[tbhb-agent]

Summary

agentrepo retro new preserves the transcript of the current session as evidence for a retrospective, matching the behavior of the shell script that does this today. No shell entry point calls the command yet.

Why

The shell-to-Go migration needs a Go command with the same behavior before any caller can switch to it. This slice adds the first retro command and leaves the shell script and its callers untouched, so the switch can be a separate, reviewable change. Later slices add the preflight command and the caller switch, and they close the issue.

Verification

The commands below ran on the head of this pull request.

Criterion	Command	Result
AC4, AC5, AC7, AC11 creation behavior	go test ./internal/retro ./internal/agentrepo -count=1	passes
AC8 existing shell suite	mise run test-hooks tests/skills/analyze-session	passes
AC8 generated copies	mise run check-generated-skills	passes
AC9 shell lint	mise run lint-shell	passes
AC10 coverage floors	mise run cover-go	passes
Windows build	GOOS=windows go vet ./internal/retro ./internal/agentrepo	passes
Lint	mise run lint-go	passes

Risk

Existing shell workflows don't invoke the new command. When someone invokes it directly, a defect could leave retrospective evidence incomplete or write it to the wrong location. The tests replace the report tool and a failing path lookup with stand-ins on PATH, because the CI test job has no uv. They check the arguments, environment, files written, exit statuses and output fields, and they don't check the real report text. Reverting the change removes the command and its tests.

Compatibility: none

Loosens: no

Related

Related to #331, #1031, #1032

[tbhb-dev-reviews]

Reviewed head: 5bc7f35
Review: CHANGES REQUESTED
Loosens: no
Owner-tier: no

Reviewed the live #331 body and all five Refinement comments, including the latest slice crosswalk. Reviewed the exact pinned checkout against origin/main (merge base 8bcd565). No gates or tests were rerun; execution evidence below comes from existing CI logs. No delegation was used.

Acceptance criteria

AC	Status	Evidence on this head
AC4 [BASE-1]	Met	Main checkout derives from the environment-isolated Git common directory at internal/retro/new.go:78; metadata naming and harness placement at internal/retro/new.go:97; LC_ALL=C and PYTHONUTF8=1 at internal/retro/new.go:150. Both harnesses are asserted at internal/retro/retro_test.go:83; worktree isolation at internal/retro/retro_test.go:123; precedence at internal/retro/retro_test.go:233.
AC5 [BASE-1]	Met	Existing RETRO.md is retained at internal/retro/new.go:248; local exclude deduplication at internal/retro/new.go:261. Byte preservation and one exclude entry are asserted at internal/retro/retro_test.go:112; an existing .gitignore remains unchanged at internal/retro/retro_test.go:149.
AC7 [both slices]	Met for BASE-1 on Unix; Windows evidence incomplete	Creation scenarios and repeat creation remain at internal/retro/retro_test.go:80. Failed Git output is discarded before metadata or file creation at internal/retro/new.go:78; nonzero/no-output/no-retro assertions remain at internal/retro/retro_test.go:226. However, internal/retro/retro_test.go:215 skips that failure-safety scenario on Windows (F2 below). BASE-2 owns the preflight equivalents.
AC8 [each slice]	Met	No maintained/generated skill or shell-test path changes. The unchanged suite has 20 cases at tests/skills/analyze-session/scripts/analyze-session.bats:56. Existing Ubuntu Bats log records all 20 as ok 282 through ok 301, including all six entry scenarios. Both Bats jobs succeeded. Existing lint log records check-generated-skills succeeding; its wiring is unchanged.
AC9 [each slice]	Met	The complete eight-path diff below contains only Go and JSONL fixtures. No shell, generator or gate change. Existing lint job records lint-shell, lint-shell-allowlist, lint-shell-fmt and check-generated-skills completing successfully.
AC10 [each slice]	Not met: required threshold evidence absent	Floors remain 75/80/85 at .testcoverage.yml:12, and no exclusion changes. Fresh CI execution includes internal/retro (85.4% package statements) and internal/agentrepo (92.2% on Ubuntu). But the coverage job only merges/renders (.github/workflows/ci.yml:281; .devtools/tasks/go.toml:214); it does not run the threshold gate at .devtools/tasks/go.toml:228. Neither PR Verification nor comments supplies an exact-head fresh-profile threshold pass. See F1.
AC11 [BASE-1]	Met	Usage/status 2 at internal/retro/new.go:35 and internal/agentrepo/retro_test.go:18; validation/status 1 at internal/retro/new.go:64 and internal/retro/retro_test.go:179; four ordered summary fields at internal/retro/new.go:116 and internal/retro/retro_test.go:94.

Struck/moved crosswalk: AC1 -> removal #1003; AC2 -> CHILD-B #1032; AC6 -> CHILD-A #1031; AC12 -> CHILD-B #1032; AC13 merged into AC6/#1031; AC14 merged into AC4 above. AC3 is BASE-2 only. The PR contains only BASE-1 registration, creation, shared helpers, creation tests and fixtures: internal/agentrepo/command.go:97, internal/agentrepo/retro.go:17, internal/retro/new.go:33. There is no preflight implementation, CHILD-A correction or CHILD-B parity implementation. The inherited copying/reporting scaffolding is explicitly permitted by #331. The PR has no closing issues and only Related to #331, correctly leaving #331 open for BASE-2.

Size and changed paths

Measured with scc 4.1.0 on the exact head, additions plus deletions versus origin/main...HEAD, with new files in full and JSONL counted as JSON code. Tests count as code. The one added command.go statement was measured separately as Go. Prose means prose files; code comments/blanks are reported separately.

Name-status / path	Added + deleted raw lines	scc code	Prose	Owner-tier
M internal/agentrepo/command.go	1 + 0	1	0	no
A internal/agentrepo/retro.go	27 + 0	21	0	no
A internal/agentrepo/retro_test.go	25 + 0	18	0	no
A internal/retro/fake_uv_test.go	98 + 0	82	0	no
A internal/retro/new.go	328 + 0	305	0	no
A internal/retro/retro_test.go	264 + 0	244	0	no
A internal/retro/testdata/claude.jsonl	7 + 0	7	0	no
A internal/retro/testdata/codex.jsonl	8 + 0	8	0	no
Total	758 + 0	686	0	no

There are additionally 19 code-comment lines and 53 blank lines. Even counting comments as prose gives 19, below 2,000. Code is below 800; no size/split finding. No tests should be trimmed. No listed protected policy, CODEOWNERS, agent-skills.conf, review instructions, automode/settings, Codex configuration, workflow/action, library, guard or merge-script path changes. The default CODEOWNERS ownership of all files does not make ordinary Go changes owner-tier.

Loosening inspection: no gate/check/exclusion/floor/allowlist change; zero occurrences of the configured loosening markers in added lines. Git grep of the changed head files and the merge-base diff confirms zero nolint directives. Comparison to cad4a75 confirms all eight former directives were removed. Thus Loosens: no is accurate for the diff, but the live PR description does not actually contain a Loosens declaration.

Findings

• F1 [P2] Supply the missing fresh-profile threshold result for AC10. A green coverage context is not that result: .github/workflows/ci.yml:284 invokes the merge task, whose body at .devtools/tasks/go.toml:223 only merges coverage and renders reports. The threshold command is separate at .devtools/tasks/go.toml:230. The existing coverage log also records an HTTP 403 upload failure tolerated by the unchanged fail-on-error setting at .github/workflows/ci.yml:292. Provide existing exact-head evidence that all three unchanged floors pass; this review did not rerun a gate and does not allege that the floors fail.
• F2 [P2] Remove the Windows hole in failed-path-resolution coverage at internal/retro/retro_test.go:215. There is still a t.Skip for the Python-shebang Git fixture. It existed at cad4a75 and was not introduced by the uv replacement, but it contradicts this review's explicit no-skip requirement and means a green windows-2025 job never exercises these AC7 failure-safety assertions. Use a portable fake Git subprocess, retaining the test name and every outcome assertion.
• F3 [P2] Remove the one-call subprocess wrapper at internal/retro/new.go:146. It adds no behavior, injection point or AC-required boundary. Its sole caller at internal/retro/new.go:137 can directly call exec.CommandContext with the same arguments. Retain the existing lint gate and avoid reintroducing suppressions.
• F4 [P2] Complete the PR delivery evidence (D1/D9). The live Verification table lists targeted commands but no AC IDs, AC8/AC9 delivery results or AC10 threshold evidence; Related lists only #331 and omits the filed deferred-work owners #1031 and #1032. Add the per-AC status/evidence and the deferred-issue links. Also add the requested truthful Loosens: no declaration. No repository file edit is needed for this description correction.

Test preservation, formats and Windows paths

All seven original TestRetro* names and TestRetroCommand remain. The no-skip check fails: exactly one t.Skip remains, at internal/retro/retro_test.go:216 (F2). No test-body outcome assertion was weakened or removed versus cad4a75. A literal claim of zero assertion-line differences would be inaccurate: setup helpers changed to remove suppressions. The complete assertion-line changes are:

• write helper: os.MkdirAll/os.WriteFile NoError checks became OpenRoot/Rel NoError checks plus Root.MkdirAll/Root.WriteFile checks (internal/retro/retro_test.go:27).
• worktree setup: an additional NoError checks EvalSymlinks (internal/retro/retro_test.go:127).
• executable Git fixture setup: os.Chmod NoError became OpenRoot NoError plus Root.Chmod NoError (internal/retro/retro_test.go:221).
• the three embedded Python asserts for LC_ALL, PYTHONUTF8 and format arguments moved to fakeUV's nonzero-return checks at internal/retro/fake_uv_test.go:34. They are still observed by the unchanged success assertions.

No other original assertion line changed; internal/agentrepo/retro_test.go is identical to cad4a75. Additional non-assertion setup changes are the fake-uv link, removal of the Python stub, explicit DEVTOOLS_HARNESS=claude in two tests, and ToSlash on the harness-precedence input at internal/retro/retro_test.go:245. Within TestRetroFailuresAndFormat itself, the only setup changes are the auto format-loop entry and RETRO_FAKE_META. The live PR description discloses the stand-in and that real report text is not tested.

Explicit ANALYZE_SESSION_FORMAT=auto is accepted at internal/retro/new.go:128 and exercised at internal/retro/retro_test.go:189; it reaches metadata and reports as auto, as the shell accepts at packages/agents/source/skills/analyze-session/shared/scripts/new-retro.sh:76 and packages/agents/source/skills/analyze-session/shared/scripts/session.py:812. An unsupported nonempty value now exits 1 with new-retro: unsupported ANALYZE_SESSION_FORMAT, rather than Python argparse's invalid-choice diagnostic/status 2 propagated by the shell. Examined AC4, AC5, AC7, AC8, AC9, AC10 and AC11: none requires invalid-format diagnostic/status parity in BASE-1. AC7 expressly excludes additional AC12 edge cases; AC8 requires retention of the unchanged shell, not full Go parity. AC12 is deferred to #1032. This difference is outside the current slice's ACs, not an additional parity finding.

The Windows path change is already present in this pinned head; comparison with 1f36240 shows the ToSlash production change plus worktree-root normalization and the harness-precedence setup normalization. Comparing the pinned head with 5bc7f35 itself naturally gives no diff. internal/retro/new.go:102 renders the same forward-slash common-directory/main-checkout path as the frozen shell's construction at packages/agents/source/skills/analyze-session/shared/scripts/new-retro.sh:105 and print at line 156. On Linux and macOS filepath.Separator is '/', so ToSlash is the identity; splitting the same concatenation into name changes no value. By source comparison, successful stdout there is byte-identical to the prior 1f36240 head for the same inputs/state. This is a source proof, not a rerun. On Windows ToSlash replaces backslashes with the shell/Git's forward slashes.

EvalSymlinks and ToSlash are confined to the test's initial main-checkout root setup at internal/retro/retro_test.go:125. They do not normalize the command output. The unchanged assertion at line 152 still requires that specific main checkout root in stdout; line 154 still requires no retro in the distinct linked checkout. This proves the intended main-checkout placement outside the calling linked worktree and cannot turn placement in the linked worktree into success.

Over-engineering

F3 is the unnecessary layer. The shared copying/scaffolding helpers and test subprocess stand-in otherwise have concrete call sites or retained scenarios. No additional speculative error-handling finding.

Checks, provenance and definition of done

The one requested status snapshot reports every context COMPLETED SUCCESS. Coverage is COMPLETED SUCCESS, not skipped; its log proves fresh merge/render but not the threshold gate, and the upload reports tolerated HTTP 403. CodeQL and both analyze jobs are COMPLETED SUCCESS. The four OS jobs (ubuntu-24.04, ubuntu-24.04-arm, macos-15, windows-2025) are all COMPLETED SUCCESS on the pinned head. Mergeability is MERGEABLE; mergeStateStatus is BLOCKED.

GitHub reports the sole head commit signature verified=true, reason=valid. Both Assisted-by trailers are present, as is the Signed-off-by trailer; the PR has provenance:agent-assisted. Help text matches the added command (internal/agentrepo/retro.go:19; internal/retro/new.go:38). Existing supported entry points remain unchanged, so Compatibility: none is appropriate.

• D1: Unmet author delivery evidence: no per-AC PR Verification crosswalk and no threshold-pass result; merged-head verification remains a merge-time duty (F1/F4).
• D2: Unmet: this independent review requests changes on the exact head; findings await resolution.
• D3: All contexts green on this head; no loosened gate. The separate AC10 evidence gap remains F1.
• D4: Gitleaks passed and CI's repository leak scan passed, but CI explicitly loaded no private values. This review body is scanned with check-leaks-with-lists before posting. An exact-head repository scan using its private-strings list is not evidenced in the PR.
• D5: The sole current commit has a GitHub-verified signature. A future squash signature remains a merge-time duty.
• D6: New command help matches the dormant behavior at internal/agentrepo/retro.go:19 and internal/retro/new.go:38; no supported entry-point/doc contract changes.
• D7: Current commit Assisted-by trailers and provenance label verified. A future squash trailer remains a merge-time duty.
• D8: 686 code / 0 prose (19 if code comments counted as prose), below 800/2,000.
• D9: Deferred owners exist as #1031/#1032 with Problem and numbered ACs, but the PR Related section does not list them (F4).
• D10: Intentionally pending BASE-2 and merge: this BASE-1 PR must not close #331. No closing reference exists; final Done/merge-status evidence cannot yet exist.

Round: 1 of 3

[tbhb-dev-reviews]

Reviewed head: 65dff0e
Review: APPROVED
Loosens: no
Owner-tier: no

No new blocking findings. F1-F4 from round 1 are resolved for this BASE-1 delivery. Approval covers this exact head and this slice; the outstanding merge-time and private-list evidence duties are identified in the checklist below.

Read the live #331 body and all six comments, including every Refinement comment and the latest slice crosswalk. Examined the pinned Git objects in an isolated source snapshot; the operator's unrelated working-tree edits were left untouched. Compared against origin/main at 8bcd565 and inspected the complete amend from 5bc7f35. No tests or mechanical gates were rerun. No delegation was used.

Round-1 findings

Finding	Resolution on this head
F1: fresh coverage threshold result	Fixed by the live PR Verification edit: it now explicitly reports AC10, mise run cover-go, passes, and says the listed commands ran on the PR head. This is author-supplied exact-head execution evidence. The unchanged task generates a fresh profile at .devtools/tasks/go.toml:167 and invokes the threshold checker at .devtools/tasks/go.toml:171. Floors remain 75/80/85 at .testcoverage.yml:12. I did not independently rerun that command. The green coverage context alone is still not evidence of threshold execution.
F2: Windows path-failure skip	Fixed: internal/retro/retro_test.go:213 now unconditionally runs the scenario. internal/retro/fake_uv_test.go:28 emits partial path output and exits 9; internal/retro/fake_uv_test.go:100 installs the actual test executable as Git, with .exe on Windows. All four failure outcome assertions remain at internal/retro/retro_test.go:218. No t.Skip remains in the changed files. The existing Windows job passes internal/retro with this fixture.
F3: one-call subprocess wrapper	Fixed: internal/retro/new.go:137 directly invokes exec.CommandContext; the wrapper was deleted. Environment handling remains at internal/retro/new.go:138. No suppression was added.
F4: PR delivery evidence	Fixed in the live description: Verification now names AC4/5/7/11, AC8 shell and generated-copy results, AC9 shell lint and AC10 fresh coverage, all as passing. Related now includes #331, #1031 and #1032, and Risk explicitly says Loosens: no. The description continues to disclose the subprocess stand-ins and lack of real-report-text assertions. These are description edits, not repository-file changes.

Acceptance and scope

Criterion	Status and evidence
AC4 [BASE-1]	Met. Git common-directory lookup and failed-output rejection: internal/retro/new.go:78. Metadata name and harness location: internal/retro/new.go:97. Inherited Git/CDPATH variables are removed and report locale fixed at internal/retro/new.go:146. Both harnesses: internal/retro/retro_test.go:82. Linked-worktree placement and hostile inherited environment: internal/retro/retro_test.go:122. Report environment and format checks: internal/retro/fake_uv_test.go:41. Harness precedence/default location: internal/retro/retro_test.go:224 and internal/retro/retro_test.go:246.
AC5 [BASE-1]	Met. Existing RETRO.md remains untouched at internal/retro/new.go:244; exclude deduplication at internal/retro/new.go:271. Repeated creation preserves exact bytes and one exclude entry at internal/retro/retro_test.go:111. Existing .gitignore preservation remains at internal/retro/retro_test.go:148.
AC7 [both slices]	Met for BASE-1's creation and failed-path-resolution portion. The three creation scenarios remain at internal/retro/retro_test.go:79; partial-path failure outcomes remain at internal/retro/retro_test.go:213 on every platform. The other three entry scenarios belong to BASE-2. All six remain in the unchanged shell suite, tests/skills/analyze-session/scripts/analyze-session.bats:184.
AC8 [each slice]	Met. No maintained/generated skill or shell-suite path changes. The suite still has exactly 20 cases beginning at tests/skills/analyze-session/scripts/analyze-session.bats:56. Existing Ubuntu Bats job 124855119819 records all 20 as ok 282-301, including all six entry scenarios. Both Bats contexts succeed. Existing lint job 124855119790 records check-generated-skills completing successfully.
AC9 [each slice]	Met. The complete eight-path diff below has no shell, generator or gate changes. Existing lint job 124855119790 records lint-shell, lint-shell-allowlist, lint-shell-fmt and check-generated-skills completing successfully.
AC10 [each slice]	Met on the author's exact-head mise run cover-go passing result now published under Verification; task freshness and threshold invocation are .devtools/tasks/go.toml:167 and .devtools/tasks/go.toml:171. The floors and exclusions are unchanged. Existing Ubuntu CI also records internal/retro 85.3% and internal/agentrepo 92.2% statements, confirming the new package participates, but those package numbers are not a substitute for the reported full threshold pass.
AC11 [BASE-1]	Met. Usage/status 2: internal/retro/new.go:35 and internal/agentrepo/retro_test.go:18. Missing input/template/repository diagnostics and status 1: internal/retro/retro_test.go:178. Successful four-field order: internal/retro/new.go:116 and internal/retro/retro_test.go:93.

AC3 is BASE-2. AC1 remains moved to removal #1003; AC2/AC12 to CHILD-B #1032; AC6 and merged AC13 to CHILD-A #1031; AC14 is merged into AC4. Those deferred guarantees are not claimed here. Command registration at internal/agentrepo/command.go:97 and internal/agentrepo/retro.go:17 adds only retro new. There is no preflight implementation, caller activation, CHILD-A correction or CHILD-B correction. Shared copying/reporting scaffolding is explicitly permitted by #331. The PR relates to #331 and has no closing reference, correctly leaving it open for BASE-2. Existing supported shell entry points remain unchanged, consistent with Compatibility: none.

Preservation and amend review

All seven TestRetro* cases plus TestRetroCommand remain. No test-body outcome assertion was removed or weakened against round 1 or the earlier cad4a75 base. The amend replaces only the failing-Git fixture setup, removes its Windows skip and obsolete runtime import, shares the already-existing executable-link setup between uv and Git, and deletes the production one-call wrapper. The former fixture chmod check is replaced by the executable-path/link NoError checks at internal/retro/fake_uv_test.go:102. The four path-failure outcomes are unchanged at internal/retro/retro_test.go:218. The fake-Git branch deliberately precedes fake-uv dispatch because that scenario inherits the uv fixture flag; it therefore actually produces the failed Git output under test. The shared helper has two concrete callers, not speculative extension points.

Compared with cad4a75, the earlier setup refactors retain the embedded locale/format assertions as nonzero subprocess outcomes at internal/retro/fake_uv_test.go:41; success assertions still observe them. Worktree-root normalization at internal/retro/retro_test.go:126 does not normalize the command output, and the assertion still demands the main checkout path while rejecting a retro in the linked checkout at internal/retro/retro_test.go:151. Test bodies and fixtures were not shortened to fit size.

Size and ownership

Measured with scc 4.1.0, additions plus deletions against origin/main...the pinned head, new files counted in full, tests counted as code and JSONL counted as JSON. The modified command.go contributes its one added Go statement. The table lists every name-status path from that diff; HEAD here means the reviewed pinned head, not the operator's separate working checkout.

Name-status / path	Added + deleted raw lines	Code	Prose files	Comments	Blank	Owner-tier
M internal/agentrepo/command.go	1 + 0	1	0	0	0	no
A internal/agentrepo/retro.go	27 + 0	21	0	2	4	no
A internal/agentrepo/retro_test.go	25 + 0	18	0	2	5	no
A internal/retro/fake_uv_test.go	110 + 0	92	0	9	9	no
A internal/retro/new.go	324 + 0	302	0	4	18	no
A internal/retro/retro_test.go	255 + 0	235	0	3	17	no
A internal/retro/testdata/claude.jsonl	7 + 0	7	0	0	0	no
A internal/retro/testdata/codex.jsonl	8 + 0	8	0	0	0	no
Total	757 + 0	684	0	20	53	no

684 code is below 800; 0 prose-file lines, or 20 including code comments, is below 2,000. No size finding or further split is required.

All changed paths fail to match the owner, golden, constitutional and gate-configuration patterns in .agents/agent-skills.conf:4. No policy, CODEOWNERS, workflow, settings, protected library, guard or merge-script change exists. There are zero configured loosening-marker occurrences in any changed file, using .agents/agent-skills.conf:6, including zero nolint. No floor, exclusion, gate or assertion is loosened. Loosens: no; Owner-tier: no.

Over-engineering

No remaining P2 guard, option, abstraction or layer without an acceptance need was found. The removed command wrapper resolves the prior finding. The skill-directory option at internal/agentrepo/retro.go:15 supplies the deployed scripts/template boundary; environment isolation implements AC4; artifact helpers retain the explicitly allowed base scaffolding; the shared executable-link fixture serves both uv and Git.

Checks and definition of done

The single requested status snapshot reports every context COMPLETED SUCCESS. Mergeability: MERGEABLE; mergeStateStatus: BLOCKED. All four OS test contexts pass, including windows-2025. Existing logs were read; no gate or test was rerun.

Coverage: COMPLETED SUCCESS, not skipped. The existing coverage job merges/renders at .github/workflows/ci.yml:281; upload still reports HTTP 403 because repository code quality is not enabled, tolerated by the unchanged fail-on-error: false at .github/workflows/ci.yml:292. The author's separate fresh-profile threshold-pass result resolves F1; I do not infer that result from the green coverage context.

CodeQL: CodeQL, analyze (go) and analyze (actions) are all COMPLETED SUCCESS.

GitHub verifies the sole commit signature (verified=true, reason=valid). Both Assisted-by trailers and Signed-off-by are present; provenance:agent-assisted is present.

• D1: Slice AC verification and evidence are now published in the PR and covered above; verification on the eventual merged head remains a merge-time duty.
• D2: This independent review approves the exact current head and F1-F4 are resolved; approval of the eventual merged head remains a merge-time duty.
• D3: All contexts pass on the pinned head; no loosened gate or owner exception.
• D4: Gitleaks passes and CI's check-leaks passes, but CI explicitly loaded no private values. An exact-head repository scan with its private-strings list is not evidenced in the PR and remains an unmet delivery-evidence duty. This post is leak-reviewed and scanned by gh-post before publication; no repository gate was rerun.
• D5: The current commit is GitHub-verified signed; the future squash signature remains a merge-time duty.
• D6: Help matches the new dormant command at internal/agentrepo/retro.go:19 and internal/retro/new.go:38; supported shell workflows are unchanged.
• D7: Current Assisted-by trailers and provenance label are verified; the future squash trailer remains a merge-time duty.
• D8: 684 code and 0 prose-file lines (20 including code comments), below 800/2,000.
• D9: #1031 and #1032 each have a Problem and numbered ACs; the live PR Related section now lists both. Deferred ownership remains explicit in #331 and the crosswalk above.
• D10: Intentionally pending BASE-2 and merge. BASE-1 must leave #331 open; no closing reference exists. The final Done/merge-status evidence cannot yet exist.

Round: 2 of 3