babysit-pr: replace legacy fixer loop with one-off implement-phase BRC cycle

[james-in-a-box]

Replace the legacy egg-babysit fixer/reviewer loop with a lean, one-off implement-phase BRC cycle run against an existing PR's diff. Closes #1748.

The legacy egg_babysit package ran a multi-round fixer→reviewer loop that accumulated state across respawns and had no clear termination signal. The new design fires a single concurrent BRC cycle (same five-agent roster as issue-mode) scoped to the PR's diff, with full gateway integration, staging-branch isolation, and MCP skill exposure.

All implement-phase work (phases 1–7) was produced and BRC-consensus-confirmed by five concurrent agents (coder, documenter, tester, reviewer_code, reviewer_contract) in a single cycle.

Changes

Orchestrator (orchestrator/):

• routes/pipelines.py — get_pr_base_branch() / _resolve_base_ref() 3-level fallback; babysit route plumbing (pr-{N} pipeline ID, 400/409 early exits, skip_pr_creation); _verify_pr_head_unchanged() final-push guard
• concurrent_executor.py — BABYSIT-aware orient/review prompts with gh pr checkout step (0); staging-branch egg/babysit-pr/{pr}/{sha}/{role}; per-cycle BRC history ID pr-{N}-{sha}
• models.py + state_store.py — Pipeline.has_contract / pr_head_sha fields; PipelineMode.BABYSIT docstring update (enum string unchanged)
• mcp_tools.py — /babysit-pr MCP skill; POST /api/v1/pipelines + /start, returns {task_id, status, message}
• health_checks/context.py + tier1/phase_output.py — stop hardcoding origin/main

Shared (shared/):

• egg_contracts/agent_roles.py — has_contract: bool param to drop REVIEWER_CONTRACT from babysit roster
• pyproject.toml — remove egg-babysit console script + egg_babysit* package glob
• Deleted 17 modules under shared/egg_babysit/

Docs / skills:

• docs/guides/babysit-pr.md — full rewrite with BRC-cycle flow, D1–F13 contract trace, migration guide
• skills/babysit-pr/SKILL.md — lean MCP skill entry point
• Cross-ref updates in docs/index.md, docs/architecture/orchestrator.md, docs/development/STRUCTURE.md, docs/guides/github-automation.md, docs/guides/sdlc-pipeline.md, shared/README.md

Tests:

• 262 new tests across 11 new files covering all phases 1–7; full suite 4310 passed, ruff clean
• Removed legacy shared/tests/test_egg_babysit/ suite and obsolete integration_tests/test_babysit_pr/test_cli.py

Test Plan

• Automated: pytest 4310 passed, 13 skipped; ruff check + format clean on all touched files
• Known follow-up: test_pr_base_branch.py still imports deprecated get_pr_base_branch shim — NB1 tracked for post-merge cleanup
• Manual: POST /api/v1/pipelines with mode=babysit, pr_number=N, repo=owner/repo should launch a BRC cycle against the PR's diff and return pipeline_id=pr-{N}

Issue: #1748

Authored-by: egg

[james-in-a-box]

Autofix tracking

{"Lint/Python": 1}

[james-in-a-box]

Overseer note — PR created manually (orchestrator post-consensus bug)

BRC consensus was reached at 21:00:48Z with all 5 agents CONFIRMED. The orchestrator's auto_create_pr handler failed to run due to a StateStore path error (/home/egg/repos instead of /home/egg/repos/egg). This PR was opened by the overseer agent as a fallback.

The orchestrator is currently returning HTTP 500 on all pipeline endpoints — the pipeline cannot advance to a terminal state on its own. Once the PR is reviewed/merged, the issue-1748 pipeline may need to be manually marked complete or cleaned up.

— Authored by egg

[egg-reviewer]

No agent-mode design concerns.

The replacement of the legacy egg_babysit fixer/reviewer loop with a standard BRC cycle is a net improvement for agent-mode alignment. Key observations:

• Orient prompts provide setup mechanics, not pre-fetched data. Agents gh pr checkout and read diffs themselves rather than having content baked into prompts.
• Infrastructure constraints stay in infrastructure. Head-move guard (_verify_pr_head_unchanged) is enforced in orchestrator code, not prompt instructions. Branch policy enforced by gateway.
• No post-processing pipelines or structured-output-for-humans anti-patterns. Agents act directly via BRC consensus.
• Producer orientation steps (0–5) are reasonable. They provide necessary git setup mechanics for the babysit context while leaving implementation work unconstrained.

— Authored by egg

[egg-reviewer]

Review: babysit-pr BRC cycle replacement

Good architectural direction — replacing the bespoke fixer/reviewer loop with a standard BRC cycle is the right call. The PR is well-structured and thoroughly tested (262 new tests, legacy suite removed). The following issues need attention before merge.

---

Blocking

B1. branch / base_branch lack input validation — git flag injection risk

orchestrator/routes/pipelines.py lines 682-683, 759-762

source_branch is validated with re.match(r"^[a-zA-Z0-9_./-]+$") (line 691), but branch and base_branch are never validated — neither when supplied directly by the caller (line 682-683) nor when auto-populated from GitHub API data (lines 759-762). These values flow into subprocess.run argument lists (e.g., ["git", "fetch", "origin", branch] at line 4418, ["git", "rev-parse", f"origin/{branch}"] at line 4425) and into agent prompt text.

While subprocess is called with shell=False (good), a branch name starting with -- could be interpreted as a git flag. A value like --upload-pack=malicious in a git fetch argument list is a real vector.

Fix: Apply the same regex validation used for source_branch to branch and base_branch after they are populated from any source (user input or PR state). Add this validation block around line 763 (after the auto-populate from PR state):

for _ref_name, _ref_val in [("branch", branch), ("base_branch", base_branch)]:
    if _ref_val is not None:
        if not re.match(r"^[a-zA-Z0-9_./-]+$", _ref_val) or ".." in _ref_val:
            return make_error_response(
                f"Invalid {_ref_name}: {_ref_val!r}",
                status_code=400,
            )

B2. _verify_pr_head_unchanged fails open — silently bypasses the head-move guard

orchestrator/routes/pipelines.py lines 4393-4450

The head-move guard is the last line of defense against pushing stale work on top of concurrent changes. But it returns (True, None) on any git failure — network timeout, misconfigured remote, corrupted repo. The docstring explicitly states this design: "git/subprocess failures are treated as 'unknown' (returns (True, None))."

This means a transient git fetch failure silently disables the entire safety mechanism. For a guard whose purpose is specifically to prevent data loss, fail-open is the wrong default. A transient fetch failure should retry or block, not silently allow the push.

Fix: Either:

1. Retry the fetch once before giving up (the timeout is already 30s, so one retry is reasonable).
2. Return (False, None) on fetch failure so the caller treats it as "unknown but unsafe" — the HITL escalation is cheap compared to pushing over someone's work.
3. At minimum, use --force-with-lease=<branch>:<expected-sha> on the actual push command so the guard is atomic with the push, eliminating the TOCTOU window entirely.

B3. pr_head_sha has no format validation on the Pydantic model

orchestrator/models.py lines 487-495

pr_head_sha is declared as str | None with no field_validator or pattern constraint. This field is used to construct branch names (egg/babysit-pr/{pr}/{short_sha}/{role} in concurrent_executor.py line 147) and BRC history identifiers (pr-{pr}-{sha[:7]} in pipelines.py line 329). A non-hex value (e.g., containing /, .., or --) could produce path-traversal branch names or unexpected filenames.

Fix: Add a field validator:

@field_validator("pr_head_sha")
@classmethod
def _validate_sha(cls, v: str | None) -> str | None:
    if v is not None and not re.fullmatch(r"[0-9a-f]{7,40}", v):
        raise ValueError("pr_head_sha must be a 7-40 char hex string")
    return v

B4. PipelineMode.BABYSIT docstring incorrectly lists tester as a reviewer

orchestrator/models.py lines 38-42

The docstring says "reviewers (reviewer_code, tester)" — tester is a producer, not a reviewer. The PR's own documentation and the actual agent_roles.py code correctly categorize tester as a producer. Since this PR is modifying this docstring area (added line 33), fix it here.

Fix: Change to "reviewers (reviewer_code)" or "reviewers (reviewer_code only; reviewer_contract is filtered out)".

---

Non-blocking

NB1. repo parameter lacks format validation in MCP handler

orchestrator/mcp_tools.py lines 878-880

The repo parameter is only checked for presence and string type, not for valid owner/name format. It's passed directly to the orchestrator API. While the orchestrator eventually passes it to gh --repo, a regex check at the MCP layer would provide defense-in-depth:

if not re.match(r'^[a-zA-Z0-9._-]+/[a-zA-Z0-9._-]+$', repo):
    return {"error": "repo must be in owner/name format"}

NB2. Statefile/BRC operations run after pipeline marked FAILED

orchestrator/routes/pipelines.py lines ~9745-9781

When the head-move guard trips, phase_failed = True and the pipeline is marked FAILED. But execution continues through _ensure_statefiles_on_branch and _rewrite_brc_history_for_pr. These are wasted work against a failed pipeline and could have unexpected side effects. Consider guarding them with if not phase_failed.

NB3. Duplicated _resolve_base_ref implementations

orchestrator/health_checks/context.py lines 117-132 and orchestrator/health_checks/tier1/phase_output.py lines 198-231

Both files implement _resolve_base_ref with the same 3-step fallback logic but different error handling (the phase_output.py version checks returncode, the context.py version does not). Consider extracting to a shared utility to prevent divergence.

NB4. dag_visualizer.py missing has_contract parameter

dag_visualizer.py line 170

The DAG visualizer calls get_agent_roles() without passing has_contract, so for babysit-pr pipelines it will show REVIEWER_CONTRACT in the graph even though that agent won't be spawned. Cosmetic but misleading.

NB5. Minor grammar in SKILL.md

skills/babysit-pr/SKILL.md line 259

"functional equivalent" should be "functionally equivalent".

---

Observations (no action needed)

• TOCTOU window: Even with B2 fixed, there's an inherent race between the head-move check and the push. The --force-with-lease approach in B2 option 3 would eliminate this entirely.
• _fetch_pr_state fail-open: Returns {} on any gh error, skipping all pre-flight checks. This is acceptable for the pre-flight checks (merged/closed/fork/empty) since downstream agents will surface the issue. The head-move guard (B2) is the one that must not fail open.
• Pipeline ID collision: pr-{pr_number} is not unique across multiple babysit runs against the same PR. The 409 conflict check mitigates this for concurrent runs, and the BRC history ID includes the SHA for historical separation. Acceptable.
• Test coverage is strong: 262 new tests across 11 files covering pipeline creation, staging branches, orient prompts, BRC history IDs, health checks, and the MCP skill. Good coverage of the new code paths.

— Authored by egg

[james-in-a-box]

Contract Verification: All Acceptance Criteria Verified

Comprehensive contract verification of all 8 phases (28 tasks) against the plan in .egg-state/drafts/1748-plan.md. All acceptance criteria are met.

Phase 1 — Foundation: PR base-branch parameterization (TASK-1-1 to 1-4)

Task	Status	Evidence
TASK-1-1	Verified	get_pr_base_branch() at pipelines.py:4292 — type-annotated, docstring describes 3-level fallback (gh pr view → _detect_default_branch() → "main"). Returns bare branch name.
TASK-1-2	Verified	grep -n 'origin/main' orchestrator/routes/pipelines.py returns only docstrings/comments and _detect_default_branch() internals. _resolve_origin_ref() wired at 7+ call sites.
TASK-1-3	Verified	Both context.py and phase_output.py implement _resolve_base_ref() reading pipeline.base_branch first, probing origin/HEAD second, falling back to origin/main last.
TASK-1-4	Verified	3 test files exist: test_pr_base_branch.py (11 tests), test_pipelines_origin_main_parameterization.py (18 tests), test_health_checks_base_ref.py (29 tests). All pass.

Phase 2 — Pipeline.has_contract + roster filter (TASK-2-1 to 2-4)

Task	Status	Evidence
TASK-2-1	Verified	Pipeline.has_contract at models.py:496, default True. PipelineMode.BABYSIT docstring (lines 33-42) describes repurposed meaning.
TASK-2-2	Verified	get_roles_for_phase() at agent_roles.py:1002 accepts has_contract: bool = True; filters REVIEWER_CONTRACT when False (lines 1038-1041).
TASK-2-3	Verified	Call sites at pipelines.py:7362 and concurrent_executor.py:110 pass has_contract=getattr(pipeline, "has_contract", True).
TASK-2-4	Verified	test_agent_roles_has_contract.py (14 tests) and test_pipeline_has_contract_field.py (11 tests). All pass.

Phase 3 — Mode-aware orientation prompts (TASK-3-1 to 3-4)

Task	Status	Evidence
TASK-3-1	Verified	_build_reviewer_preparation() (lines 6368-6519) has BABYSIT branch emitting "read the PR diff" with resolved base ref via _resolve_origin_ref().
TASK-3-2	Verified	_build_producer_orientation() (lines 6521-6674) BABYSIT branch contains "rebase", "your role's scope", "do not refactor outside the diff", and "conflict_resolver".
TASK-3-3	Verified	Both builders accept mode and pr_number parameters; _build_agent_prompt() threads them through at lines 6150-6151 and 6184-6185.
TASK-3-4	Verified	test_orient_prompts_babysit_pr.py — 47 tests covering babysit text, issue-mode regression, base-branch interpolation. All pass.

Phase 4 — Staging-branch isolation + route plumbing (TASK-4-1 to 4-4)

Task	Status	Evidence
TASK-4-1	Verified	get_worktree_branch() in concurrent_executor.py:114-158 generates egg/babysit-pr/{pr}/{sha}/{role} for BABYSIT mode. Issue-mode unchanged.
TASK-4-2	Verified	Pipeline-creation route (lines 649-948): 400 on missing pr_number, fork, merged/closed; 409 on empty diff and duplicate pr-{N}; 201 on happy path. No PR comments on error paths.
TASK-4-3	Verified	Final-push consolidates staging branches to PR head. BRC history records multiple proposals.
TASK-4-4	Verified	test_concurrent_executor_staging_branch.py (15 tests) and test_pipeline_creation_babysit_pr.py (27 tests). All pass.

Phase 5 — BRC-history identifier + final-push guard (TASK-5-1 to 5-3)

Task	Status	Evidence
TASK-5-1	Verified	_brc_history_identifier() at pipelines.py:308-333 returns pr-{N}-{short_sha} for babysit mode. Called from _rewrite_brc_history_for_pr() at line 9770.
TASK-5-2	Verified	_verify_pr_head_unchanged() at pipelines.py:4393-4450 compares stored SHA vs remote HEAD. On mismatch: HITL escalation with both SHAs and recovery suggestion (lines 9721-9727). Pipeline marked FAILED.
TASK-5-3	Verified	test_brc_history_identifier_babysit_pr.py (16 tests) and test_final_push_head_move_guard.py (23 tests). All pass.

Phase 6 — babysit-pr MCP skill (TASK-6-1 to 6-2)

Task	Status	Evidence
TASK-6-1	Verified	skills/babysit-pr/SKILL.md exists (268 lines) with metadata header (name, description, argument-hint). Confirm-Submit-Watch flow with POST to /api/v1/pipelines. Deprecation note for egg-babysit CLI present.
TASK-6-2	Verified	integration_tests/test_babysit_pr/test_skill.py (266 lines) covers argument validation, happy-path POST, 409 duplicate, 400 early-exits (fork, merged, closed, empty diff), tool registration. All pass.

Phase 7 — Remove legacy shared/egg_babysit/ (TASK-7-1 to 7-5)

Task	Status	Evidence
TASK-7-1	Verified	shared/egg_babysit/ directory does not exist. orchestrator/tests/test_babysit_pipeline_creation.py deleted.
TASK-7-2	Verified	grep -n 'babysit' shared/pyproject.toml returns no matches. Console script and package glob removed.
TASK-7-3	Verified	shared/tests/test_egg_babysit/ directory does not exist. All 12 legacy test files removed.
TASK-7-4	Verified	integration_tests/test_babysit_pr/test_cli.py deleted. Rewritten suite: test_pipeline.py, test_gateway.py, test_escalation.py, test_skill.py, conftest.py. 40 tests pass.
TASK-7-5	Verified	shared/egg_babysit/README.md does not exist.

Phase 8 — Documentation refresh (TASK-8-1 to 8-2)

Task	Status	Evidence
TASK-8-1	Verified	docs/guides/babysit-pr.md (391 lines) documents new BRC flow end-to-end. "What Changed" migration section present. Contract/decision trace appendix covers all 13 items (D1-D7, F8-F13) with section back-references.
TASK-8-2	Verified	grep -rn 'egg-babysit' docs/ returns matches only in migration/deprecation sections of babysit-pr.md. Cross-references in github-automation.md, sdlc-pipeline.md, index.md, orchestrator.md, STRUCTURE.md updated.

Test Results

251 tests passed, 0 failed across all 11 new/rewritten test files:

Test File	Passed
test_pr_base_branch.py	11
test_pipelines_origin_main_parameterization.py	18
test_health_checks_base_ref.py	29
test_agent_roles_has_contract.py	14
test_pipeline_has_contract_field.py	11
test_orient_prompts_babysit_pr.py	47
test_concurrent_executor_staging_branch.py	15
test_pipeline_creation_babysit_pr.py	27
test_brc_history_identifier_babysit_pr.py	16
test_final_push_head_move_guard.py	23
integration_tests/test_babysit_pr/	40

Decision Trace Compliance

All 13 resolved HITL decisions from .egg-state/contracts/1748.json decision-2 are implemented and documented:

• D1 (PipelineMode swap): BABYSIT enum string unchanged, docstring updated
• D2 (has_contract): New field + roster filter
• D3 (lean skill): Single Confirm-Submit-Watch flow
• D4 (mid-cycle commits): Final-push guard with HITL escalation
• D5 (conflict_resolver): On-demand mention in producer prompt
• D6 (base-branch sweep): Full parameterization across all call sites
• D7 (inline branch): No orient-builder refactor
• F8 (reviewer pre-filters): Only reviewer_contract filtered
• F9 (fork-PR UX): Fail-fast stderr, no PR comment
• F10 (concurrency): 409 on duplicate pr-{N}
• F11 (scope guardrails): Soft prompt hint
• F12 (BRC-history ID): pr-{N}-{short-sha} content-addressed
• F13 (CLI removal): egg-babysit removed, no deprecation shim

Known Follow-up

PR description notes: test_pr_base_branch.py still imports deprecated get_pr_base_branch shim — NB1 tracked for post-merge cleanup. This is non-blocking.

Verdict

All 28 tasks across 8 phases meet their acceptance criteria. All 251 new tests pass. No contract violations found.

— Authored by egg

[james-in-a-box]

Addressed all blocking items and most non-blocking items in e27bc8e:

Blocking — all fixed

• B1 — Added branch/base_branch regex validation (same pattern as source_branch) after auto-populate from PR state. Rejects -- prefixed values and .. path traversal.
• B2 — Changed _verify_pr_head_unchanged from fail-open to fail-closed with one retry. Transient git failures now return (False, None) → HITL escalation instead of silently allowing the push. Updated docstring and 7 existing tests; added 2 new retry-success tests.
• B3 — Added field_validator for pr_head_sha enforcing [0-9a-f]{7,40} hex format. Empty strings are normalized to None. Updated 3 test files that used non-hex placeholder values to use SimpleNamespace instead (matching the existing pattern for bypassing validation in edge-case tests).
• B4 — Fixed docstring: removed tester from the reviewer list. Tester is a producer.

Non-blocking — fixed

• NB1 — Added owner/name regex validation for repo param in the MCP handler.
• NB2 — Wrapped _ensure_statefiles_on_branch, _cleanup_agent_outputs_for_pr, and _rewrite_brc_history_for_pr inside if not phase_failed guard.
• NB5 — Fixed "functional" → "functionally" in SKILL.md.

Non-blocking — deferred

• NB3 (duplicated _resolve_base_ref): Agree this should be extracted to a shared utility, but it's a refactor that touches health check internals beyond this PR's scope. Better as a follow-up.
• NB4 (dag_visualizer missing has_contract): Cosmetic — the DAG graph is an internal debugging tool. Will address in a follow-up.

— Authored by egg

[egg-reviewer]

No agent-mode design concerns in the delta since 4db59bca.

The new commit (e27bc8e) is entirely orchestrator-side hardening: input validation for branch names and repo params, fail-closed semantics for the head-move guard, SHA format validation, and corresponding test updates. None of these changes affect agent prompts, pre-fetching, or agent autonomy.

— Authored by egg

[james-in-a-box]

Contract Re-verification: APPROVED

Delta reviewed: 7e6293f..e27bc8e — "Address review feedback: input validation, fail-closed guard, test updates"

All blocking issues from prior review are resolved:

Blocking (B) Items — Resolved

• B1 (branch validation): branch and base_branch now validated with the same regex as source_branch — rejects git flag injection (--upload-pack=...) and path-traversal (..).
• B2 (fail-closed guard): _verify_pr_head_unchanged now returns (False, None) on transient git failures after retry, escalating via HITL rather than silently allowing push. Docstring and implementation aligned.
• B3 (SHA field validator): pr_head_sha field has a field_validator enforcing 7–40 char hex format via re.fullmatch(r"[0-9a-f]{7,40}", v).
• B4 (BABYSIT docstring): Correctly lists tester as a producer, not a reviewer.

Non-blocking (NB) Items — Resolved

• NB1: MCP handler validates repo param with ^[a-zA-Z0-9._-]+/[a-zA-Z0-9._-]+$ regex.
• NB2: Statefile/BRC operations guarded behind if not phase_failed check (line 9798).
• NB5: "functional" → "functionally" fixed in SKILL.md.

Test verification

All 315 new/modified tests pass:

• test_final_push_head_move_guard.py — 25 pass (fail-closed semantics verified)
• test_brc_history_identifier_babysit_pr.py — 16 pass
• test_pipeline_has_contract_field.py — 11 pass
• test_pipeline_creation_babysit_pr.py — 23 pass
• test_pr_base_branch.py, test_pipelines_origin_main_parameterization.py, test_health_checks_base_ref.py, test_mcp_tools.py, test_orient_prompts_babysit_pr.py, test_concurrent_executor_staging_branch.py — 240 pass

Contract compliance

No regressions to previously verified criteria. The origin/main hardcoding sweep is complete — remaining references in health checks are fallback-path only (correct behaviour when no base branch is resolvable).

— Authored by egg

[egg-reviewer]

Re-review: e27bc8e (delta from 4db59bc)

All four blocking issues and three of five non-blocking items from the previous review have been addressed. Verified each fix against the surrounding code and test coverage.

Blocking items — all resolved

• B1 (branch/base_branch validation): Fixed correctly. Validation at lines 764-773 runs after auto-population from PR state, using the same regex as source_branch. Both user-supplied and GitHub API-derived values are covered.
• B2 (fail-open → fail-closed with retry): Fixed correctly. The retry loop (lines 4429-4489) retries once on any failure path (fetch failure, rev-parse failure, exception, empty output), then returns (False, None) on the second failure. Docstring updated to match. Tests updated to supply mock side effects for both attempts, and two new retry-success tests (test_rev_parse_succeeds_on_retry, test_exception_on_first_attempt_succeeds_on_retry) verify the happy retry path.
• B3 (pr_head_sha field validator): Fixed correctly. Validator at lines 498-505 enforces [0-9a-f]{7,40} hex format and normalizes empty strings to None. Tests that previously used non-hex placeholder SHAs ("short", "") correctly switched to SimpleNamespace to bypass validation while still testing the downstream fallback code paths.
• B4 (docstring misclassifying tester): Fixed. Line 39 now reads "reviewers (reviewer_code)" only.

Non-blocking items — resolved

• NB1 (repo format validation in MCP): Added at mcp_tools.py:879-880.
• NB2 (statefiles/BRC after pipeline FAILED): Wrapped in if not phase_failed guard at lines 9798-9828. Clean implementation.
• NB5 (grammar in SKILL.md): Fixed.

Non-blocking items — deferred (acceptable)

• NB3 (duplicated _resolve_base_ref): Agreed — extracting a shared utility is a separate refactor.
• NB4 (dag_visualizer missing has_contract): Cosmetic, fine as follow-up.

New observations (non-blocking)

NB6. The retry loop in _verify_pr_head_unchanged re-runs git fetch on rev-parse failure (the continue at line 4480 restarts from the top of the loop). This is harmless — a redundant fetch is cheap — but a future refactor could separate the fetch retry from the rev-parse retry for clarity. Also, there's no backoff between retries, which means both attempts fire back-to-back. For a 30-second-timeout fetch, this is unlikely to matter in practice.

NB7. test_large_pr_number_with_full_sha (line 63 of test_brc_history_identifier_babysit_pr.py) was corrected from 41 chars to 40 chars to comply with the new validator — good catch, no issues.

No new blocking issues found. All previous concerns have been properly addressed, not just superficially patched.

— Authored by egg

[james-in-a-box]

egg is addressing review feedback...

[james-in-a-box]

Thanks for the thorough re-review. Acknowledged NB6 (no-backoff retry) and NB7 (SHA length fix) — agree both are fine as-is for the reasons noted. NB6 can be addressed in a future cleanup if the retry logic is ever revisited.

— Authored by egg

[james-in-a-box]

egg feedback addressed. View run logs

10 previous review(s) hidden.