Folds three prior unsigned commits (8fe5853 switch summoner LLM, 377ae66 abstract hook protocol, ddd94cc/d8d3295/6875892 memory freshness+multi-harness+lineage+preflight) into a single signed commit for the WIP PR. abstract hook protocol (issue #50): - core/models/protocol/events.py: AbstractEventSource, AbstractEventType (snake_case values), ToolCallPayload, AbstractHookEvent (ConfigDict(extra='forbid')) - core/protocol/dispatch.py: dispatch_event + emit_event_from_stdin - core/protocol/session.py: SessionManager(source, state_root) with legacy migration from ~/.claude/slopometry/ - core/protocol/adapters/base.py: HookEventAdapter Protocol + ADAPTERS registry - core/protocol/adapters/claude_code.py: ClaudeCodeAdapter + ToolType StrEnum + resolve_tool_type - core/protocol/adapters/opencode.py: OpenCodeAdapter + resolve_opencode_event_type - core/hook_handler.py + core/opencode_handler.py: thin dispatchers - core/database.py: save_event(AbstractHookEvent), get_session_events returns AbstractHookEvent - core/models/hook.py: removed HookEvent/HookEventType/EventSource/ ToolType/HookInputUnion/PreToolUseInput/PostToolUseInput/ NotificationInput/StopInput/SubagentStopInput (harness-specific schemas no longer needed) - core/migrations.py: Migration015AbstractEventTypeValues translates legacy PascalCase event_type values to snake_case canonical memory extraction across harnesses (issue #46): - find_memories discovers + extracts from Claude Code AND OpenCode sessions, per-session source routing via DiscoveredTranscript, processed_memory_sessions.source column (Migration017 backfills existing rows to claude_code) - MemoryExtractor.extract_memories_from_opencode_session walks OpenCode's message/<sid>/<message_id>.json + part/<mid>/<part_id>.json tree, emits same USER:/ASSISTANT:/[TOOL: ...] format as Claude parser; skips step-start + reasoning parts - TranscriptFinder.find_opencode_storage_root resolves XDG/Windows/Linux paths; find_opencode_sessions matches by project.worktree OR session.directory being a descendant (subdirectory sessions) - TranscriptTruncationConfig (Pydantic BaseModel, ConfigDict(extra='forbid')) with tool_input_chars/tool_output_chars/tool_result_chars defaults that preserve pre-refactor behavior freshness validator (4-action LLM judge): - ProjectSimilarityDistribution derives dedupe threshold from project's own pairwise similarity quantiles (p75 clamped to [FLOOR=0.45, CEILING=0.95]) — no hardcoded similarity threshold for candidate generation - _judge_reconciliation asks the LLM to choose keep_both / merge / supersede / dedupe (with merged_content on merge only); all open-ended judgment delegated to the LLM - scripts/calibrate_freshness_threshold.py shows Qwen3-VL-Embedding-8B on slopometry's 58 memories: n=1653 mean=0.45 p75=0.50 → current clamps reasonable lineage (memory superseded_by): - MemoryEntry.superseded_by: str | None + schema column + CRUD + Migration016 + index in migration only (avoid CREATE INDEX on non-existent column for upgraded DBs) - CLI no longer flips retained on supersede; instead links existing → superseded_by=<new_id>. merge rewrites candidate content; dedupe annotates metadata and skips save; keep_both unchanged. All decisions record freshness_action/reason/pair_with. preflight (explicit failure on unreachable endpoints): - src/slopometry/solo/cli/preflight.py preflight_endpoints validates both chat LLM + embedding via GET /v1/models at start of find_memories; aborts with click.ClickException listing all failures (no partial batch completion). --dry-run skips the check. misc: - 12 enums migrated to StrEnum via ruff --fix --unsafe-fixes - pyrightconfig.json (basic mode) - version bump 2026.4.15 → 2026.6.23 - removed unused processed_key, redundant or None, MIN_CANDIDATE_CONTENT_LEN - added id: int | None to AbstractHookEvent tests (165 new): - 27 freshness tests (ProjectSimilarityDistribution, 4-action judge, threshold derivation cases) - 9 transcript_finder tests (worktree + subdirectory matching, storage-root resolution, XDG handling) - 10 memory_extractor tests (user/assistant/tool reconstruction, step-start/reasoning skip, truncation config, ordering) - 6 opencode_memory_integration tests (real OpenCode session extraction + stubbed LLM end-to-end supersede/merge/dedupe/ keep_both pipeline with fresh DB fixture) - 8 preflight tests (endpoint check, exception wrapping, dry-run skip, offline_mode precedence) - existing test updates: hook_handler, opencode_handler, database, sessions_performance, list_sessions_performance, plan_analyzer, notebookread_integration, posttooluse_validation, transcript_token_analyzer, models, baseline_service, current_impact_service, qpe_calculator, implementation_comparator, memory_service, git_tracker, protocol_events, claude_code_adapter, opencode_adapter, protocol_session verification: - 976 tests pass (was 916 before this batch); only pre-existing stale --help assertion failure in test_user_story_generation_cli_integration - ruff clean, pyright clean on all changed files - End-to-end verified on /mnt/terradump/code/microagi/moelite with MicroAGI MiniMax-M3 endpoint from opencode.json (provider 'microagi-minimax-m3-mxfp4'): 11 sessions processed, 8 memories saved, freshness validator ran (KEEP_BOTH, sim=0.59), project distribution n=55 mean=0.48 p75=0.57 → derived threshold 0.57 - Bad endpoint (https://nonexistent.invalid/v1) → immediate abort with explicit error Signed with key 34A5902CFFFC0171037ABFE2F810974410C09E22 (Alex Korolev (HOT) <lxk@droidcraft.org>).

Implement a stable abstract hook protocol instead of harness-specific mappings #50

Description

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions