pi-lossless 0.1.1 → 0.1.3

package/CHANGELOG.md

@@ -1,5 +1,23 @@
 # pi-lossless
 
+## 0.1.3
+
+### Patch Changes
+
+- 470a7b7: Verify file bytes and sha256 against the linked message text before inserting file rows on import. A repeated entry id with different text, or a file row whose text differs from the stored message, also refuses the import without writing. Existing rows stay untouched.
+- 470a7b7: Signal a watermark miss on the ingest path. A full rescan now writes one `ingest-rescan` metric row and, in interactive sessions, one UI warning. Empty entries stay silent and ingest behavior is unchanged.
+- 470a7b7: Add a mechanical tier for noisy tools. Name tools in `mechanicalTools` in summarizer-config, `/lcm` config, or the OpenCode plugin options. Their spans compact into deterministic `mechanical` leaves with zero summarizer spend at compaction time. Every message stays stored verbatim and covered, so the integrity coverage check passes, and heal re-derives a skipped leaf from provenance on demand. There is no ingest filtering.
+- 470a7b7: Report orphan sidecars in doctor. `integritySnapshot` now lists sidecar files and stale `wal` and `shm` files with no live session (`.sqlite` by default, `.db` for Pi). `checkIntegrity` reports them as a read-only `orphan-sidecar` finding through `lcm doctor` and the diagnose bundle. Report only. Deletion stays manual.
+- 470a7b7: Upgrade recall search. `lcm_grep` takes `role`, `date_start`/`date_end`, and `tool_name` filters. Role and time are SQL predicates on the stored columns with a covering index. Tool scope is an FTS rewrite against `[tool:Name]` markers with no new table or index. `lcm_expand` takes an `anchor` entry id that centers the listing window with edge clamping. The recall manifest and the retrieval brief order by stop-word stripped query-word overlap over summary headers. A read-only files view lists each linked file once with an incomplete-coverage flag and a `large-inline` mark. No migration and no new writer.
+- 470a7b7: Add a read-only static sidecar dump. `/lcm dump` and `/lossless-dump` open the session sidecar `queryOnly` and reuse the recall views and metric rows. The dump redacts displayed seams, labels removed rows, and reports old generations as named skips. Output is capped at 16 KB.
+- 0c70748: Ask providers for a whole number of tokens. A leaf target was half of the calibrated chunk budget, so it was fractional whenever that budget was odd, and a provider that validates `max_tokens` as an unsigned integer rejected the call with a 422. The ladder then stored the deterministic truncate: in one week of Pi sessions that was 1,315 of 5,358 summaries, with 540 rejected calls naming values such as `max_tokens: 5543.5`. The request builder now rounds to nearest, never below one, and treats a non-finite budget as that floor, before the reasoning headroom is added. Integer budgets produce the same request as before. No config key, migration or metric row changes.
+
+## 0.1.2
+
+### Patch Changes
+
+- de419eb: Ship a versioned Intent skill with each package (`skills/<name>/SKILL.md` in the published tarball): install, command and recall vocabulary, config essentials, and common mistakes, stamped with the package version on release. Adds `skills` to each package's published files and validates skills on publish and in CI.
+
 ## 0.1.1
 
 ### Patch Changes

package/README.md

@@ -32,6 +32,8 @@ once context crosses the soft threshold (default: 70% of the model's window).
 /lcm backup     # copy the live store to a timestamped file
 /lcm export     # write the full store as JSONL (verbatim; see Security notes)
 /lcm report     # metrics summary for the current session
+/lcm dump       # read-only dump of the session sidecar; add --include-removed
+                # to keep removed-branch rows, --save <path> to write mode 0600
 /lcm doctor     # diagnostics: store checks, thresholds, summarizer chain,
                 # footprint; add --repair to rebuild the index and drop the rows
                 # nothing can reach
@@ -61,7 +63,9 @@ a call, or a span needs, because an index cannot hold punctuation. A pattern is
 case-insensitive unless `case_sensitive` is set. Both paths page: hits come 20
 at a time, `offset` reaches the next page, and the query path reports its total
 while the pattern path says whether more matches exist rather than claiming a
-count it did not compute.
+count it did not compute. Four filters narrow the search: `role` keeps
+only rows stored with that role, `date_start` and `date_end` bound rows by
+timestamp, and `tool_name` keeps only tool-result rows for that tool.
 
 All three take an address when you have one, so a known id does not have to be
 searched for: `lcm_grep` accepts `summary_id` to search inside one node, and