Switching among Claude Code, Codex, Copilot, and Cursor creates fragmented session history. A restart can leave the developer remembering the project but not the harness, folder, session identifier, or exact wording needed to find it.
The first version of SS searched local transcripts. That was useful, but it didn't solve recovery. Results could be technically relevant while failing to explain what happened, what remained, or how to reopen the work.
SS now indexes local harness records into SQLite and combines exact full-text search, character-level fuzzy matching, whole-session embeddings, and per-turn embeddings. Results are grouped by native session and ranked with bounded recency, source intent, query coverage, and filters for worker sessions, expanded instructions, and transcript dumps.
Each result includes the original folder, session purpose, visible progress, continuation evidence, and the correct next action. Claude Code and Codex can reopen their native sessions. Cross-harness continuation creates a local context packet. Archive state removes completed work from the active dashboard without deleting native history.
SS keeps all retrieval local. The first release also chose deterministic cards over generated summaries because they were fast, traceable, private, and reproducible. That held until real dashboards showed the cost: extractive About lines were technically correct and still vague.
The second release added model summaries without giving up the original guarantees, and each one became a mechanism rather than a promise. Summaries are off until the user sets an explicit switch, because an API key in the environment is not consent. Credential shapes are stripped at the one place the outbound payload is assembled, so no future call site can forget. Every card records whether its summary came from the model or from local evidence, the dashboard only trusts the model kind, and a card built during an outage is rebuilt automatically instead of being cached as if it were good. Routine backfill stops at the newest sixty sessions, so the tool never pays to summarize work nobody reads. The deterministic cards remain as the universal fallback and the local-only mode.
Unsupported editor behavior appears in an executable capability matrix instead of being implied by documentation.
Reliability work covered nested transactions, atomic migrations, backups, schema validation, lock contention, process termination, read-only storage, SQLite full conditions, selector isolation, malformed input, and archive-state recovery.
The public corpus contains 120 synthetic sessions and 80 frozen queries split across exact, fuzzy, semantic, and time-oriented recovery.
| Mode | Top-one accuracy | Recall at five | MRR at ten |
|---|---|---|---|
| Full text | 60.00% | 61.25% | 0.606 |
| Local fuzzy | 68.75% | 70.00% | 0.694 |
| Hybrid | 90.00% | 95.00% | 0.918 |
Hybrid search preserved perfect exact-identifier retrieval, raised fuzzy top-one accuracy from 15% to 95%, and raised semantic top-one accuracy from 35% to 65%.
At 1,000 synthetic sessions, hybrid query latency measured 191 ms p50 and 220 ms p95 after embeddings were available. The one-time embedding build took 6.8 seconds on the recorded Apple Silicon Mac.
Adversarial testing repeatedly found defects that the original suite missed: nested rollbacks erased outer work, partial migrations were trusted, pasted commands could archive sessions, typo prefilters disagreed with the parser, and weak evidence produced malformed cards.
The summary work went through the same treatment. A review of the first implementation proved that a failed model call was cached forever as a good card, that raw transcript lines could reach the dashboard when the cache missed, and that two different truncation rules disagreed about typographic ellipses. Mutation testing then killed sixteen surviving mutants in the suite itself, including a privacy check that silently skipped files nobody listed.
Those failures were retained as regressions. Independent archive-intent batches were recorded even when they failed rather than being relabeled after the fact.
This release supports macOS and Linux. Claude Code, Codex, and Pi have proven native reopening. VS Code/Copilot and Cursor remain searchable and support packet continuation, but exact native reopening isn't claimed. Codex assistant output is partially available because its low-level telemetry doesn't yet provide a stable, complete thread mapping.