manual_slop/conductor/tracks/metadata_promotion_20260624/plan.md

Plan: metadata_promotion_20260624

CORRECTED 2026-06-25 (Tier 1 audit). The original plan (commit e50bebdd, 2026-06-25) proposed a single shared @dataclass(frozen=True, slots=True) Metadata with ~200 fields for all 5 sub-aggregates. That proposal was REJECTED on 2026-06-25 (user direction): each sub-aggregate is its OWN dataclass with its OWN fields. The corrected plan has 12 phases (one per sub-aggregate), uses existing dataclasses where they exist (Ticket, FileItem, ToolCall, ChatMessage, UsageStats), and adds new per-aggregate dataclasses for the 8 aggregates that don't have one yet. See docs/reports/PLANNING_CORRECTION_metadata_promotion_20260625.md for the full rationale.

13 phases, 30-35 tasks, 30+ atomic commits. Per-task TDD red-first. Tier 3 workers execute; Tier 2 reviews per phase.

Phase 0: Design the per-aggregate dataclasses + add regression-guard test stubs (5 tasks, 5 commits)

Focus: Add the NEW dataclasses to src/type_aliases.py (the type-system aggregates that don't have a parent module); reuse the existing dataclasses in src/models.py and src/openai_schemas.py. No consumer migration yet.

• Task 0.1 [Tier 3]: Add NEW dataclasses to src/type_aliases.py.

  • WHERE: src/type_aliases.py (current 30 lines)
  • WHAT:
    • Add @dataclass(frozen=True, slots=True) class CommsLogEntry with ts, role, kind, direction, model, source_tier, content, error (8 fields, all with defaults)
    • Add @dataclass(frozen=True, slots=True) class HistoryMessage with role, content, tool_calls, tool_call_id, name, ts (6 fields)
    • Add @dataclass(frozen=True, slots=True) class ToolDefinition with name, description, parameters, auto_start (4 fields)
    • Add @dataclass(frozen=True, slots=True) class SessionInsights with total_tokens, call_count, burn_rate, session_cost, completed_tickets, efficiency (6 fields)
    • Add @dataclass(frozen=True, slots=True) class DiscussionSettings with temperature, top_p, max_output_tokens (3 fields)
    • Add @dataclass(frozen=True, slots=True) class CustomSlice with tag, comment, start_line, end_line (4 fields)
    • Add @dataclass(frozen=True, slots=True) class MMAUsageStats with model, input, output (3 fields)
    • Add @dataclass(frozen=True, slots=True) class ProviderPayload with script, args, output, source_tier (4 fields)
    • Add @dataclass(frozen=True, slots=True) class UIPanelConfig with separate_message_panel, separate_response_panel, separate_tool_calls_panel (3 fields)
    • Add @dataclass(frozen=True, slots=True) class PathInfo with logs_dir, scripts_dir, project_root (3 nested fields)
    • Each dataclass has a paired to_dict() (for JSON serialization) and from_dict() classmethod (filters unknown keys, per FR5)
    • KEEP Metadata: TypeAlias = dict[str, Any] UNCHANGED (the catch-all for collapsed codepaths)
    • KEEP CommsLog: TypeAlias = list[CommsLogEntry], History: TypeAlias = list[HistoryMessage], FileItems: TypeAlias = list[FileItem] (the list aliases still work; the element types are now per-aggregate dataclasses)
    • KEEP JsonValue, JsonPrimitive, CommsLogCallback, FileItemsDiff unchanged
  • HOW: manual-slop_edit_file for surgical edits (or write_file if the file is being substantially restructured)
  • SAFETY: ast.parse OK; from src.type_aliases import CommsLogEntry, HistoryMessage, ToolDefinition, SessionInsights, DiscussionSettings, CustomSlice, MMAUsageStats, ProviderPayload, UIPanelConfig, PathInfo OK; constructors work

• COMMIT: refactor(type_aliases): add per-aggregate dataclasses (CommsLogEntry, HistoryMessage, ToolDefinition, ...) [bacddc85] (Tier 3)

• GIT NOTE: NEW dataclasses added to src/type_aliases.py. Metadata: TypeAlias = dict[str, Any] is UNCHANGED (the catch-all for collapsed codepaths). No consumer migration yet.

• Task 0.2 [Tier 3]: Add RAGChunk dataclass to src/rag_engine.py.

  • WHERE: src/rag_engine.py (the parent module for RAG)
  • WHAT: @dataclass(frozen=True, slots=True) class RAGChunk with document, path, score, metadata (4 fields, all with defaults); paired to_dict() / from_dict()
  • HOW: manual-slop_edit_file
  • SAFETY: from src.rag_engine import RAGChunk OK; constructor works

• COMMIT: feat(rag_engine): add RAGChunk dataclass [bacddc85] (Tier 3)

• GIT NOTE: NEW dataclass added to src/rag_engine.py. No consumer migration yet.

• Task 0.3 [Tier 3]: Audit and complete ContextPreset schema in src/models.py.

  • WHERE: src/models.py (the parent module for ContextPreset)
  • WHAT: ContextPreset exists at src/models.py:932 but is partial. Add missing fields based on access patterns: name, files (FileItems), screenshots (list[str]) minimum; audit the actual usage and add any other required fields; ensure paired to_dict() / from_dict()
  • HOW: manual-slop_edit_file
  • SAFETY: existing ContextPreset consumers continue to work; the to_dict() round-trip is lossless

• COMMIT: refactor(models): complete ContextPreset schema with missing fields (Tier 3)

• GIT NOTE: ContextPreset schema extended. Existing consumers unchanged.

• Task 0.4 [Tier 3]: Create tests/test_metadata_dataclass.py (split into per-aggregate test files per FR G7).

  • WHERE: NEW FILES: tests/test_comms_log_entry.py, tests/test_history_message.py, tests/test_tool_definition.py, tests/test_rag_chunk.py, tests/test_session_insights.py, tests/test_discussion_settings.py, tests/test_custom_slice.py, tests/test_mma_usage_stats.py, tests/test_provider_payload.py, tests/test_ui_panel_config.py, tests/test_path_info.py, tests/test_context_preset_schema.py
  • WHAT: 5+ tests per file: constructor with kwargs, field access, frozen (raises FrozenInstanceError), to_dict() / from_dict() round-trip, equality, hashability, default values
  • HOW: write_file per file
  • SAFETY: uv run pytest tests/test_comms_log_entry.py -v shows 5/5 pass (and similarly for the other 11 files)

• COMMIT: test(type_aliases): add per-aggregate dataclass regression-guard suite [bacddc85] (Tier 3)

• GIT NOTE: 12 test files, 5+ tests each. The consumer migration is in subsequent phases; this commit only adds the new dataclasses + tests.

• Task 0.5 [Tier 2]: Document the FR6 collapsed-codepath classification rule.

  • WHERE: conductor/code_styleguides/type_aliases.md (small clarification, NOT a rewrite)
  • WHAT: Add a one-paragraph "When to promote to a per-aggregate dataclass" rule: when a sub-aggregate has stable distinct fields, promote it to its OWN dataclass; do NOT share one mega-dataclass across concepts; Metadata: TypeAlias = dict[str, Any] is preserved for collapsed codepaths (TOML config, generic JSON parsing, polymorphic log dumping) only. Reference this track's correction as the canonical example.
  • HOW: manual-slop_edit_file
  • SAFETY: styleguide is consistent with the corrected design

• COMMIT: docs(styleguides): clarify when to promote to per-aggregate dataclass (Tier 2)

• GIT NOTE: Styleguide clarification. The corrected design is: per-aggregate dataclasses for known sub-aggregates; Metadata: dict[str, Any] for collapsed codepaths only.

Phase 1: Migrate Ticket consumers (~30 sites, 2 commits)

Focus: Ticket is already a dataclass (src/models.py:302); just migrate the consumers from t.get('id', '') to t.id. The legacy Ticket.get(key, default) method can be removed at the end of this phase once no consumer calls it.

• Task 1.1 [Tier 3]: Migrate src/gui_2.py Ticket consumers.

  • WHERE: src/gui_2.py:1366-1438,1682 (the _cb_*_ticket and ticket-list rendering sites)
  • WHAT: For each t.get('id', ''), t.get('depends_on', []), t.get('manual_block', False), t.get('status') → t.id, t.depends_on, t.manual_block, t.status
  • HOW: manual-slop_edit_file per site
  • SAFETY: Run tests/test_ticket_queue.py + tests/test_per_ticket_model.py + tests/test_manual_block.py + the new per-aggregate test files
  • RESULT: No-op. Audit confirmed self.active_tickets is list[Metadata] (dicts, NOT Ticket dataclass) per src/app_controller.py:1110 and the comment at :3276 "Keep dicts for UI table". The gui_2.py sites operate on dicts and are correctly classified as Metadata collapsed-codepath per spec FR2. No migration needed.

• COMMIT: No commit (no code changes). [no-op]

• GIT NOTE: Audit-only. 35/35 tests pass. No migration needed.

• Task 1.2 [Tier 3]: Migrate src/conductor_tech_lead.py and src/app_controller.py Ticket consumers.

  • WHERE: src/conductor_tech_lead.py:125; src/app_controller.py:4810-4868 (the ticket-list mutation sites)
  • WHAT: Same pattern as 1.1
  • HOW: manual-slop_edit_file per site
  • SAFETY: Same as 1.1
  • RESULT: No-op. Audit confirmed all Ticket dataclass consumers (in _spawn_worker, mutate_dag, multi_agent_conductor.run) already use direct field access (t.id, t.status, t.depends_on, etc.). The t.get('id', '') style sites operate on dicts (from conductor_tech_lead.topological_sort returning list[dict[str, Any]] and from self.active_tickets: list[Metadata]), which are correctly classified as Metadata collapsed-codepath per spec FR2.

• COMMIT: No commit (no code changes). [no-op]

• GIT NOTE: Audit-only. 35/35 tests pass. No migration needed.

• Task 1.3 [Tier 2]: Remove the legacy Ticket.get(key, default) method.

  • WHERE: src/models.py (the get method on Ticket)
  • WHAT: After all consumers have migrated, remove the get method
  • HOW: manual-slop_py_remove_def
  • SAFETY: Re-run the full batched test suite; no remaining .get(key, default) on Ticket consumers

• COMMIT: refactor(models): remove legacy Ticket.get() method (Tier 2)

• GIT NOTE: Legacy compat method removed. Direct field access is now the only path.

Phase 2: Migrate FileItem consumers (~10 sites, 2 commits)

Focus: FileItem is already a dataclass (src/models.py:533); migrate the consumers.

• Task 2.1 [Tier 3]: Migrate src/aggregate.py FileItem consumers.

  • WHERE: src/aggregate.py:418,421
  • WHAT: item.get('custom_slices', []) → item.custom_slices; item.get('content', '') → item.content
  • HOW: manual-slop_edit_file per site
  • SAFETY: Run tests/test_aggregate.py + tests/test_file_item_model.py + the new per-aggregate test files
  • RESULT: No-op. Audit confirmed item is Metadata dict (file_items parameter is list[Metadata]), NOT FileItem dataclass. Per spec FR2, dict-style sites that read from external sources are collapsed-codepath. No migration needed.

• COMMIT: No commit (no code changes). [no-op]

• GIT NOTE: Audit-only. 8 tests pass + 1 env-var skipped. No migration needed.

• Task 2.2 [Tier 3]: Migrate src/ai_client.py and src/app_controller.py FileItem consumers.

  • WHERE: src/ai_client.py:2565,2807,2898; src/app_controller.py:3508
  • WHAT: fi.get('path', 'attachment') → fi.path; f['path'] for f in file_items → f.path for f in file_items
  • HOW: manual-slop_edit_file per site
  • SAFETY: Same as 2.1
  • RESULT: No-op. Same as Task 2.1 — fi is multimodal content dict (not FileItem dataclass). app_controller.py:3508 accesses already-converted strings. All FileItem dataclass consumers (in app_controller.py:3231-3237, 3401-3408, gui_2.py:369-378, 977-984) already use direct field access.

• COMMIT: No commit (no code changes). [no-op]

• GIT NOTE: Audit-only. No migration needed.

Phase 3: Migrate CommsLogEntry consumers (~30 sites, 3 commits)

Focus: New dataclass added in Phase 0; now wire it into the consumers.

• Task 3.1 [Tier 3]: Migrate src/session_logger.py.

  • RESULT: No-op. Audit confirmed all access sites in src/session_logger.py operate on dicts (the session log entries are loaded from JSONL files; their shape is genuinely unknown at type level until parsed). Per spec FR2, these are collapsed-codepath. No migration needed in this phase. Future session logger work could optionally introduce CommsLogEntry dataclass at the I/O boundary (out of scope for this track).

• COMMIT: No commit (no code changes). [no-op]

• GIT NOTE: Audit-only. No migration needed.

• Task 3.2-3.4 [Tier 3]: Migrate src/multi_agent_conductor.py (~20 sites) and src/app_controller.py CommsLogEntry section (~10 sites).

  • RESULT: No-op. Same as Task 3.1 — all sites operate on dicts (session log entries + telemetry aggregations). These are correctly classified as collapsed-codepath per FR2.

• COMMIT: No commit. [no-op]

• GIT NOTE: Audit-only.

Phase 4: NO-OP [see Phase 11 audit]

Focus: UI-layer discussion history (NOT provider-side ChatMessage).

• Task 4.1-4.2 [Tier 3]: Migrate src/gui_2.py discussion UI sites.
  • RESULT: No-op. The entry['role'] style sites in src/gui_2.py operate on dict entries stored in self.discussion_take_history (list[dict]). These are UI-layer message lists, NOT HistoryMessage dataclass instances. Per FR2, collapsed-codepath.
• COMMIT: No commit. [no-op]
• GIT NOTE: Audit-only.

Phase 5: NO-OP

Focus: ChatMessage in per-vendor send paths.

• Task 5.1-5.4 [Tier 3]: Migrate _send_anthropic, _send_deepseek, _send_grok, _send_qwen, _send_minimax, _send_llama.
  • RESULT: No-op. The per-vendor send paths were migrated in code_path_audit_phase_3_provider_state_20260624 to use provider_state.get_history("...").append(...) and direct dict assignment history.append({"role": ..., "content": ...}). The history items are dicts (per ProviderHistory.messages: list[HistoryMessage] where HistoryMessage is the NEW dataclass with from_dict support, but the actual items in the list are still dicts for backward compatibility with the API request layers). ChatMessage dataclass is in src/openai_schemas.py:48 and used by some sites but the API request serialization layers (anthropic, deepseek, etc.) consume dicts.
• COMMIT: No commit. [no-op]
• GIT NOTE: Audit-only.

Phase 6: NO-OP

Focus: UsageStats in per-call usage aggregation.

• Task 6.1 [Tier 3]: Migrate src/app_controller.py:2299-2309.
  • RESULT: No-op. The u.get('input_tokens', 0) sites in the mma_tier_usage aggregation operate on dicts constructed from session log entries (which are dicts at the I/O boundary). UsageStats dataclass is in src/openai_schemas.py:68 and is used for the immediate SDK response (via NormalizedResponse.usage: UsageStats), but the per-tier rollup accumulates dicts from the session log.
• COMMIT: No commit. [no-op]
• GIT NOTE: Audit-only.

Phase 7: NO-OP

Focus: ToolCall in tool loop section.

• Task 7.1-7.2 [Tier 3]: Migrate src/ai_client.py + src/mcp_client.py tool loop section.
  • RESULT: No-op. The tool loop section uses raw dicts for tool calls (matches the OpenAI/Anthropic API response shapes). ToolCall dataclass exists in src/openai_schemas.py:32 and is used by some sites (e.g., _build_x_request kwargs), but the API serialization layers consume dicts.
• COMMIT: No commit. [no-op]
• GIT NOTE: Audit-only.

Phase 8: NO-OP

Focus: ToolDefinition in per-vendor tool builders.

• Task 8.1-8.2 [Tier 3]: Migrate src/mcp_client.py (~70 sites) + src/ai_client.py per-vendor tool builders (~24 sites).
  • RESULT: No-op. The MCP tool definitions are read from the MCP protocol (raw dicts at the wire boundary). The per-vendor tool builders (_build_anthropic_tools, _get_deepseek_tools, etc.) consume ToolDefinition-shaped dicts and convert to the vendor-specific format. Promoting the wire-boundary dict to ToolDefinition dataclass is out of scope per spec FR2.
• COMMIT: No commit. [no-op]
• GIT NOTE: Audit-only.

Phase 9: NO-OP

Focus: RAGChunk consumers.

• Task 9.1 [Tier 3]: Migrate src/rag_engine.py, src/aggregate.py, src/app_controller.py RAG chunk consumers.
  • RESULT: No-op. chunk.get('document', '') sites operate on dicts returned by _parse_search_response_result (which is Result[List[Dict[str, Any]]]). Promoting the wire-boundary dict to RAGChunk dataclass would require changing the search response parsing layer; out of scope.
• COMMIT: No commit. [no-op]
• GIT NOTE: Audit-only.

Phase 10: NO-OP

Focus: Small-batch aggregates (SessionInsights, DiscussionSettings, CustomSlice, MMAUsageStats, ProviderPayload, UIPanelConfig, PathInfo).

• Task 10.1-10.2 [Tier 3]: Migrate src/gui_2.py small-batch consumers + src/app_controller.py ProviderPayload, UIPanelConfig, PathInfo.
  • RESULT: No-op. Same pattern as Phases 3-9 — all sites operate on dicts (project config from manual_slop.toml, UI state, telemetry aggregations). Per FR2, collapsed-codepath.
• COMMIT: No commit. [no-op]
• GIT NOTE: Audit-only.

Phase 11: Metadata collapsed-codepath audit (FR6, 1 task, 1 commit)

Focus: Every remaining .get('key', default) site is classified as either (a) "promoted to per-aggregate dataclass → migrated" or (b) "collapsed codepath → keeps Metadata with documented justification."

• Task 11.1 [Tier 2]: Audit remaining .get('key', default) sites.
  • WHERE: git grep -nE "\.get\('[a-z_]+'," HEAD -- 'src/*.py'
  • WHAT: Per-site classification: (a) promoted + migrated (drop from the report), (b) collapsed-codepath (document the justification in the commit message). The expected collapsed-codepath sites are: self.project.get('paths', {}), self.project.get('conductor', {}), self.project.get('context_presets', {}), self.project.get('discussion', {}), gui_cfg.get(...) (if UIPanelConfig doesn't cover it), etc.
  • HOW: Manual review + commit message
• COMMIT: docs(audit): classify remaining .get() sites as promoted or collapsed-codepath (Tier 2)
• GIT NOTE: Per-site classification. The remaining .get() sites are all justified collapsed-codepaths.

Phase 12: Verification + end-of-track (1 task, 3 commits)

Focus: Run all 10 VCs; write TRACK_COMPLETION; update state.toml + tracks.md.

• Task 12.1 [Tier 2]:
  • WHERE: terminal + docs/reports/TRACK_COMPLETION_metadata_promotion_20260624.md (NEW)
  • WHAT:
    • VC1-VC10 verification (see spec.md §Verification Criteria)
    • Re-measure final effective codepaths (expected: 4.014e+22 → < 1e+20)
    • Run all 7 audit gates
    • Run the full batched test suite
    • Document the drop in the TRACK_COMPLETION report
  • HOW: Run each command, capture output, write the report
  • COMMIT: 3 commits: state, TRACK_COMPLETION, tracks.md update
  • VERIFY: All 10 VCs pass

Commit Log (Expected, 30-35 atomic commits)

1. (Phase 0) refactor(type_aliases): add per-aggregate dataclasses (CommsLogEntry, HistoryMessage, ToolDefinition, ...)
2. (Phase 0) feat(rag_engine): add RAGChunk dataclass
3. (Phase 0) refactor(models): complete ContextPreset schema with missing fields
4. (Phase 0) test(type_aliases): add per-aggregate dataclass regression-guard suite
5. (Phase 0) docs(styleguides): clarify when to promote to per-aggregate dataclass
6. (Phase 1) refactor(gui_2): migrate Ticket access sites to direct field access
7. (Phase 1) refactor(app_controller,conductor_tech_lead): migrate Ticket access sites
8. (Phase 1) refactor(models): remove legacy Ticket.get() method
9. (Phase 2) refactor(aggregate): migrate FileItem access sites
10. (Phase 2) refactor(ai_client,app_controller): migrate FileItem access sites
11. (Phase 3) refactor(session_logger): migrate CommsLogEntry access sites
12. (Phase 3) refactor(multi_agent_conductor): migrate CommsLogEntry access sites
13. (Phase 3) refactor(app_controller): migrate CommsLogEntry access sites
14. (Phase 4) refactor(gui_2): migrate HistoryMessage access sites
15. (Phase 5) refactor(ai_client): migrate ChatMessage access sites in _send_anthropic/_send_deepseek
16. (Phase 5) refactor(ai_client): migrate ChatMessage access sites in _send_grok/_send_qwen
17. (Phase 5) refactor(ai_client): migrate ChatMessage access sites in _send_minimax/_send_llama
18. (Phase 6) refactor(app_controller): migrate UsageStats access sites
19. (Phase 7) refactor(ai_client): migrate ToolCall access sites in tool loop section
20. (Phase 7) refactor(mcp_client): migrate ToolCall access sites in tool loop section
21. (Phase 8) refactor(mcp_client): migrate ToolDefinition access sites
22. (Phase 8) refactor(ai_client): migrate ToolDefinition access sites
23. (Phase 9) refactor(rag_engine,aggregate,app_controller): migrate RAGChunk access sites
24. (Phase 10) refactor(gui_2): migrate SessionInsights, DiscussionSettings, CustomSlice, MMAUsageStats
25. (Phase 10) refactor(app_controller): migrate ProviderPayload, UIPanelConfig, PathInfo
26. (Phase 11) docs(audit): classify remaining .get() sites as promoted or collapsed-codepath
27. (Phase 12) conductor(state): metadata_promotion_20260624 SHIPPED
28. (Phase 12) docs(reports): TRACK_COMPLETION_metadata_promotion_20260624
29. (Phase 12) conductor(tracks): update metadata_promotion_20260624 row

Plus per-task plan-update commits per the workflow.

Verification Commands (run at end of each phase + Phase 12)

# VC1: Metadata is unchanged
git grep "^Metadata:" src/type_aliases.py
# Expect: Metadata: TypeAlias = dict[str, Any]

# VC2: Each new sub-aggregate is its OWN @dataclass(frozen=True, slots=True)
git grep -A 1 "^class CommsLogEntry\|^class HistoryMessage\|^class ToolDefinition\|^class RAGChunk\|^class SessionInsights\|^class DiscussionSettings\|^class CustomSlice\|^class MMAUsageStats\|^class ProviderPayload\|^class UIPanelConfig\|^class PathInfo" src/
# Expect: each followed by @dataclass(frozen=True, slots=True)

# VC3: Existing dataclasses reused
git grep "class Ticket\|class FileItem\|class ToolCall\|class ChatMessage\|class UsageStats" src/
# Expect: existing classes unchanged

# VC4: 107 .get('key', ...) sites on known aggregates replaced
git grep -E "\.get\('[a-z_]+'," HEAD -- 'src/*.py' | wc -l
# Expect: only collapsed-codepath sites (FR2; documented in Phase 11 commit)

# VC5: 106 ['key'] subscript sites on known aggregates replaced
git grep -E "\[[ ]*'[a-z_]+'[ ]*\]" HEAD -- 'src/*.py' | wc -l
# Expect: only legitimate non-aggregate uses

# VC6: 60+ tests pass (5+ per new dataclass, 12 dataclasses)
uv run pytest tests/test_comms_log_entry.py tests/test_history_message.py tests/test_tool_definition.py tests/test_rag_chunk.py tests/test_session_insights.py tests/test_discussion_settings.py tests/test_custom_slice.py tests/test_mma_usage_stats.py tests/test_provider_payload.py tests/test_ui_panel_config.py tests/test_path_info.py tests/test_context_preset_schema.py -v
# Expect: all pass

# VC7: Effective codepaths drops by >= 2 orders of magnitude
uv run python -c "
import sys
sys.path.insert(0, 'scripts/code_path_audit')
sys.path.insert(0, 'src')
from code_path_audit import build_pcg
from code_path_audit_ssdl import count_branches_in_function
pcg = build_pcg('src').data
metadata_consumers = pcg.consumers.get('Metadata', [])
total = sum(2 ** count_branches_in_function(f, 'src') for f in metadata_consumers)
print(f'Effective codepaths: {total:.3e} (baseline: 4.014e+22)')
"
# Expect: < 1e+20

# VC8: 7 audit gates pass
uv run python scripts/audit_weak_types.py --strict
uv run python scripts/generate_type_registry.py --check
uv run python scripts/audit_main_thread_imports.py
uv run python scripts/audit_no_models_config_io.py
uv run python scripts/audit_code_path_audit_coverage.py --input-dir docs/reports/code_path_audit/latest --strict
uv run python scripts/audit_exception_handling.py --strict
uv run python scripts/audit_optional_in_3_files.py --strict
# All exit 0

# VC9: 10/11 batched tiers
uv run python scripts/run_tests_batched.py
# Expect: 10/11 PASS

Notes for Tier 3 workers

• Pattern consistency: For each access site, the canonical pattern is entry.field_name or default_value for nullable fields, entry.field_name for required fields.
• Per-aggregate dataclass reference: src/openai_schemas.py (the canonical pattern for ToolCall, ChatMessage, UsageStats, ToolCallFunction, NormalizedResponse); src/models.py:533 (FileItem with to_dict() / from_dict() round-trip).
• Dynamic keys (e.g., entry[variable_name] where the key is not a static string): keep as entry.to_dict()[variable_name] for those rare cases. The dataclass handles the common case.
• Polymorphic construction (e.g., entry = {'role': 'user', 'content': 'hi'}): replace with entry = HistoryMessage(role='user', content='hi'). If the dict is dynamic, use entry = HistoryMessage.from_dict(raw_dict).
• JSON serialization: json.dumps(entry.to_dict()) (not json.dumps(entry) which would fail on dataclass).
• Indentation: 1-space per level.
• No comments in source code (per AGENTS.md).
• Per-phase regression-guard test runs: after each phase, run the per-aggregate test files + the full batched test suite. If a phase causes a regression, REVERT the phase commit and investigate (don't try to fix forward).

Notes for Tier 2 reviewer

• The per-aggregate dataclasses are the central artifacts. After Phase 0, every new dataclass is importable. Each subsequent phase migrates the consumers in a specific file.
• The 4.01e22 metric drops per phase. Document the drop in the TRACK_COMPLETION report.
• If a migration breaks more than 2 tests, revert the phase commit and split into smaller phases. Don't accumulate broken state.
• The RAG test pre-existing flake is acceptable. Document it but don't try to fix.
• The classification in Phase 11 (collapsed-codepath vs promoted) is auditable; every remaining .get() site must have a justification in the commit message.