman/agent/executor_native.go	89.83%	3 Missing and 3 partials ⚠️

📢 Thoughts on this report? Let us know!

[joryirving]

This is clean and I'd merge it. The precedence is right in both seams and prBodySource reports it accurately in all three branches, the copy-not-mutate of the reviewer's extra is the correct call and the comment says why, and the digit-bounded reference regex holds up on every boundary I threw at it including the Refs #7 multi-slice case. Verified r.Extra is non-nil via modelDecidedResult so the new prBodySource write can't panic.

One optional thing: the newest-coder-task pick at executor_native.go:2237 has no tiebreaker for equal CreationTimestamp, so two issue-fix tasks on the same branch in the same second would resolve by list order. It can't happen in the real revision flow, so it's just hardening — a Name tiebreaker would close it.

Tests are thorough and green locally: built the touched packages, go vet clean, and ran the DescriptionBody and PR suites.