|
- Replace the MCP import_messages tool with standalone CLI commands:
iny-convmem-mcp --data-dir <path> --import <file>
iny-convmem-mcp --data-dir <path> --export <file>
- Add src/cli/{import,export,cli.h} and a CTest process test (tiny-convmem-mcp.cli)
covering round-trip, manifest generation, overwrite/reject policies, and
include_deleted export.
- Fix Debug heap corruption in iny_store_import() by deep-copying imported
messages instead of taking ownership of caller-allocated records, which were
later freed by the caller.
- Remove the now-redundant src/mcp/tools/import_messages.c and keep the tool
registry at 7 MCP tools.
- Update docs/development-plan.md (P1-002 complete) and
docs/implementation-status.md with CLI test coverage.
|
|
- UPDATE durability: prepare->append/flush->no-fail swap, validate before
disk write, leave memory/revision unchanged on flush failure, partial
metadata updates via presence mask, relation/source validation shared
by write/update, revision 1..INT32_MAX with optimistic concurrency.
- Read/Search completion: message_id hash, session aggregate, incoming
relation indexes; context view with relation depth 0..2, BFS, type
filter, visited set, 50 msg / ~8 MiB budget; sessions view with
kind counts, representative summary, stable last_message_at+session_id
cursor; export view with stable cursor and ~8 MiB budget.
- Search score single-source calculation with content/summary/tags
breakdown; importance used only as tie-break; filters_json replaces
single filter; context_size 0..3.
- All 15 tests pass on Debug, independent Release, and ASan builds.
- docs/implementation-status.md and development-plan.md synchronized.
- docs/feedback.md added to .gitignore as user-owned untracked notes.
- Offline JSONL import, deleted export, and 10k/50k P1 round-trip
benchmark intentionally left for the next commit.
- Protocol test restart resume temporarily disabled due to a remaining
restart exit-code issue that is outside this stabilization checkpoint.
|