Triage from dev

[jinhongkuan]

Summary

Triage release per DEV_CYCLE.md §10.5. Forwards a curated v0 subset of dev to main between full releases. v1 architecture (Layer A governance, Layer B CodeGenome, semantic-status pre-classifier, HITL bypass, LLM drift judge — issues #44, #60, #61, #109, #110, #112) is intentionally held back per §10.5.1 eligibility ("not triage-eligible: schema-migrating changes, breaking public-API changes, multi-PR feature epics").

All five P0 bugfixes are on this branch.

Linked issues

Closes — P0 bugfixes (auto-close on merge to main):

• Closes [P0] Preflight skill does not instruct agent to capture refinements when user prompt contradicts surfaced decisions #154 — preflight Step 5.6 (capture refinements when prompt contradicts surfaced decision)
• Closes fix(skill): preflight does not auto-fire on natural refactor prompts in headless Claude Code sessions #146 — preflight auto-fire on natural refactor prompts in headless Claude Code
• Closes feat(skill): session-end auto-capture of uningested decisions — research + observable validation #147 — SessionEnd auto-capture of uningested decisions (research + observable validation)
• Closes post-commit hook syncs drift but does not auto-resolve — pending-compliance state accumulates for out-of-session committers #135 — post-commit hook syncs drift but does not auto-resolve (dashboard tooltip nudge)
• Closes [P0] Telemetry Layer 1: local-only tool usage counters (privacy-first) #39 — local telemetry counters + first-boot consent

Refs — supporting work:

• Refs [P1] SessionEnd capture-corrections hook is silently broken — design pivot to next-session surfacing #156 — Flow 4 advisory pivot (design pivot, not Flow 2a)
• Refs Add deterministic escalation policy engine for semantic drift outcomes #108 — v0 user-flow e2e CI (multi-commit harness landing)
• Refs docs: development cycle reference + demos/guides/training scaffolding #93 — DEV_CYCLE.md spec
• Refs chore: CI Phase 1 — Windows matrix + ruff/mypy + secret scan + merged-to-dev labeller #102 — CI Phase 1 (Windows matrix + Tier-1 gates)

P0 status (verified)

#	Issue	Triage commit
#154	preflight Step 5.6	c95c6a8
#147	SessionEnd capture-corrections	cf48270 + 17907fb + 82a493e
#146	preflight auto-fire	aa74510
#135	dashboard tooltip nudge	667a3b9
#39	local telemetry counters	eba9812 (shipped in v0.13.5)

Triage commits (per §10.5.4)

Curated v0 subset (this release's payload)

triage SHA	dev SHA	issue/PR	subject
0b79e35	(bulk copy)	#108, #93, #102	feat: carry over v0 CI workflows + demo recording + dev-cycle docs from dev
e7323c8	d8ac94d (PR #164)	—	test(e2e): rewrite demo flow prompts in realistic per-role voice
c95c6a8	51e631d (PR #163)	#154	fix(skill): capture refinements when prompt contradicts surfaced decision
e72a418	48a0e92	—	refactor(e2e): single source of truth for harness + recording setup
975dc83	cd9b7d2	#156	test(e2e): point Flow 4 advisory at #156 (design pivot) instead of #154
82a493e	17923b6	#147	test(e2e): bootstrap .bicameral/ + pass --mcp-config to SessionEnd subprocess
17907fb	8af60f3	#147	test(e2e): add Flow 4 path-X-(b) ledger validation
cf48270	d76b419	#147	fix(hooks): SessionEnd hook drift — re-entrancy guard + --auto-ingest
e961cad	87b996b	—	style: ruff format tests/e2e/run_e2e_flows.py
f97ddab	5e8f7c0	—	test(e2e): split Flow 2 into auto-fire (Flow 2) + correction-capture loop (Flow 2a)
697dc6e	e3250cf	—	fix(hook): emit hookSpecificOutput envelope so additionalContext reaches model
a50d723	daf9e49	—	fix(e2e): materialize UserPromptSubmit hook into test target settings
d014299	80c4219	—	style: ruff format scripts/hooks/preflight_intent.py
c5c86f7	79927c7	—	fix(setup): install preflight UserPromptSubmit hook for end users
aa74510	ca02b68	#146	fix(skill): resolve preflight auto-fire failure on natural refactor prompts

Pre-§10.5 carry-over (sunk-cost SHAs — content already on main via the v0.13.6 release path)

These 5 commits exist on triage-from-dev with different SHAs from the matching commits on main. Per §10.5.3, the lane is published — history is not rewritten; the audit trail re-converges going forward.

triage SHA	main SHA (equivalent)	issue/PR	subject
ad3e440	f6695c6	#135, #108	chore: bump to v0.13.6 — triage release
6163002	29846d6	#108	fix: portable repo-root resolution in sim_issue_108_flows.py
78b6c09	430a1b1	#108	style: ruff format scripts/sim_issue_108_flows.py + docstring sync
aebd94b	e651233	#108	feat: end-to-end sim + capture-corrections skill correction
667a3b9	7b17e74	#135	feat: dashboard tooltip nudges out-of-session committers to /bicameral-sync

Eligibility (per §10.5.1)

Each new-payload commit is small, self-contained, and lands one of: bug fix on a supported workflow (#146, #147, #154), test/e2e harness substrate (#108, #156), CI infrastructure (#102), demo recording for the v0 user flow, or documentation reference (#93). No schema migrations, no breaking public-API changes.

The bulk-copy commit (0b79e35) carries 15 files (CI workflows + demo recording scripts + DEV_CYCLE.md docs) directly from origin/dev rather than via cherry-pick -x. This is an explicit §10.5.3 adaptation: the underlying dev commits diverged substantially from triage's pre-§10.5 SHAs and prior triage-adapt: workarounds (preflight_telemetry imports, schema migrations gated on codegenome). A clean snapshot of each file faithfully preserves the v0 content without fighting SHA history. The commit body lists every file copied with provenance.

Diverged-surface conflicts during the cherry-pick portion (tests/e2e/run_e2e_flows.py, tests/e2e/record_demo_interactive.sh, the five demo-prompt files) were resolved per §10.5.3's adaptation clause — accepting the cherry-picked content where the file simply hadn't yet landed on this branch's line. No new logic was invented.

Held back from this triage release

Issue / area	Reason
#44 LLM drift judge	v1 Layer A Phase 2
#60 CodeGenome Phase 3 (continuity)	v1 Layer B
#61 CodeGenome Phase 4 (semantic drift)	v1 Layer B + Layer A Phase 1
#109, #110 governance contracts + escalation engine	v1 Layer A core
#112 preflight HITL bypass flow	v1 Layer A bypass
#111 governance architecture docs	v1 docs
#65 preflight telemetry capture loop	depends on v1 escalation feedback; existing triage-adapt: markers on triage explicitly skip its imports
#102 CI Phase 1 codegenome refactor	bundled with CI Phase 1 (CI portion carried over via 0b79e35; codegenome refactor portion held)
#76 dashboard decision_level surfacing	deferred (separate review)
#77 decision_level classifier + CLI	deferred (separate review)
#48 pre-push drift hook + branch-scan CLI	deferred (separate review)
#49 sticky PR-comment drift report	deferred (#966cdcc partial revert flagged need for review)
#97 event vocabulary extension	deferred (separate review)

Plan / Audit / Seal

Triage release roll-up — per §10.5.4, individual Plan/Audit/Seal references live on the upstream commits / PRs:

• fix(skill): capture refinements when prompt contradicts surfaced decision (#154) #163 (Step 5.6) — single-skill change; risk:L2; trace via skills/bicameral-preflight/SKILL.md diff
• test(e2e): rewrite demo flow prompts in realistic per-role voice #164 (demo prompts) — test-data change; risk:L1; trace via tests/e2e/prompts/flow-*.md diff
• feat(skill): session-end auto-capture of uningested decisions — research + observable validation #147 (SessionEnd capture-corrections) — multi-commit; trace via the linked issue
• fix(skill): preflight does not auto-fire on natural refactor prompts in headless Claude Code sessions #146 (preflight auto-fire) — single-skill change
• Add deterministic escalation policy engine for semantic drift outcomes #108 (v0 user-flow e2e CI) — multi-commit harness landing; trace via the linked issue
• post-commit hook syncs drift but does not auto-resolve — pending-compliance state accumulates for out-of-session committers #135 (dashboard tooltip) — dashboard-only; risk:L1
• chore: CI Phase 1 — Windows matrix + ruff/mypy + secret scan + merged-to-dev labeller #102 (CI Phase 1) — workflow-only portion carried via 0b79e35; risk:L1

META_LEDGER does not gain a new entry for the triage release itself — the chain advances on the upstream feature PRs.

Test plan

Tier 2 release gates per §4.5.2:

• All Tier 1 gates green on this PR (lint, mypy, regression Linux + Windows, schema persistence, module imports, secret scan, pip check)
• Full regression including slow markers — pytest -m "not bench"
• Preflight eval — drift precision must not regress vs main's last successful run
• Schema migration validation against persistent DB with seed data (no row loss; roundtrip works)
• Performance regression check (drift detection p50, ingest throughput, search latency — fail if > 15% regression)
• Security scan (bandit, pip-audit, GitHub Dependency Review)
• CHANGELOG enforcement — ## Unreleased content moved under a new ## [v0.13.7] block before merge
• Version monotonicity — pyproject.toml version > v0.13.6 (current main tag)
• MCP protocol live smoke (spawn server, exercise each tool over stdio, assert response shape)
• e2e flow assertions PASS against the pinned desktop/desktop commit (Flow 2a should now PASS with fix(skill): capture refinements when prompt contradicts surfaced decision (#154) #163's Step 5.6 landed)
• v0 user-flow e2e CI workflow (v0-user-flow-e2e.yml) is now wired — confirm at least one assertion run completes green on this PR

Pre-merge checklist

Before this PR can satisfy the §10.5.4 release-PR contract:

• Title rename to release: v0.13.7 (triage) (current "Triage from dev" is non-compliant with §10.5.4 + §4.2)
• flow:release label applied (mandatory per §4.1.1)
• Version bump commit — pyproject.toml 0.13.6 → 0.13.7 + CHANGELOG.md ## Unreleased → ## [v0.13.7] block
• After merge to main: tag v0.13.7, publish GitHub Release, sync main back to dev per §10

Summary by CodeRabbit

• New Features

  • Added contextual hook reminders during code work: preflight prompts, post-commit sync cues, and collision-capture flow.
  • Expanded decision discovery via code-dependency graph to surface related decisions across imports.
  • Improved contradiction-capture workflow with user-guided resolution (supersede/keep both/unrelated).

• Tests

  • Added comprehensive end-to-end test suite for canonical user workflows with Claude Code CLI.
  • Added unit and integration test coverage for hooks, intent classification, and graph expansion.

• Documentation

  • Documented development cycle, demos, feature guides, and training concepts.

• Chores

  • Version bumped to 0.13.6.
  • Added GitHub Actions workflows for CI/CD (lint, type-check, secret scan, e2e validation).

[coderabbitai]

Note

Reviews paused

It looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the reviews.auto_review.auto_pause_after_reviewed_commits setting.

Use the following commands to manage reviews:

• @coderabbitai resume to resume automatic reviews.
• @coderabbitai review to trigger a single review.

Use the checkboxes below for quick actions:

• ▶️ Resume reviews
• 🔍 Trigger review

📝 Walkthrough

Walkthrough

This PR implements a comprehensive preflight gating and contradiction-capture system, adds code-graph expansion for preflight region anchoring, updates hook infrastructure with a re-entrancy guard on SessionEnd, and introduces a complete end-to-end test harness for five canonical user flows. It also documents the repository's development cycle, demo/training/guide requirements, and includes updated CI workflows.

Changes

Preflight Gating, Reminder System, and Hook Infrastructure

Layer / File(s)	Summary
Intent Classifier scripts/hooks/preflight_intent.py	Deterministic regex-based classifier with canonical IMPLEMENTATION_VERBS, INDIRECT_INTENT_PHRASES, and SKIP_PATTERNS that decides when to fire preflight based on prompt content.
UserPromptSubmit Hook Script scripts/hooks/preflight_reminder.py	Hook reads stdin JSON, calls should_fire_preflight(prompt), and emits hookSpecificOutput with system reminder to run bicameral.preflight before write operations when prompted intent is detected.
Hook Command Setup & Wiring setup_wizard.py, .claude/settings.json	Install/update UserPromptSubmit hook in Claude settings; filter stale preflight entries before adding the deterministic reminder command via _BICAMERAL_PREFLIGHT_REMINDER_COMMAND.
Tests tests/test_preflight_intent.py, tests/test_preflight_hook.py	Unit tests for intent classifier (verb/phrase matching, skip patterns, empty/whitespace input) and subprocess tests for hook contract (JSON envelope shape, hookSpecificOutput, idempotency).

Post-Preflight Contradiction Capture & Refinement Flow

Layer / File(s)	Summary
Collision Capture Hook Script scripts/hooks/post_preflight_capture_reminder.py	PostToolUse hook for mcp__bicameral__bicameral_preflight that emits a <system-reminder> instructing the agent to ask the user (via AskUserQuestion) to select supersede/keep_both/unrelated, then mechanically call bicameral.ingest(source=agent_session) and bicameral.resolve_collision(...) based on the choice.
Skill Refinement Documentation skills/bicameral-preflight/SKILL.md	Documents the new contradiction-refinement flow: user-prompted AskUserQuestion with three-way judgment, followed by mechanical ingest and collision resolution; adds PostToolUse hook reinforcement for post-preflight context injection.
Hook Installation setup_wizard.py, .claude/settings.json	Install/update PostToolUse hook scoped to _BICAMERAL_PREFLIGHT_TOOL_NAME; filter stale bicameral/preflight entries and add collision-capture command via _BICAMERAL_COLLISION_CAPTURE_REMINDER_COMMAND.
Tests tests/test_post_preflight_capture_hook.py	Subprocess tests validating hook emits envelope when fired=True and decisions non-empty, routes judgment to user (not agent), handles tool_response as dict or JSON string, and is silent on mismatches/errors.

Code Graph Expansion for Preflight Region Anchoring

Layer / File(s)	Summary
Code Locator Adapter Expansion APIs adapters/code_locator.py	Add expand_file_paths_via_graph(file_paths, hops=1) to walk imports-only 1-hop ego graph around symbols; add neighbors_for(file_path, start_line, end_line) to resolve symbol span and return sorted neighbor addresses; store loaded _config for use by expansion logic.
Preflight Handler Integration handlers/preflight.py	Update _region_anchored_preflight to conditionally expand caller-supplied file_paths via code-graph, track surfaced_via_expansion, adjust confidence to 0.7 for non-direct matches, and return (matches, expanded_flag); update handle_preflight to append "graph" to sources_chained when expansion contributes.
Eval Dataset & Runner Updates tests/eval/preflight_dataset.jsonl, tests/eval/run_preflight_eval.py	Update M6 test case to pin decision to dependency and include graph_neighbors mapping; enhance _apply_setup to mock path-aware get_decisions_for_files via region_decisions_pinned_to and conditionally attach ctx.code_graph with graph_neighbors-driven expansion.
Documentation docs/preflight-failure-scenarios.md	Mark M6 (Transitive) scenario as closed; document graph-based expansion, reduced confidence (0.7), and sources_chained metadata.
Tests tests/test_preflight_graph_expansion.py	Unit tests for expander (1-hop inclusion, input preservation, empty input, hub-cap enforcement, imports-only filtering, uninitialized fallback) and integration tests (decision surfaces via graph expansion, sources_chained tagging).

Post-Commit Sync Reminder & SessionEnd Hook Re-entrancy

Layer / File(s)	Summary
Post-Commit Sync Script scripts/hooks/post_commit_sync_reminder.py	PostToolUse hook for Bash tool that detects git write-ops (commit, merge, pull, rebase --continue) and emits reminder envelope instructing agent to run /bicameral:sync.
Hook Command Updates setup_wizard.py, .claude/settings.json	Move PostToolUse/Bash bicameral reminder from inline python3 -c one-liner to python3 scripts/hooks/post_commit_sync_reminder.py; add _build_session_end_command(mcp_config_path) helper to construct SessionEnd command with optional MCP config flags; update SessionEnd to include re-entrancy guard (BICAMERAL_SESSION_END_RUNNING env var) and --auto-ingest flag.
Tests tests/test_post_commit_sync_hook.py, tests/test_session_end_hook_drift.py	Subprocess tests for post-commit hook (git write-op detection, silent on read-only/non-git, malformed stdin, idempotency); drift tests for SessionEnd command (re-entrancy guard presence, --auto-ingest flag, setup_wizard canonical-command matching, MCP config injection).

End-to-End Test Infrastructure and Demo Recording

Layer / File(s)	Summary
Shared Harness Setup tests/e2e/_harness_setup.py	Materialize MCP config from template via ${DESKTOP_REPO_PATH}/${LEDGER_DIR} substitution; generate hook-wired Claude settings via setup_wizard; clean ledger, reset desktop repo, bootstrap .bicameral/ config; provide setup_all() and CLI entry point for artifact materialization.
Ledger Snapshot Helpers tests/e2e/_ledger_helpers.py	Pure helper count_agent_session_decisions(snapshot) that counts decisions with source_type="agent_session", returning None on error snapshots.
Flow Orchestration tests/e2e/run_e2e_flows.py	Validate local prerequisites, bootstrap MCP/settings, define FlowSpec/FlowResult dataclasses, run 5 flows via claude -p with stream-json capture (or re-grade/skip per plan), extract tool_use calls, apply per-flow assertions (assert_flow_1–assert_flow_5), inject scaffolding for recovery, perform post-hoc ledger validation (Flow 3 lifecycle via status changes; Flow 4 via agent_session decision count), and print summary table with advisories.
Interactive Demo Recording tests/e2e/record_demo_interactive.sh	Run 5 scenes in parallel tmux sessions with ffmpeg capture, poll for dashboard port, refresh Chromium dashboard per-scene, record start/end timestamps, trim continuous MP4 into per-scene outputs, generate transition slide, concatenate into PM and Dev split-screen MP4s.
Demo Rendering & Artifact tests/e2e/demo_renderer.py, tests/e2e/record_demo.sh	Render NDJSON stream-json to human-readable transcript while recording scene boundaries; orchestrate xterm-based Claude session with Chromium dashboard polling, ffmpeg screen capture, and post-processing into split-screen MP4s.
Flow Prompts & Assertions tests/e2e/prompts/flow-1-ingest.md, flow-2-preflight.md, flow-3-commit-sync.md, flow-4-session-end.md, flow-5-history.md; tests/e2e/prompts/composite-demo.md	Define canonical user flow prompts for ingest, preflight with preflight-driven edit, commit linking, session-end resolution, and ledger history; composite demo prompt for PM/Dev/PM three-scene split-screen workflow.
MCP Configuration tests/e2e/bicameral.mcp.json	Template MCP config for e2e environment with bicameral-mcp command and environment variables (SURREAL_URL, REPO_PATH).
Documentation & Unit Tests tests/e2e/README.md, tests/test_flow4_ledger_validation.py, tests/test_e2e_asserters.py	Suite documentation (flow definitions, session structure, local setup, CI integration, spec-change guidance); Flow 4 ledger validation unit tests; Flow 1 asserter unit tests covering file-anchor variations and failure modes.

Release, Documentation, CI, and Configuration

Layer / File(s)	Summary
Development Cycle Contract docs/DEV_CYCLE.md	Comprehensive 1177-line document defining repo topology (dev/main branches), feature release workflow phases (friction → spec → harness → solution → telemetry → optimization), issue/PR conventions, CI gate tiers, merge strategy, release process, CHANGELOG structure, skill file rule, hotfix path, triage lane mechanics, roles, demo requirements, and decision guides.
Demo Documentation docs/demos/README.md, docs/demos/v0-userflow-e2e.md	Define demo purpose, authoring rules, index; document v0-flow demo (split-screen recording, scene structure, tool visibility, artifact access, recording steps, split logic).
User & Training Guides docs/guides/README.md, docs/training/README.md	Templates and authoring rules for user feature guides and long-form training docs with required sections and release-process constraints.
Version & Script Entries pyproject.toml	Bump version 0.13.5 → 0.13.6; add script entry points for bicameral-mcp-preflight-reminder, bicameral-mcp-post-commit-sync-reminder, bicameral-mcp-collision-capture-reminder; expand wheel exclude, add [tool.ruff], [tool.ruff.lint], and [tool.mypy] configuration.
CI Workflows .github/workflows/label-merged-to-dev.yml, .github/workflows/lint-and-typecheck.yml, .github/workflows/secret-scan.yml, .github/workflows/test-mcp-regression.yml, .github/workflows/v0-user-flow-e2e.yml	Add label-closed-PR-to-dev automation; add lint/ruff/mypy checks; add TruffleHog secret scanning; extend MCP regression to Windows/Linux matrix with OS-gated eval/artifacts; add e2e flow assertions job + optional recording job with manual approval gate.
Release Notes & Ignore CHANGELOG.md, .gitignore	Add unreleased section documenting graph expansion, contradiction capture, hook updates, and M6 eval changes; exclude demo MP4s from git.

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~75 minutes

Possibly related issues

• feat(preflight): expand region-anchored lookup via 1-hop code-graph traversal #173: Directly implements 1-hop code-graph file-path expansion in preflight handler and code-locator adapter, resolving the feature request.
• Reduce collision-capture reminder spam from PostToolUse hook on bicameral_preflight #170: Implements deterministic preflight-intent gating via UserPromptSubmit hook and post-preflight contradiction capture to prevent unconditional reminder spam.
• [P1] SessionEnd capture-corrections hook is silently broken — design pivot to next-session surfacing #156: Updates SessionEnd hook command with re-entrancy guard and preserves --auto-ingest flag, directly addressing hook behavior concerns.

Possibly related PRs

• BicameralAI/bicameral-mcp#62: Both implement and wire 1-hop graph expansion for preflight, update M6 eval dataset/runner, and modify handlers/preflight with identical feature scope.
• BicameralAI/bicameral-mcp#140: Related release PR sharing v0.13.6 version bump, agent-session contradiction capture, and e2e test script changes.
• BicameralAI/bicameral-mcp#37: Both modify hook installation in setup_wizard and .claude/settings.json, transition from inline Python one-liners to hook scripts, and adjust PostToolUse/SessionEnd hook behavior.

Suggested labels

flow:release, type:feature, impact:contract-change

Suggested reviewers

• Knapp-Kevin

---

🐰 Hooks and graphs and demos, oh my!
A preflight gating spree, reaching ever high,
From intent to contradiction, a user asks "why?"
Then ledger-tested flows make reason comply. ✨

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch triage-from-dev

[coderabbitai]

Actionable comments posted: 3

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

setup_wizard.py (1)

57-59: ⚠️ Potential issue | 🟠 Major | ⚡ Quick win

Remove the stray f prefixes so lint passes again.

Ruff is already failing with F541 here because these strings do not interpolate anything. This is a straight CI blocker.

Minimal fix

     raw = input(
-        f"\n  History storage path (default: same as repo — press Enter to skip):\n  > "
+        "\n  History storage path (default: same as repo — press Enter to skip):\n  > "
     ).strip()

-        print(f"\n  Note: bicameral-mcp binary not found on PATH.")
+        print("\n  Note: bicameral-mcp binary not found on PATH.")

Also applies to: 790-790

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@setup_wizard.py` around lines 57 - 59, Remove the unnecessary f-string
prefixes on the input prompts that cause Ruff F541: locate the input call
assigning to raw (the line with raw = input(f"...").strip()) and remove the
leading f so the string is a plain literal; also find the other occurrence
mentioned around line 790 and remove its stray f prefixes as well so neither
prompt uses an f-string when there is no interpolation.

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@pyproject.toml`:
- Line 7: Update the package version metadata in pyproject.toml from "0.13.6" to
"0.13.7" so the release matches the PR objective; locate the version = "0.13.6"
entry and change it to version = "0.13.7" to ensure correct package metadata for
the merge/tag flow.

In `@scripts/hooks/post_preflight_capture_reminder.py`:
- Around line 64-72: In _format_reminder, validate and sanitize each item in
decisions before building bullets: ensure each entry is a dict (skip or coerce
non-dicts), read decision_id and description safely (fall back to '<unknown>' /
'<no description>'), strip or replace dangerous characters like '<', '>', and
newline characters and trim to a reasonable max length to avoid breaking the
<system-reminder> envelope, and then join the sanitized values to form the
bullets string; make these checks inside the generator (or a small helper within
the same function) so malformed items never raise when calling d.get(...) and
the reminder wrapper remains intact.

In `@skills/bicameral-preflight/SKILL.md`:
- Around line 330-376: Step 5.6 in SKILL.md inaccurately says captures happen
only when the user's prompt contradicts a surfaced decision; update the prose to
reflect the new mechanical behavior (always ingest when preflight surfaces
decisions) and clarify that only the bicameral.resolve_collision action choice
depends on user direction; reference the onboarding symbols that implement this
behavior (bicameral.ingest, bicameral.resolve_collision, the
mcp__bicameral__bicameral_preflight PostToolUse hook in
scripts/hooks/post_preflight_capture_reminder.py, and the wiring points
setup_wizard._install_claude_hooks and materialize_settings_with_hooks) so
readers know the change is intentional and the hook will always inject the
reminder but the LLM/agent decides supersede|keep_both|link_parent.

---

Outside diff comments:
In `@setup_wizard.py`:
- Around line 57-59: Remove the unnecessary f-string prefixes on the input
prompts that cause Ruff F541: locate the input call assigning to raw (the line
with raw = input(f"...").strip()) and remove the leading f so the string is a
plain literal; also find the other occurrence mentioned around line 790 and
remove its stray f prefixes as well so neither prompt uses an f-string when
there is no interpolation.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 74710b5a-f3e3-492c-94c9-b244f9c7c63f

📥 Commits

Reviewing files that changed from the base of the PR and between 6cb0e5f and 2b20bb2.

📒 Files selected for processing (11)

• .claude/settings.json
• pyproject.toml
• scripts/hooks/post_commit_sync_reminder.py
• scripts/hooks/post_preflight_capture_reminder.py
• setup_wizard.py
• skills/bicameral-preflight/SKILL.md
• tests/e2e/_harness_setup.py
• tests/e2e/run_e2e_flows.py
• tests/test_e2e_asserters.py
• tests/test_post_commit_sync_hook.py
• tests/test_post_preflight_capture_hook.py

🚧 Files skipped from review as they are similar to previous changes (1)

• tests/e2e/_harness_setup.py

[coderabbitai]

⚠️ Potential issue | 🟠 Major | ⚡ Quick win

Version still points at the previous release cut.

The PR objectives say this triage release must ship as v0.13.7. Leaving 0.13.6 here will produce the wrong package metadata for the merge/tag flow.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@pyproject.toml` at line 7, Update the package version metadata in
pyproject.toml from "0.13.6" to "0.13.7" so the release matches the PR
objective; locate the version = "0.13.6" entry and change it to version =
"0.13.7" to ensure correct package metadata for the merge/tag flow.

[coderabbitai]

⚠️ Potential issue | 🟠 Major | ⚡ Quick win

Sanitize and validate decision text before injecting it into <system-reminder>.

Line 66 currently promotes raw ledger text into a system-level wrapper. That makes stored decision_id / description values capable of breaking the reminder envelope or smuggling prompt text via characters like <, >, or newlines. It also assumes every list item is a dict; a malformed item will raise on d.get(...), which breaks the file's "never blocks a user" contract.

Suggested hardening

+def _safe_text(value: object, *, default: str) -> str:
+    text = default if value is None else str(value)
+    text = " ".join(text.splitlines())
+    return text.replace("&", "&amp;").replace("<", "&lt;").replace(">", "&gt;") or default
+
+
 def _format_reminder(decisions: list[dict]) -> str:
+    safe_decisions = [d for d in decisions if isinstance(d, dict)]
     bullets = "\n".join(
-        f"  - {d.get('decision_id', '<unknown>')}: {d.get('description', '<no description>')}"
-        for d in decisions
+        f"  - {_safe_text(d.get('decision_id'), default='<unknown>')}: "
+        f"{_safe_text(d.get('description'), default='<no description>')}"
+        for d in safe_decisions
     )
     return (
         "<system-reminder>\n"
-        f"bicameral.preflight surfaced {len(decisions)} prior decision(s):\n"
+        f"bicameral.preflight surfaced {len(safe_decisions)} prior decision(s):\n"
         f"{bullets}\n"

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@scripts/hooks/post_preflight_capture_reminder.py` around lines 64 - 72, In
_format_reminder, validate and sanitize each item in decisions before building
bullets: ensure each entry is a dict (skip or coerce non-dicts), read
decision_id and description safely (fall back to '<unknown>' / '<no
description>'), strip or replace dangerous characters like '<', '>', and newline
characters and trim to a reasonable max length to avoid breaking the
<system-reminder> envelope, and then join the sanitized values to form the
bullets string; make these checks inside the generator (or a small helper within
the same function) so malformed items never raise when calling d.get(...) and
the reminder wrapper remains intact.

[coderabbitai]

Actionable comments posted: 2

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

handlers/preflight.py (1)

214-249: ⚠️ Potential issue | 🟠 Major | ⚡ Quick win

Aggregate duplicate decision rows before deciding is_direct.

This loop dedupes on decision_id before it finishes provenance. If the ledger returns one row for a direct bind and another for an expanded-path bind, the first row wins. That means an expanded-path row arriving first will incorrectly downgrade a direct hit to confidence=0.7 and flip sources_chained to "graph" even though the caller pinned the decision directly.

Please union all bound paths for a decision_id first, then compute is_direct from that merged set.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@handlers/preflight.py` around lines 214 - 249, The loop currently dedupes on
decision_id early (seen_ids) before computing provenance, which can misclassify
a decision when the ledger returns both direct and expanded-path rows; change
the logic to first aggregate/merge all rows for each decision_id (collecting
union of bound_paths from d.get("code_regions") and top-level region_dict)
before computing is_direct and surfaced_via_expansion; modify the processing
around raw, seen_ids, bound_paths, region_dict and is_direct so you accumulate
per-decision bound_paths (and any other relevant flags) across all rows and only
after the union compute status/is_direct/surfaced_via_expansion and emit the
decision summary.

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@adapters/code_locator.py`:
- Around line 217-224: The call to self._ensure_initialized() is outside the try
in neighbors_for(), so initialization failures propagate instead of returning an
empty tuple; wrap the initialization call inside the same try/except (or expand
the try to include it) so that any exception from self._ensure_initialized(),
self._resolve_symbol_id_for_span, or self._neighbors_tool.execute results in
returning () as intended, referencing the neighbors_for(), _ensure_initialized,
_resolve_symbol_id_for_span, and _neighbors_tool.execute symbols.

In `@skills/bicameral-preflight/SKILL.md`:
- Around line 142-148: Update the SKILL.md text to reflect the actual bicameral
preflight contract: remove references to a topic-only fuzzy fallback and
per-decision confidence values (confidence=0.7/0.9) since the handler no longer
exposes them; instead explain that history() provides semantic recall, supplying
context, that passing file_paths enables region-anchored lookup, and that
provenance is exposed via PreflightResponse.decisions (BriefDecision) through
sources_chained rather than per-decision confidence; make the same edits for the
second block noted (lines ~171-189) so callers are not encouraged to omit
file_paths or rely on a nonexistent field.

---

Outside diff comments:
In `@handlers/preflight.py`:
- Around line 214-249: The loop currently dedupes on decision_id early
(seen_ids) before computing provenance, which can misclassify a decision when
the ledger returns both direct and expanded-path rows; change the logic to first
aggregate/merge all rows for each decision_id (collecting union of bound_paths
from d.get("code_regions") and top-level region_dict) before computing is_direct
and surfaced_via_expansion; modify the processing around raw, seen_ids,
bound_paths, region_dict and is_direct so you accumulate per-decision
bound_paths (and any other relevant flags) across all rows and only after the
union compute status/is_direct/surfaced_via_expansion and emit the decision
summary.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 76aad221-ac98-4725-8af5-2101d7409db6

📥 Commits

Reviewing files that changed from the base of the PR and between 2b20bb2 and b3fb654.

📒 Files selected for processing (14)

• CHANGELOG.md
• adapters/code_locator.py
• docs/preflight-failure-scenarios.md
• handlers/preflight.py
• scripts/hooks/post_preflight_capture_reminder.py
• scripts/hooks/preflight_reminder.py
• skills/bicameral-preflight/SKILL.md
• tests/e2e/prompts/flow-2-preflight.md
• tests/e2e/run_e2e_flows.py
• tests/eval/preflight_dataset.jsonl
• tests/eval/run_preflight_eval.py
• tests/test_post_preflight_capture_hook.py
• tests/test_preflight_graph_expansion.py
• tests/test_preflight_hook.py

✅ Files skipped from review due to trivial changes (2)

• docs/preflight-failure-scenarios.md
• CHANGELOG.md

🚧 Files skipped from review as they are similar to previous changes (5)

• scripts/hooks/preflight_reminder.py
• tests/test_preflight_hook.py
• scripts/hooks/post_preflight_capture_reminder.py
• tests/test_post_preflight_capture_hook.py
• tests/e2e/run_e2e_flows.py

[coderabbitai]

⚠️ Potential issue | 🟠 Major | ⚡ Quick win

Catch initialization failures inside neighbors_for().

neighbors_for() says it returns () on resolution/execution failure, but self._ensure_initialized() is outside the try. If the index is missing or stale, this method raises instead of degrading, which can break callers that expect the Jaccard signal to just drop to zero.

🩹 Minimal fix

     def neighbors_for(
         self,
         file_path: str,
         start_line: int,
         end_line: int,
     ) -> tuple[str, ...]:
         """Return 1-hop neighbor symbol addresses for a code span.
@@
-        self._ensure_initialized()
         try:
+            self._ensure_initialized()
             sym_id = self._resolve_symbol_id_for_span(file_path, start_line, end_line)
             if sym_id is None:
                 return ()
             neighbors = self._neighbors_tool.execute({"symbol_id": sym_id})
         except Exception:
             return ()

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	self._ensure_initialized()
	try:
	sym_id = self._resolve_symbol_id_for_span(file_path, start_line, end_line)
	if sym_id is None:
	return ()
	neighbors = self._neighbors_tool.execute({"symbol_id": sym_id})
	except Exception:
	return ()
	try:
	self._ensure_initialized()
	sym_id = self._resolve_symbol_id_for_span(file_path, start_line, end_line)
	if sym_id is None:
	return ()
	neighbors = self._neighbors_tool.execute({"symbol_id": sym_id})
	except Exception:
	return ()

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@adapters/code_locator.py` around lines 217 - 224, The call to
self._ensure_initialized() is outside the try in neighbors_for(), so
initialization failures propagate instead of returning an empty tuple; wrap the
initialization call inside the same try/except (or expand the try to include it)
so that any exception from self._ensure_initialized(),
self._resolve_symbol_id_for_span, or self._neighbors_tool.execute results in
returning () as intended, referencing the neighbors_for(), _ensure_initialized,
_resolve_symbol_id_for_span, and _neighbors_tool.execute symbols.

[coderabbitai]

⚠️ Potential issue | 🟠 Major | ⚡ Quick win

This section still describes a bicameral.preflight contract the handler no longer exposes.

The current handler does not fall back to topic-only fuzzy decision lookup, and the returned PreflightResponse.decisions are BriefDecisions, so the per-decision confidence=0.7/0.9 guidance here is not something the agent can actually inspect. Leaving this prose in the skill prompt nudges callers toward omitting file_paths and reasoning about a nonexistent field.

Please rewrite this around the real contract: history() provides semantic recall, file_paths unlock region-anchored lookup, and graph provenance is observable via sources_chained rather than per-decision confidence.

Also applies to: 171-189

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@skills/bicameral-preflight/SKILL.md` around lines 142 - 148, Update the
SKILL.md text to reflect the actual bicameral preflight contract: remove
references to a topic-only fuzzy fallback and per-decision confidence values
(confidence=0.7/0.9) since the handler no longer exposes them; instead explain
that history() provides semantic recall, supplying context, that passing
file_paths enables region-anchored lookup, and that provenance is exposed via
PreflightResponse.decisions (BriefDecision) through sources_chained rather than
per-decision confidence; make the same edits for the second block noted (lines
~171-189) so callers are not encouraged to omit file_paths or rely on a
nonexistent field.

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@CHANGELOG.md`:
- Around line 6-29: The file CHANGELOG.md contains unresolved Git merge conflict
markers (<<<<<<<, =======, >>>>>>>) around the Unreleased section; remove the
conflict markers and preserve the intended content (the "## [Unreleased]" block
and its Added/Changed entries shown between the markers) so the changelog is a
single coherent section; verify and keep the imports-only expansion text,
adapters/code_locator.py::RealCodeLocatorAdapter.expand_file_paths_via_graph
entry, SKILL.md changes, and test updates as the final content and delete the
leftover markers (<<<<<<< triage-from-dev, =======, >>>>>>> main) so no merge
markers remain.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 97a6f2c6-7dfa-49de-a5a7-e04450cbcf3a

📥 Commits

Reviewing files that changed from the base of the PR and between b3fb654 and c7d1274.

📒 Files selected for processing (1)

• CHANGELOG.md

[coderabbitai]

⚠️ Potential issue | 🟠 Major | ⚡ Quick win

Resolve leftover merge-conflict markers in CHANGELOG before merge.

CHANGELOG.md still contains unresolved markers (<<<<<<<, =======, >>>>>>>) at Line 6, Line 28, and Line 29. This is a release blocker because it leaves the changelog in an invalid merge state.

✅ Suggested fix

-<<<<<<< triage-from-dev
 ## [Unreleased]

 ### Added

 - `handlers/preflight.py` — `_region_anchored_preflight` now expands caller-supplied `file_paths` by 1 hop along the code-locator graph's **import edges** before the `binds_to` lookup. Lifts the strict exact-match recall ceiling so a decision bound to `app/src/lib/git/reorder.ts` surfaces when the caller passes the structurally-near `app/src/ui/multi-commit-operation/reorder.tsx`. Decisions reached only via expansion carry `confidence=0.7` (vs `0.9` for direct pins). `sources_chained` includes `"graph"` (alongside `"region"`) when expansion contributed at least one hit. Bounded per `#64`: ≤10 input seeds × `max_neighbors_per_result` neighbors per seed. Closes `#173` (and supersedes `#64`).
 - `adapters/code_locator.py::RealCodeLocatorAdapter.expand_file_paths_via_graph` — public method backing the expansion. Filters to ``imports`` edges only (file-level structural dependency); ``invokes`` / ``inherits`` / ``contains`` are symbol-level edges that over-broaden the file-level expansion. Returns `(expanded, added)` so callers can mark provenance.
 - `skills/bicameral-preflight/SKILL.md` Step 2 — documents the imports-only expansion + caller-side `confidence` and `sources_chained` semantics.
 - `tests/eval/preflight_dataset.jsonl` — M6 row flipped from XFAIL → live. Setup updated to specify graph-neighbor topology (`graph_neighbors`) and pinned-decision targets (`region_decisions_pinned_to`); the asserter now tests true graph-expansion semantics rather than mock-returns-decision-regardless-of-input.
 - `tests/eval/run_preflight_eval.py` — `_apply_setup` extended with `region_decisions_pinned_to` (path-aware decision lookup) and `graph_neighbors` (stub code_graph) so M6-style scenarios can be expressed in the dataset.

 ### Changed

 - `skills/bicameral-preflight/SKILL.md` Step 5.6 — judgment for contradiction-capture moves from the agent to the user via `AskUserQuestion` (Step 5.6.1). The agent no longer infers whether the prompt contradicts a surfaced decision; it asks the user (`supersede` / `keep_both` / `unrelated`) and acts mechanically on the answer (Step 5.6.2 — ingest + resolve_collision). The PostToolUse hook reminder now templates the disambiguation question rather than the bare ingest+resolve_collision sequence. Closes `#175`.
 - `tests/e2e/run_e2e_flows.py::assert_flow_2a` — pass criterion changed from "ingest+resolve_collision fired" to "`AskUserQuestion` invoked with disambiguation shape after preflight surfaced ≥1 decision." The user-side response can't be driven in headless `claude -p`, so the testable signal is the question invocation. The mechanical capture (Step 5.6.2) only fires after a human answers and is exercised in interactive Claude Code sessions, not CI.

 ### Fixed

 ### Schema

 ### Security
-
-=======
->>>>>>> main

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	<<<<<<< triage-from-dev
	## [Unreleased]
	### Added
	- `handlers/preflight.py` — `_region_anchored_preflight` now expands caller-supplied `file_paths` by 1 hop along the code-locator graph's **import edges** before the `binds_to` lookup. Lifts the strict exact-match recall ceiling so a decision bound to `app/src/lib/git/reorder.ts` surfaces when the caller passes the structurally-near `app/src/ui/multi-commit-operation/reorder.tsx`. Decisions reached only via expansion carry `confidence=0.7` (vs `0.9` for direct pins). `sources_chained` includes `"graph"` (alongside `"region"`) when expansion contributed at least one hit. Bounded per #64: ≤10 input seeds × `max_neighbors_per_result` neighbors per seed. Closes #173 (and supersedes #64).
	- `adapters/code_locator.py::RealCodeLocatorAdapter.expand_file_paths_via_graph` — public method backing the expansion. Filters to ``imports`` edges only (file-level structural dependency); ``invokes`` / ``inherits`` / ``contains`` are symbol-level edges that over-broaden the file-level expansion. Returns `(expanded, added)` so callers can mark provenance.
	- `skills/bicameral-preflight/SKILL.md` Step 2 — documents the imports-only expansion + caller-side `confidence` and `sources_chained` semantics.
	- `tests/eval/preflight_dataset.jsonl` — M6 row flipped from XFAIL → live. Setup updated to specify graph-neighbor topology (`graph_neighbors`) and pinned-decision targets (`region_decisions_pinned_to`); the asserter now tests true graph-expansion semantics rather than mock-returns-decision-regardless-of-input.
	- `tests/eval/run_preflight_eval.py` — `_apply_setup` extended with `region_decisions_pinned_to` (path-aware decision lookup) and `graph_neighbors` (stub code_graph) so M6-style scenarios can be expressed in the dataset.
	### Changed
	- `skills/bicameral-preflight/SKILL.md` Step 5.6 — judgment for contradiction-capture moves from the agent to the user via `AskUserQuestion` (Step 5.6.1). The agent no longer infers whether the prompt contradicts a surfaced decision; it asks the user (`supersede` / `keep_both` / `unrelated`) and acts mechanically on the answer (Step 5.6.2 — ingest + resolve_collision). The PostToolUse hook reminder now templates the disambiguation question rather than the bare ingest+resolve_collision sequence. Closes #175.
	- `tests/e2e/run_e2e_flows.py::assert_flow_2a` — pass criterion changed from "ingest+resolve_collision fired" to "`AskUserQuestion` invoked with disambiguation shape after preflight surfaced ≥1 decision." The user-side response can't be driven in headless `claude -p`, so the testable signal is the question invocation. The mechanical capture (Step 5.6.2) only fires after a human answers and is exercised in interactive Claude Code sessions, not CI.
	### Fixed
	### Schema
	### Security
	=======
	>>>>>>> main
	## [Unreleased]
	### Added
	- `handlers/preflight.py` — `_region_anchored_preflight` now expands caller-supplied `file_paths` by 1 hop along the code-locator graph's **import edges** before the `binds_to` lookup. Lifts the strict exact-match recall ceiling so a decision bound to `app/src/lib/git/reorder.ts` surfaces when the caller passes the structurally-near `app/src/ui/multi-commit-operation/reorder.tsx`. Decisions reached only via expansion carry `confidence=0.7` (vs `0.9` for direct pins). `sources_chained` includes `"graph"` (alongside `"region"`) when expansion contributed at least one hit. Bounded per `#64`: ≤10 input seeds × `max_neighbors_per_result` neighbors per seed. Closes `#173` (and supersedes `#64`).
	- `adapters/code_locator.py::RealCodeLocatorAdapter.expand_file_paths_via_graph` — public method backing the expansion. Filters to ``imports`` edges only (file-level structural dependency); ``invokes`` / ``inherits`` / ``contains`` are symbol-level edges that over-broaden the file-level expansion. Returns `(expanded, added)` so callers can mark provenance.
	- `skills/bicameral-preflight/SKILL.md` Step 2 — documents the imports-only expansion + caller-side `confidence` and `sources_chained` semantics.
	- `tests/eval/preflight_dataset.jsonl` — M6 row flipped from XFAIL → live. Setup updated to specify graph-neighbor topology (`graph_neighbors`) and pinned-decision targets (`region_decisions_pinned_to`); the asserter now tests true graph-expansion semantics rather than mock-returns-decision-regardless-of-input.
	- `tests/eval/run_preflight_eval.py` — `_apply_setup` extended with `region_decisions_pinned_to` (path-aware decision lookup) and `graph_neighbors` (stub code_graph) so M6-style scenarios can be expressed in the dataset.
	### Changed
	- `skills/bicameral-preflight/SKILL.md` Step 5.6 — judgment for contradiction-capture moves from the agent to the user via `AskUserQuestion` (Step 5.6.1). The agent no longer infers whether the prompt contradicts a surfaced decision; it asks the user (`supersede` / `keep_both` / `unrelated`) and acts mechanically on the answer (Step 5.6.2 — ingest + resolve_collision). The PostToolUse hook reminder now templates the disambiguation question rather than the bare ingest+resolve_collision sequence. Closes `#175`.
	- `tests/e2e/run_e2e_flows.py::assert_flow_2a` — pass criterion changed from "ingest+resolve_collision fired" to "`AskUserQuestion` invoked with disambiguation shape after preflight surfaced ≥1 decision." The user-side response can't be driven in headless `claude -p`, so the testable signal is the question invocation. The mechanical capture (Step 5.6.2) only fires after a human answers and is exercised in interactive Claude Code sessions, not CI.
	### Fixed
	### Schema
	### Security

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@CHANGELOG.md` around lines 6 - 29, The file CHANGELOG.md contains unresolved
Git merge conflict markers (<<<<<<<, =======, >>>>>>>) around the Unreleased
section; remove the conflict markers and preserve the intended content (the "##
[Unreleased]" block and its Added/Changed entries shown between the markers) so
the changelog is a single coherent section; verify and keep the imports-only
expansion text,
adapters/code_locator.py::RealCodeLocatorAdapter.expand_file_paths_via_graph
entry, SKILL.md changes, and test updates as the final content and delete the
leftover markers (<<<<<<< triage-from-dev, =======, >>>>>>> main) so no merge
markers remain.