fix(go): gate resolver permission-mode to claude, truthful push reporting, test-before-push prompt

[AbirAbbas]

Rebuilt on top of current main after a deep review of the original branch. Three fixes land here; a fourth was dropped (see below).

1. Gate the resolver's writable permission mode to the claude runtime

run_pr_resolver was handed cfg.PermissionMode verbatim, which defaults to "". Under the claude harness an empty permission mode omits --permission-mode entirely, so claude --print falls back to its interactive prompting default — with nobody to answer the prompt, every write is denied and the resolver finishes having produced no commits.

The default is now "auto" (→ bypassPermissions), but only when the resolved provider is claude. Checked against sdk/go/harness/:

provider	empty mode	"auto"	verdict
claude	no flag → CLI prompts → read-only under -p	bypassPermissions	fix is real
codex	--sandbox workspace-write — already writable	--dangerously-bypass-approvals-and-sandbox	"auto" would be harmful: drops the sandbox around the whole machine for no benefit
opencode	PermissionMode is never read	never read	no-op; leave empty rather than imply a guarantee we do not make

An explicitly configured permission mode is still passed through unchanged for every provider.

2. Truthful push reporting

When the harness returns nothing parseable, run_pr_resolver hands back a deterministic all-false PRResolveResult. The orchestrator took that at face value — so a run where the agent committed and pushed but failed to emit a final structured answer was reported as "nothing happened", and the CI gate was skipped because pushed was false.

The head branch is now sampled on the remote (git ls-remote, since the agent pushes from its own process and leaves this workspace's tracking refs stale) before and after the agent runs, and only what that comparison can actually prove is recorded:

• report_invalid: true is set whenever the sentinel report is seen, independently of the remote, so callers can never mistake a reconstructed result for a first-hand one.
• Remote moved and the new tip == this workspace's HEAD → pushed=true. fixed stays false: a landed commit is not evidence that CI passes or that review comments were addressed. Overall success is fixed && pushed, so it correctly stays false.
• Remote moved to some other commit (a third party pushed while we ran) → remote_advanced: true and nothing is attributed. Their push is not ours to claim and their commits are not ours to describe.
• Remote did not move → nothing is claimed at all.

Commit/file lists are reconstructed only after git fetch origin <branch> makes the post-push objects resolvable locally. If that fetch fails, the lists stay empty and verification_partial: true marks them as unknown rather than as "nothing changed".

The sentinel string is now the exported ci.InvalidResolverReport const, referenced from both the fallback that emits it and the orchestrator that matches on it, so the two copies cannot drift.

3. Make the resolver test before it pushes

The resolver's instructions told it to re-run the failing tests, which says nothing about the tests its own edits break. Added an explicit test-and-lint-before-commit requirement to both halves of the prompt — as a properly numbered task step (incrementing the step counter, so it does not render as a bare paragraph wedged between two numbered steps) and as item 6 of the system prompt's "You are NOT done until". Goldens regenerated.

Dropped: the existing-branch-builds fix

All changes to go/internal/fast/build.go, go/internal/prompts/gitops/git_init.go and go/internal/roles/gitops/workspace.go from the original branch are removed. That fix targeted fast.Build, which is not the deployed path; it also contradicted the unmodified git-init system prompt and would double-open PRs. It will be redone separately against orch/build.go as a follow-up.

Test coverage

New tests in go/internal/orch/resolve_test.go:

• TestResolverPermissionModeGatedToClaude — drives real BuildConfigs for claude_code / codex / open_code and asserts auto / "" / "".
• TestResolverPermissionModeRespectsExplicitConfig — an explicit mode is never overridden.
• TestResolveSendsGatedPermissionMode — end-to-end: the permission_mode kwarg the reasoner actually receives is the gated value.
• TestClassifyRemoteAdvance — table over unchanged / unknown-before / unknown-after / ours / third-party / unknown-local-HEAD.
• TestResolverReportInvalidMatchesCISentinel — both fields, plus a real report is not flagged.
• TestResolveInvalidReportWithOurPushReportsPushedNotFixed — pushed=true, fixed=false, report_invalid=true, commit/file lists populated, overall success=false.
• TestResolveInvalidReportDegradesWhenFetchFails — pushed=true, verification_partial=true, no invented detail.
• TestResolveInvalidReportDoesNotAttributeThirdPartyPush — pushed stays false, remote_advanced=true, no commits described.
• TestResolveInvalidReportRemoteUnchangedClaimsNothing — nothing claimed.

Gates run locally (CI's go job, replicated verbatim on Go 1.23.0, GOWORK=off, with the pinned SDK sparse clone as a sibling)

gofmt -l . clean · go build ./... OK · go vet ./... OK · go test -race -count=1 ./... all 27 packages ok.

No Python files are touched, so the test job is unaffected.

🤖 Generated with Claude Code

[CLAassistant]

All committers have signed the CLA.