Use semantic dedupe for approved future follow-up filing

[wwind123]

Summary

• Add deterministic-first, bounded semantic reconciliation for approved future follow-ups.
• Search open repository follow-up trackers using explicit source context, validate identities, revalidate open state, and reuse only high-confidence duplicates.
• Preserve conservative fallback behavior, creation-cap accounting, replay gates, interruption recovery, usage accounting, and sanitized publication text across planning, PR, and skill-mode publishers.
• Add CLI/configuration controls and regression coverage for the Follow up future plan-review note: Untracked-only salvage: following the issue guardrail, a failed run whose only local  #484/Follow up future review note: Untracked-only partial work currently produces no salvage at all, since capture keys off ` #488 and Follow up future plan-review note: Track coder-dispatch guardrails and credential isolation or API broker design in a de #746/Follow up future review note: Track coder-dispatch guardrails and credential isolation or an API broker in a dedicated f #748 wording differences.

Verification

• Focused follow-up/orchestration/skill/recovery tests: 585 passed.
• Backend/protocol/docs/agent-loop tests: 720 passed.
• Final managed-CI and follow-up safety tests: 124 passed.
• Full suite: 2,551 passed; two containment tests remain environment-specific because this runner inherits systemd-cgroup-v2-inherited rather than creating the direct scope those tests assert.

Fixes #490

[wwind123]

Expected closing issues for PR #751: #490.

Origin flow: approved-plan-implementation
Contract hash: def1a70ddfd414d36fb8396e2442925040e567452022c3ac1beebcbf02e68953

-- coding-review-agent-loop

[wwind123]

Issue implementation

Implemented the approved semantic follow-up dedupe plan, committed it as ad517fe, pushed the branch, and opened PR #751. Focused tests and all non-containment tests passed; the full suite has two environment-specific inherited-cgroup containment failures.

Result

Pull request reported: #751.

Tests run

• python3 -m pytest tests/test_followups.py tests/test_orchestrator_issue.py tests/test_orchestrator_pr.py tests/test_skill_helpers.py tests/test_split_materialization.py tests/test_decomposition.py -q (agent-loop instrumented; whole-command timeout 180s)
• python3 -m pytest tests/test_backends.py tests/test_protocol_markers.py tests/test_protocol.py tests/test_docs_guidance.py tests/test_agent_loop.py -q (agent-loop instrumented; whole-command timeout 180s)
• python3 -m pytest tests/test_managed_ci.py tests/test_followups.py -q (agent-loop instrumented; whole-command timeout 120s)
• python3 -m pytest -q (agent-loop instrumented; whole-command timeout 180s)

-- OpenAI Codex: gpt-5.6-luna (xhigh)

[wwind123]

Review verdict: Blocking

The two-stage design is sound: deterministic dedupe stays the first pass, the model call is bounded, quota-reset escapes cleanly, replay is checked before any provider/search activity, and only high confidence suppresses. Three problems block merge: the antigravity semantic backend can never run because the isolated config it builds is rejected by AgentLoopConfig validation (and leaks its temp dir on that path); PR-filed follow-up trackers never render the lookup context that was plumbed in for them, so they are invisible to the parent-issue searches this feature relies on; and two acceptance-criteria behaviors (in-batch semantic merge, medium/low-confidence 'uncertain' filing) have no test. GitHub checks are still pending (test in_progress), so I have no global test result; per instructions I did not run tests locally.

Blocking issues

• semantic_dedupe._isolated_provider_config is broken for --semantic-followup-backend antigravity: it sets antigravity_model while dataclasses.replace carries the already-resolved multi-entry antigravity_models chain forward, so AgentLoopConfig.__post_init__ always raises "Cannot specify both antigravity_model and a custom antigravity_models chain." (default chain has three entries; an explicit --semantic-followup-model hits the same check). The exception is swallowed by the broad except Exception in _find_existing_for_group / _semantic_batch_matcher, so one of the four advertised backends silently degrades to deterministic-only on every call with no user-visible error. Set antigravity_models=(model or config.antigravity_models[0],) and leave antigravity_model unset instead. Related on the same path: tempfile.mkdtemp runs before the try/finally in default_semantic_transport, so every failure inside _isolated_provider_config leaks the temp directory — move the mkdtemp inside the protected region or clean it up on the raise path.
• followups._followup_issue_body accepts the new source_context parameter but never renders it, while _plan_followup_issue_body emits a - Lookup context: ... line. As a result PR-review-filed trackers record only "...approved review on PR #" and contain no parent issue number or plan identity. _search_followup_trackers narrows primarily with "#<parent>" "follow-up" queries and _candidate_context_matches looks for #<number> in the candidate text, so a tracker filed from PR review cannot be found by a later plan-stage run on the same parent issue, nor by a later PR on that issue — only the bounded topic-term fallback query can reach it. Render source_context.render() in _followup_issue_body as the plan body does (or drop the parameter and state the asymmetry deliberately).
• Two acceptance-criteria behaviors introduced here are untested. (a) In-batch semantic merge: nothing exercises reconcile_approved_followups(..., semantic_matcher=...) or _semantic_batch_matcher, so the group-N identity round-trip through parse_semantic_match and back into existing_index is unverified, as is the requirement that only high confidence merges. (b) The medium/low-confidence path: no test covers status == "uncertain", i.e. that the item is still filed, that the body gains the sanitized "Possible duplicate" section, and that the summary reports it as filed-and-uncertain rather than reused. Add a focused test for each in tests/test_followups.py.

Same-PR follow-ups

• helpers/skill_runner.py uses inspect.signature(_publish_approved_followups) at call time to decide whether to pass source_context. That adapter exists only to keep two test doubles in tests/test_skill_helpers.py (fake_publish at ~line 1839 and ~line 1884) working with the old signature. Update those doubles to accept **kwargs / source_context and delete the runtime introspection from production code.
• helpers/skill_runner._publish_pr_followups builds a FollowupSourceContext with related_pr_numbers only and no parent_issue_numbers, so the skill-mode flow loses the parent-issue query that is the primary narrowing signal for the plan-to-PR case this PR is meant to fix. Thread the issue number through from the caller if it is available there.
• SemanticDedupeMatcher.match records usage with outcome="succeeded" and sets validation_status = "validated" before parse_semantic_match runs, so a provider response that fails contract validation is still counted as a validated success. Record the result after parsing, or mark it invalid when parsing raises.
• In reconcile_approved_followups, the semantic branch accepts both a group-N string (1-based) and a bare int (treated as 0-based) for duplicate_of. The int form is unreachable through parse_semantic_match (batch candidate identities are strings, so an int fails the allowed-ids check) and the two indexing conventions in adjacent lines are easy to misread. Drop the int branch or make the indexing consistent and document it.
• _create_approved_followup_issues, _create_plan_approved_followup_issues, and _format_created_followup_issue_summary in followups.py now have no callers; the first and third are still imported (unused) at orchestrator.py:343 and :349. Remove the dead functions and imports.
• _format_publication_summary is always given the heading "Created approved-review future follow-up issues for PR #N:" / "Created approved-plan future follow-up issues:", including when every group was reused and nothing was created. Similarly _publish_plan_approved_followups labels the reused URLs under "Filed future follow-up issues:". Adjust the wording so a reuse-only publication does not claim issues were filed.

-- Anthropic Claude: claude-opus-5 (medium)

[wwind123]

Review verdict: Blocking

Blocking review: GitHub PR checks are currently pending (test in progress). The semantic deduplication design is well-structured and handles existing tracker reuse, but there are several blocking correctness and documentation issues: parent issues are not excluded from the candidate search in PR review mode (risking suppression of future follow-ups against the very issue being closed), build_semantic_prompt candidate budgeting overflows prompt_char_limit and blindly slices the prompt (dropping up to 75% of candidates), _format_plan_approval_summary_with_followups produces contradictory reconciliation counts and misreports reused trackers as newly filed, user-facing CLI flags are missing from README.md, and in-batch semantic deduplication lacks test coverage.

Blocking issues

• In src/coding_review_agent_loop/followups.py, _search_followup_trackers defines excluded_numbers as only {source_context.source_number} when source_context.source_kind == 'plan'. In PR mode (source_kind == 'pr'), parent_issue_numbers is not excluded. When searching for open follow-ups via repo:OWNER/REPO is:issue is:open "#{parent}" "follow-up", the PR's open parent issue can be returned. If the parent issue was itself created as a follow-up or contains follow-up headings, it matches _looks_like_followup_tracker and can be matched deterministically or semantically as a duplicate. This marks the follow-up as reused against the parent issue, so no new issue is filed; when the PR merges and closes the parent issue, the unfiled follow-up is lost. excluded_numbers must include *source_context.parent_issue_numbers.
• In src/coding_review_agent_loop/semantic_dedupe.py, build_semantic_prompt calculates candidate_budget = max(256, prompt_char_limit // len(candidates)) without accounting for prompt instructions, source context, or proposed text (~3,300 chars overhead), and allocates candidate_budget // 3 for title plus candidate_budget for body per candidate (summing to ~1.33x candidate budget per candidate). With batches near the default candidate cap (50), the rendered prompt exceeds 20,000 characters before [:prompt_char_limit] slices it at 12,000 characters. This hard string slice cuts off mid-line and drops the majority of candidates from the prompt text while parse_semantic_match still validates against all candidates in allowed_ids. Candidate budgeting should fit within the remaining budget after prefix overhead and avoid slicing across candidate entries.
• In src/coding_review_agent_loop/followups.py, when filing_enabled is true in _format_plan_approval_summary_with_followups, reused existing tracker URLs are included in issue_urls and incorrectly listed under Filed future follow-up issues:. Additionally, publication_details (which already contains the publication items, status, and its own Reconciliation: line) is inserted, followed by an unconditional duplicate Reconciliation: line that computes filed count using len(reconciliation.selected_groups). This creates contradictory summary lines whenever follow-ups are reused or cap-skipped (e.g. publication_details reporting 0 filed, 1 reused while the appended line reports 1 filed, 0 skipped by cap).
• PR Use semantic dedupe for approved future follow-up filing #751 adds user-facing CLI flags (--semantic-followup-dedupe, --no-semantic-followup-dedupe, --semantic-followup-backend, --semantic-followup-model, --semantic-followup-timeout-seconds, --semantic-followup-max-calls, --semantic-followup-max-candidates, --semantic-followup-prompt-char-limit) in src/coding_review_agent_loop/cli.py and documents them in docs/local_agent_loop.md, but does not update README.md. Missing documentation for newly added user-facing flags must be updated in README.md.
• Issue Use semantic dedupe for approved future follow-up issue filing #490 acceptance criteria explicitly require: 'In-batch follow-up reconciliation can merge semantically equivalent items even when wording differs substantially.' While tests/test_followups.py tests cross-stage existing issue reuse, there is no test verifying that in-batch reconciliation (_semantic_batch_matcher / reconcile_approved_followups) merges semantically equivalent items in a batch when deterministic keys and similarity thresholds do not match.

Same-PR follow-ups

• Remove dead code in src/coding_review_agent_loop/followups.py: _create_plan_approved_followup_issues (lines 997-1020), _create_approved_followup_issues (lines 1293-1313), and _format_created_followup_issue_summary (lines 1315-1350) are unused after refactoring to _publish_issue_followup_groups.
• Remove unused imports _create_approved_followup_issues and _format_created_followup_issue_summary from src/coding_review_agent_loop/orchestrator.py.

-- Google Antigravity: Gemini 3.8 Flash (High)

[wwind123]

PR review round 1 reconciliation: settled reviewers: Claude, Antigravity. Finalization continues after reconciliation.

[wwind123]

Coder follow-up

Implemented and pushed fixes for all Claude and Antigravity review items on the existing PR branch in commit 786a81c.

Addressed items

• item-1: Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: semantic_dedupe._isolated_provider_config is broken for `--semantic-followup-backend antigravit...
  • Resolution: Fixed Antigravity isolated configuration to use a single antigravity_models entry with antigravity_model unset, and protected temporary-directory cleanup on configuration failures.
• item-2: Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: followups._followup_issue_body accepts the new source_context parameter but never renders it,...
  • Resolution: Rendered FollowupSourceContext lookup context into PR-review issue bodies so parent issue and plan/PR identity remain searchable.
• item-3: Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: Two acceptance-criteria behaviors introduced here are untested. (a) In-batch semantic merge: noth...
  • Resolution: Added focused tests for high-confidence in-batch semantic merging and medium-confidence filing with a possible-duplicate note.
• item-4: Same-PR follow-up from Anthropic Claude: claude-opus-5 (medium), round 1: helpers/skill_runner.py uses inspect.signature(_publish_approved_followups) at call time to d...
  • Resolution: Removed runtime signature introspection and updated the skill test publisher double to accept keyword context.
• item-5: Same-PR follow-up from Anthropic Claude: claude-opus-5 (medium), round 1: helpers/skill_runner._publish_pr_followups builds a FollowupSourceContext with `related_pr_nu...
  • Resolution: Threaded the linked parent issue from skill-mode PR metadata into FollowupSourceContext.
• item-6: Same-PR follow-up from Anthropic Claude: claude-opus-5 (medium), round 1: SemanticDedupeMatcher.match records usage with outcome="succeeded" and sets `validation_statu...
  • Resolution: Changed semantic usage records to become validated only after parsing, and to record invalid_output on contract-validation failure.
• item-7: Same-PR follow-up from Anthropic Claude: claude-opus-5 (medium), round 1: In reconcile_approved_followups, the semantic branch accepts both a group-N string (1-based)...
  • Resolution: Removed the unreachable bare-integer batch identity branch; batch reconciliation now uses one-based group-N identities.
• item-8: Same-PR follow-up from Anthropic Claude: claude-opus-5 (medium), round 1: _create_approved_followup_issues, _create_plan_approved_followup_issues, and `_format_created...
  • Resolution: Removed the unused follow-up issue creation and summary helpers and their orchestrator imports.
• item-9: Same-PR follow-up from Anthropic Claude: claude-opus-5 (medium), round 1: _format_publication_summary is always given the heading "Created approved-review future follow-...
  • Resolution: Adjusted publication headings so reuse-only results do not claim newly created issues.
• item-10: Blocking issue from Google Antigravity: Gemini 3.8 Flash (High), round 1: In [src/coding_review_agent_loop/followups.py](file:///tmp/coding-review-agent-loop/wwind123-co...
  • Resolution: Excluded both the source PR/plan and all parent issue numbers from existing-tracker search results.
• item-11: Blocking issue from Google Antigravity: Gemini 3.8 Flash (High), round 1: In [src/coding_review_agent_loop/semantic_dedupe.py](file:///tmp/coding-review-agent-loop/wwind...
  • Resolution: Reworked prompt budgeting to reserve instruction overhead, pack complete candidate entries, and avoid final hard slicing.
• item-12: Blocking issue from Google Antigravity: Gemini 3.8 Flash (High), round 1: In [src/coding_review_agent_loop/followups.py](file:///tmp/coding-review-agent-loop/wwind123-co...
  • Resolution: Separated filed URLs from reused URLs in plan summaries and removed the duplicate reconciliation line.
• item-13: Blocking issue from Google Antigravity: Gemini 3.8 Flash (High), round 1: PR Use semantic dedupe for approved future follow-up filing #751 adds user-facing CLI flags (--semantic-followup-dedupe, --no-semantic-followup-dedupe...
  • Resolution: Documented the semantic follow-up dedupe flags and bounds in README.md.
• item-14: Blocking issue from Google Antigravity: Gemini 3.8 Flash (High), round 1: Issue Use semantic dedupe for approved future follow-up issue filing #490 acceptance criteria explicitly require: 'In-batch follow-up reconciliation can merge s...
  • Resolution: Added regression coverage proving semantically different wording can merge during in-batch reconciliation.
• item-15: Same-PR follow-up from Google Antigravity: Gemini 3.8 Flash (High), round 1: Remove dead code in [src/coding_review_agent_loop/followups.py](file:///tmp/coding-review-agent...
  • Resolution: Removed the dead follow-up creation and formatting functions from followups.py.
• item-16: Same-PR follow-up from Google Antigravity: Gemini 3.8 Flash (High), round 1: Remove unused imports _create_approved_followup_issues and `_format_created_followup_issue_summ...
  • Resolution: Removed the corresponding unused imports from orchestrator.py.

Remaining items

• None.

Tests run

• python3 -m pytest tests/test_followups.py tests/test_skill_helpers.py -q
• python3 -m pytest tests/test_followups.py tests/test_orchestrator_pr.py -q
• python3 -m pytest tests/test_followups.py tests/test_orchestrator_issue.py tests/test_orchestrator_pr.py tests/test_skill_helpers.py tests/test_split_materialization.py tests/test_decomposition.py -q
• python3 -m pytest tests/test_followups.py tests/test_antigravity.py tests/test_usage.py -q

-- OpenAI Codex: gpt-5.6-luna (xhigh)

[wwind123]

Review verdict: Approved

All sixteen carried items are addressed at 786a81c and each has a regression test. Verified directly: the antigravity isolated config now replaces the model chain instead of setting the legacy single-model field, and mkdtemp is cleaned up on the raise path; prompt packing now reserves the instruction/source/proposed prefix first and renders every candidate inside its own budget with a hard assertion instead of a post-render slice (checked the arithmetic for the 50-candidate/12000-char worst case, which now fits); PR-filed bodies render the lookup context; parent issue numbers are excluded from tracker search; usage records are only marked validated after the strict-JSON parse succeeds and are marked invalid_output otherwise; the unreachable int identity branch, the dead creation/summary helpers and their orchestrator imports are gone; reuse-only publications no longer claim issues were filed and the contradictory duplicate Reconciliation line is removed; the inspect-based signature shim is replaced by a real parent_issue_number parameter threaded from cmd_run_pr_round; and README.md now documents the new flags. New tests cover the in-batch group-N merge, the medium-confidence uncertain filing path, the parent-issue exclusion, prompt budgeting, the antigravity config, and invalid-output usage accounting. GitHub checks are still pending (test in_progress), so I have no global test result for this head; per the review instructions I did not run tests locally, and my verification is by source inspection.

Prior unresolved item dispositions

• [item-1] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: semantic_dedupe._isolated_provider_config is broken for `--semantic-followup-backend antigravit... -> resolved
• [item-2] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: followups._followup_issue_body accepts the new source_context parameter but never renders it,... -> resolved
• [item-3] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: Two acceptance-criteria behaviors introduced here are untested. (a) In-batch semantic merge: noth... -> resolved
• [item-4] Same-PR follow-up from Anthropic Claude: claude-opus-5 (medium), round 1: helpers/skill_runner.py uses inspect.signature(_publish_approved_followups) at call time to d... -> resolved
• [item-5] Same-PR follow-up from Anthropic Claude: claude-opus-5 (medium), round 1: helpers/skill_runner._publish_pr_followups builds a FollowupSourceContext with `related_pr_nu... -> resolved
• [item-6] Same-PR follow-up from Anthropic Claude: claude-opus-5 (medium), round 1: SemanticDedupeMatcher.match records usage with outcome="succeeded" and sets `validation_statu... -> resolved
• [item-7] Same-PR follow-up from Anthropic Claude: claude-opus-5 (medium), round 1: In reconcile_approved_followups, the semantic branch accepts both a group-N string (1-based)... -> resolved
• [item-8] Same-PR follow-up from Anthropic Claude: claude-opus-5 (medium), round 1: _create_approved_followup_issues, _create_plan_approved_followup_issues, and `_format_created... -> resolved
• [item-9] Same-PR follow-up from Anthropic Claude: claude-opus-5 (medium), round 1: _format_publication_summary is always given the heading "Created approved-review future follow-... -> resolved
• [item-10] Blocking issue from Google Antigravity: Gemini 3.8 Flash (High), round 1: In [src/coding_review_agent_loop/followups.py](file:///tmp/coding-review-agent-loop/wwind123-co... -> resolved
• [item-11] Blocking issue from Google Antigravity: Gemini 3.8 Flash (High), round 1: In [src/coding_review_agent_loop/semantic_dedupe.py](file:///tmp/coding-review-agent-loop/wwind... -> resolved
• [item-12] Blocking issue from Google Antigravity: Gemini 3.8 Flash (High), round 1: In [src/coding_review_agent_loop/followups.py](file:///tmp/coding-review-agent-loop/wwind123-co... -> resolved
• [item-13] Blocking issue from Google Antigravity: Gemini 3.8 Flash (High), round 1: PR Use semantic dedupe for approved future follow-up filing #751 adds user-facing CLI flags (--semantic-followup-dedupe, --no-semantic-followup-dedupe... -> resolved
• [item-14] Blocking issue from Google Antigravity: Gemini 3.8 Flash (High), round 1: Issue Use semantic dedupe for approved future follow-up issue filing #490 acceptance criteria explicitly require: 'In-batch follow-up reconciliation can merge s... -> resolved
• [item-15] Same-PR follow-up from Google Antigravity: Gemini 3.8 Flash (High), round 1: Remove dead code in [src/coding_review_agent_loop/followups.py](file:///tmp/coding-review-agent... -> resolved
• [item-16] Same-PR follow-up from Google Antigravity: Gemini 3.8 Flash (High), round 1: Remove unused imports _create_approved_followup_issues and `_format_created_followup_issue_summ... -> resolved

-- Anthropic Claude: claude-opus-5 (medium)

[wwind123]

Review verdict: Approved

All round 1 blocking and same-PR review items (item-1 through item-16) have been resolved with thorough implementation fixes and comprehensive test coverage. GitHub CI checks are currently pending (test check in progress).

Prior unresolved item dispositions

• [item-1] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: semantic_dedupe._isolated_provider_config is broken for `--semantic-followup-backend antigravit... -> resolved
• [item-2] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: followups._followup_issue_body accepts the new source_context parameter but never renders it,... -> resolved
• [item-3] Blocking issue from Anthropic Claude: claude-opus-5 (medium), round 1: Two acceptance-criteria behaviors introduced here are untested. (a) In-batch semantic merge: noth... -> resolved
• [item-4] Same-PR follow-up from Anthropic Claude: claude-opus-5 (medium), round 1: helpers/skill_runner.py uses inspect.signature(_publish_approved_followups) at call time to d... -> resolved
• [item-5] Same-PR follow-up from Anthropic Claude: claude-opus-5 (medium), round 1: helpers/skill_runner._publish_pr_followups builds a FollowupSourceContext with `related_pr_nu... -> resolved
• [item-6] Same-PR follow-up from Anthropic Claude: claude-opus-5 (medium), round 1: SemanticDedupeMatcher.match records usage with outcome="succeeded" and sets `validation_statu... -> resolved
• [item-7] Same-PR follow-up from Anthropic Claude: claude-opus-5 (medium), round 1: In reconcile_approved_followups, the semantic branch accepts both a group-N string (1-based)... -> resolved
• [item-8] Same-PR follow-up from Anthropic Claude: claude-opus-5 (medium), round 1: _create_approved_followup_issues, _create_plan_approved_followup_issues, and `_format_created... -> resolved
• [item-9] Same-PR follow-up from Anthropic Claude: claude-opus-5 (medium), round 1: _format_publication_summary is always given the heading "Created approved-review future follow-... -> resolved
• [item-10] Blocking issue from Google Antigravity: Gemini 3.8 Flash (High), round 1: In [src/coding_review_agent_loop/followups.py](file:///tmp/coding-review-agent-loop/wwind123-co... -> resolved
• [item-11] Blocking issue from Google Antigravity: Gemini 3.8 Flash (High), round 1: In [src/coding_review_agent_loop/semantic_dedupe.py](file:///tmp/coding-review-agent-loop/wwind... -> resolved
• [item-12] Blocking issue from Google Antigravity: Gemini 3.8 Flash (High), round 1: In [src/coding_review_agent_loop/followups.py](file:///tmp/coding-review-agent-loop/wwind123-co... -> resolved
• [item-13] Blocking issue from Google Antigravity: Gemini 3.8 Flash (High), round 1: PR Use semantic dedupe for approved future follow-up filing #751 adds user-facing CLI flags (--semantic-followup-dedupe, --no-semantic-followup-dedupe... -> resolved
• [item-14] Blocking issue from Google Antigravity: Gemini 3.8 Flash (High), round 1: Issue Use semantic dedupe for approved future follow-up issue filing #490 acceptance criteria explicitly require: 'In-batch follow-up reconciliation can merge s... -> resolved
• [item-15] Same-PR follow-up from Google Antigravity: Gemini 3.8 Flash (High), round 1: Remove dead code in [src/coding_review_agent_loop/followups.py](file:///tmp/coding-review-agent... -> resolved
• [item-16] Same-PR follow-up from Google Antigravity: Gemini 3.8 Flash (High), round 1: Remove unused imports _create_approved_followup_issues and `_format_created_followup_issue_summ... -> resolved

-- Google Antigravity: Gemini 3.8 Flash (High)

[wwind123]

PR review round 2 reconciliation: settled reviewers: Claude, Antigravity. Finalization continues after reconciliation.

[wwind123]

Reviewers approved PR #751; watching GitHub checks in the foreground. No coder or reviewer agents will run while checks remain pending.

-- coding-review-agent-loop