trajectories 0.0.0 → 0.1.1

package/docs/research/ledger/docs/research/2026-09-07-confer-inspirations.md

@@ -0,0 +1,163 @@
+# Confer inspiration review — 2026-09-07
+
+## Outcome
+
+Retain published `@letta-ai/trajectory`. Independently implement writer ownership
+and transactional ingestion progress using Obelisk as a behavioral reference;
+learn recovery and supervision from Cass; use AgentsView's extraction/retraction
+semantics as the strongest inspected Dream reference. AI SDK Harnesses offer a
+live capture seam, not Ledger's persistence layer.
+
+This is source research and a proposed implementation sequence, not a claim that
+Ledger now provides these guarantees. No upstream code was copied.
+
+## Reproducibility and method
+
+Confer shell snapshot `e0d97b88dfc57c170845a906875ff7ac0ae29d60e185d031e6ec55cdfc5f6814`,
+created `2026-09-07T00:08:05.393Z`. Read the manifest, searched the five repositories,
+then inspected narrow numbered implementation and test ranges. No corpus clones.
+
+| Repository | Commit |
+|---|---|
+| letta-ai/trajectory | `21ae92d10579499b183d93929c85da0b91233167` |
+| tommy0103/obelisk | `69d9e21a90e55f310376d8f49929b1c16f5a1f0c` |
+| dicklesworthstone/cass | `a18a6261a9815204b6ef59f9d186370ebd35dc34` |
+| kenn-io/agentsview | `2298453833a5c5126a486a2554eac6c35cd2b55f` |
+| vercel/ai | `c3c189c068baefe62808b4eca5432fe092dc37cb` |
+
+Also used Confer's durable research-agent surface: job
+`dff6263d-7c31-49d3-aced-58d48b078327`, terminal state `succeeded`.
+Reported execution: oh-my-pi, Google Gemini 3.7 Flash, 41 turns. This confirms a
+research-agent execution, not multiple independent reviewers; telemetry did not
+show a multi-agent fan-out. Retrieve with `confer.py status <job-id>`.
+
+The agent report supplied useful leads on leases, index recovery and supervision.
+Shell validation corrected its overclaims: normalization output bounds do not
+prove bounded peak ingestion memory; `os.OpenRoot` is not a model execution
+sandbox; generic AI SDK generation APIs do not establish Harness durability.
+Some embedded report snippets differed from the cited bytes, so the evidence
+below uses directly inspected shell ranges instead. Coverage is targeted, not
+an exhaustive security or runtime audit. Upstream tests were read, not run.
+
+## Adopt, adapt, avoid
+
+### Trajectory: retain the dependency; own the ingestion protocol
+
+Native identities and byte-location identities support conflict detection;
+content fallback supports exact dedup only. Codex/Cursor byte offsets are chunkable
+with `baseByteOffset`, while other formats can use whole-decode ordinals. Do not
+apply generic byte chunking to every adapter.
+
+Evidence: `letta-ai/trajectory@21ae92d10579499b183d93929c85da0b91233167:CANONICAL.md#L115-L152`.
+
+The documented worker contract owns dedup by source/record identity, content
+versions, configuration hashes and serialized ingestion IDs. Ledger must persist
+these with committed records; normalization alone does not create a Dream cursor.
+
+Evidence: `letta-ai/trajectory@21ae92d10579499b183d93929c85da0b91233167:CANONICAL.md#L154-L170`.
+
+These are corpus contracts, not proof the locally installed package version
+exports every API. Verify installed-version compatibility before implementation.
+
+### Obelisk: independent implementation of ownership and commit semantics
+
+Its lease holds `BEGIN IMMEDIATE` on a separate SQLite connection, releases via
+rollback/close, and treats contention separately from unexpected errors. Its
+provider loop persists data and invokes the persisted callback inside one
+transaction, then invokes the committed callback afterwards; failures are
+reported with a skip/stop policy. Useful patterns for Ledger's cursor ownership.
+
+Evidence:
+
+- `tommy0103/obelisk@69d9e21a90e55f310376d8f49929b1c16f5a1f0c:packages/core/src/writer-lease.ts#L54-L93`
+- `tommy0103/obelisk@69d9e21a90e55f310376d8f49929b1c16f5a1f0c:packages/core/src/provider-indexing.ts#L264-L300`
+
+Do not copy AGPL code. A database lease is writer coordination, not process
+isolation, and only works if every writer participates. Cursor+records must
+commit together; a model call must not hold the write transaction open.
+
+### Cass: supervision and source-bound recovery
+
+The launchd renderer XML-escapes each argument and distinguishes incremental
+intervals from nightly schedules. Its lexical reconciliation compares a durable
+checkpoint's source identity and shape against current input before retrying;
+changed input fails with explicit recovery guidance.
+
+Evidence:
+
+- `dicklesworthstone/cass@a18a6261a9815204b6ef59f9d186370ebd35dc34:src/schedule.rs#L207-L222`
+- `dicklesworthstone/cass@a18a6261a9815204b6ef59f9d186370ebd35dc34:src/indexer/lexical_reconcile.rs#L213-L249`
+
+Adapt lifecycle rendering and recovery validation, not the entire separate
+Tantivy pipeline. The inspected shape check is not a cryptographic content hash.
+Ledger should keep lexical retrieval available when optional model work fails.
+That last sentence is a Ledger requirement, not verified Cass behavior here.
+
+### AgentsView: Dream needs invalidation, not only scheduled extraction
+
+The extraction pass reconciles newly ineligible sessions before model work, so
+endpoint failure cannot indefinitely postpone privacy retraction. Building
+generations are staged, a second reconciliation catches mid-pass changes, and
+only unrestricted passes advance the scan watermark. This directly addresses
+the risk of a batch limit permanently hiding older eligible sessions.
+
+Evidence: `kenn-io/agentsview@2298453833a5c5126a486a2554eac6c35cd2b55f:internal/recall/extract/manager.go#L227-L335`.
+
+A test checks that appending outside the evidenced range preserves provenance
+and that evidence carries a content digest:
+`kenn-io/agentsview@2298453833a5c5126a486a2554eac6c35cd2b55f:internal/recall/extract/manager_test.go#L1800-L1822`.
+
+Adapt versioned findings, digest-bound evidence, staged publication, eligibility
+rechecks and retraction. Do not treat generated findings as authoritative memory.
+The inspected watermark assignment is in-memory; it does not by itself prove
+crash-safe progress. Ledger needs explicit persisted job state.
+
+### AI SDK Harnesses: distinguish suspension from cancellation
+
+The sandbox channel's suspend operation stops dispatch/reconnect and returns an
+event cursor; the documented bridge continues the in-flight turn. A pinned
+checkpoint can replay an already delivered tail. The corresponding test checks
+that releasing a checkpoint returns suspension to the latest cursor.
+
+Evidence:
+
+- `vercel/ai@c3c189c068baefe62808b4eca5432fe092dc37cb:packages/harness/src/utils/sandbox-channel.ts#L290-L325`
+- `vercel/ai@c3c189c068baefe62808b4eca5432fe092dc37cb:packages/harness/src/utils/sandbox-channel.test.ts#L192-L212`
+
+Adapt explicit event IDs, replay dedup and lifecycle meanings for live capture.
+Do not equate disconnection with cancellation, or an event cursor with committed
+Ledger storage. This channel code does not establish isolation strength or
+complete child-agent capture across every adapter.
+
+## Ordered implementation gates
+
+1. **Ingestion correctness:** adapter-specific discovery fixtures including
+   JSONL for non-Cursor directory sources; overlapping-root dedup; bounded,
+   resumable backfill. Compare canonical output with whole-file normalization;
+   measure real-corpus peak RSS. A recent-session limit is not full coverage.
+2. **Durable ownership:** one participating writer; records and input progress
+   commit atomically. Kill/restart during a batch, race CLI against daemon, and
+   test truncation, replacement, late records and normalization-version changes.
+3. **Operational application:** stable paths and lifecycle commands, escaped
+   launchd/systemd definitions, periodic reconciliation, visible failed/pending
+   coverage. Verify actual login/reboot restart and upgrades on both OS families.
+4. **Independent retrieval:** prove lexical search with provider credentials
+   absent and embeddings failing; reindex derived data from retained evidence.
+   Use the known Netcup session as a discovery regression, without preselecting
+   its path or requiring a remembered harness.
+5. **Dream:** durable jobs with leases, bounded input, input/config/model versions,
+   retry and idempotent findings; recheck evidence at publication and retract
+   stale/excluded evidence. Kill during generation/publication; verify no lost
+   input or duplicated findings. Promotion remains a separate authorized action.
+6. **Fleet:** prove machine identity and explicit per-machine coverage before
+   claiming cross-machine search. Define synchronization/privacy policy before
+   moving any private transcripts off a machine.
+
+## Remaining research gaps
+
+Follow complete watcher callers and shutdown paths; run cross-process lease and
+crash recovery tests; inspect every adapter's identity/partial-input behavior;
+verify installed package APIs and licenses before adoption. Model-facing search
+contracts, extension points and isolation were not uniformly audited across all
+five projects. The agent report is a lead inventory, not evidence closing those
+gaps. No replacement for Letta is justified by this comparison.

package/docs/research/ledger/docs/research/2026-09-07-letta-continuous-service.md

@@ -0,0 +1,104 @@
+# Letta Code continuous-service implementation
+
+Inspected public `letta-ai/letta-code` at commit
+`701f2a5367828847313876c735ade27b9df97689` on 2026-09-07 via GitHub's contents API.
+Confer's previously inspected snapshot
+`e0d97b88dfc57c170845a906875ff7ac0ae29d60e185d031e6ec55cdfc5f6814`
+contains Trajectory but not Letta Code, so this follow-up uses pinned upstream
+files rather than a corpus clone. No upstream code was copied or executed.
+
+## Service, scheduler and reflection are separate owners
+
+1. `letta server` routes to the remote listener; `letta server --listen` routes
+   to the local App Server. They are different modes, not synonyms. The App
+   Server defaults to loopback and exposes explicit authentication configuration
+   for broader access.
+   [server.ts:8-41,109-127](https://github.com/letta-ai/letta-code/blob/701f2a5367828847313876c735ade27b9df97689/src/cli/subcommands/server.ts#L8-L41)
+2. The App Server CLI starts a server handle, then waits for SIGINT/SIGTERM.
+   An idempotent shutdown guard calls the handle's asynchronous close method;
+   close failures produce a failing exit result. This is a foreground process
+   suitable for OS supervision, not proof of an installed service.
+   [app-server.ts:32-53,118-139](https://github.com/letta-ai/letta-code/blob/701f2a5367828847313876c735ade27b9df97689/src/cli/subcommands/app-server.ts#L32-L53)
+3. Listener startup installs process event routing and starts cron with a
+   process transport, guarded by process-service generation state. Cron can be
+   disabled with `LETTA_DISABLE_CRON_SCHEDULER=1`. Do not infer ownership solely
+   from scheduler comments saying "when the WS listener connects": the caller
+   distinguishes connection state from process services.
+   [lifecycle.ts:375-399,460-497](https://github.com/letta-ai/letta-code/blob/701f2a5367828847313876c735ade27b9df97689/src/websocket/listener/lifecycle.ts#L460-L497)
+
+## Scheduling persistence and failure handling
+
+`crons.json` writes use a flushed temporary file and rename. A lock protects
+scheduler lease claims. The stored owner includes PID, token and process identity;
+a live other process prevents acquisition. Lease verification compares both PID
+and token. Linux process identity includes process start ticks and boot ID.
+This is more explicit than assuming a PID alone identifies an owner, but is not
+proof of identical crash/power-loss guarantees on every filesystem or OS.
+
+[cron-file.ts:181-221,646-682](https://github.com/letta-ai/letta-code/blob/701f2a5367828847313876c735ade27b9df97689/src/cron/cron-file.ts#L646-L682)
+
+The scheduler verifies its lease on each tick, refreshes cached tasks, deduplicates
+within the current minute and delays recurring tasks with jitter. Before delayed
+execution it rereads the task and checks active status. Failure paths persist a
+failed outcome and append run-log information. Startup retries lease acquisition
+at bounded intervals, then warns that schedules will not fire. Stop clears tick,
+GC and pending jitter timers and attempts lease release.
+
+[scheduler.ts:498-712](https://github.com/letta-ai/letta-code/blob/701f2a5367828847313876c735ade27b9df97689/src/cron/scheduler.ts#L498-L712)
+
+Limits worth testing before adopting the design:
+
+- The per-minute fired set is in memory; it does not alone prove durable
+  occurrence deduplication or exactly-once effects.
+- Lease-acquisition retry timeouts are scheduled separately from the tracked
+  pending jitter timers. Cancellation during acquisition needs its own test.
+- The delayed-fire block checks runtime/task state but does not itself verify
+  the lease token again. Full fencing requires tracing downstream dispatch.
+- Clearing timers does not by itself cancel or checkpoint an already-running
+  agent. Shutdown/drain and run recovery need separate contracts.
+
+These are source-level review questions, not reproduced production failures.
+
+## Reflection is not just cron
+
+The post-turn helper explicitly requires transcript delta capture to precede
+trigger evaluation. It handles `off`, `compaction-event`, and `step-count`, using
+steps since the last successful reflection for the latter.
+[post-turn-reflection.ts:17-77](https://github.com/letta-ai/letta-code/blob/701f2a5367828847313876c735ade27b9df97689/src/cli/helpers/post-turn-reflection.ts#L17-L77)
+
+The launcher refuses work when MemFS is disabled, when Cloud cutover assigns
+reflection to the server, when automatic launches are suppressed by configuration
+failure, or when another launch is active. Dirty parent memory and empty payloads
+also produce explicit skip reasons. It prepares a memory worktree for the agent.
+[reflection-launcher.ts:712-795](https://github.com/letta-ai/letta-code/blob/701f2a5367828847313876c735ade27b9df97689/src/cli/helpers/reflection-launcher.ts#L712-L795)
+
+Completion first finalizes worktree integration, then passes `completionSuccess`
+to transcript completion, and releases the launch reservation in `finally`.
+Thus subagent success is not simply treated as successful memory integration.
+The client cutover guard proves an ownership interface exists, not the deployed
+Cloud scheduler's implementation or rollout status.
+[reflection-launcher.ts:847-883](https://github.com/letta-ai/letta-code/blob/701f2a5367828847313876c735ade27b9df97689/src/cli/helpers/reflection-launcher.ts#L847-L883)
+
+## Tests and transfer to Ledger
+
+Located tests for App Server bind failure/retry and shared-runtime ownership,
+scheduler invalid-cron reporting and missed one-shot logs, and reflection cutover,
+dirty-parent rejection and configuration-failure suppression. Test declarations
+were inspected; this was not an upstream test run or full behavioral audit.
+
+- [App Server lifecycle tests](https://github.com/letta-ai/letta-code/blob/701f2a5367828847313876c735ade27b9df97689/src/websocket/app-server-lifecycle.test.ts)
+- [Scheduler tests](https://github.com/letta-ai/letta-code/blob/701f2a5367828847313876c735ade27b9df97689/src/cron/scheduler.test.ts)
+- [Reflection launcher tests](https://github.com/letta-ai/letta-code/blob/701f2a5367828847313876c735ade27b9df97689/src/cli/helpers/reflection-launcher.test.ts)
+
+Adopt explicit process ownership, observable skipped/failed runs, cancellation
+boundaries, and an integration-success gate before advancing analysis progress.
+Keep Ledger ingestion independent of relay connectivity or an active chat.
+Use durable job/occurrence identities and transactional progress rather than
+copying in-memory dedup or reservation assumptions. Preserve separate authority
+for promoting findings into memory or harness changes.
+
+No evidence here says a daemon is too large. No RSS/startup benchmark was run.
+This server hosts agent work; it does not prove continuous discovery/backfill of
+all local harness transcripts. A fleet-wide transcript sync protocol, Cloud
+reflection backend, OS service installers and all shutdown callees remain outside
+this targeted source pass.

package/docs/research/ledger/docs/research/2026-09-08-claude-history-retrieval.md

@@ -0,0 +1,117 @@
+# Bounded Claude history retrieval probe
+
+## Question and fixture
+
+Can Ledger recover the original June 26 request to create the ACTI eval developer
+skill after indexing a small historical slice?
+
+The source was Claude's native `~/.claude/history.jsonl`, session
+`e880706f-0c24-4ff4-9955-4f6ccab9b6c8`. The retained session transcript begins on
+July 22 and has no June 26 records; widening the normal session scan alone would
+not recover this prompt. The installed trajectory 0.3.0 Claude decoder accepts
+user-message envelopes, not history rows with `display` fields.
+
+### Follow-up: why the retained transcript starts in July
+
+On September 9, a bounded native-history search found a specific recovery request
+in Claude session `74636afc-517d-4359-8c8d-27daf41145fe`:
+
+- July 23, 14:52:04 UTC, history line 22778: locate the `e880706f` transcript.
+- 14:53:17 UTC, line 22780: the user reported it was corrupted and could not resume.
+- 14:59:51 UTC, line 22781: the user requested a wholesale backup followed by
+  keeping only the last two days of transcript to test resumption.
+- 15:13:12 UTC, line 22788: the user asked where the backup was.
+
+The current transcript's filesystem birth time is July 23, 15:01:57 UTC, its
+earliest retained message is July 22, 01:22:38 UTC, and that first message's parent
+record is absent. This strongly supports deliberate trimming during the recovery
+attempt. The actual command log and wholesale backup have not been located, so
+do not claim a verified deletion mechanism, blame automatic retention, or describe
+the older conversation as permanently lost. The prompt-history copy survived.
+
+Selection was frozen before import: June 26 in China time, represented as
+`2026-06-25T16:00:00Z <= timestamp < 2026-06-26T16:00:00Z`, one session,
+78 nonempty prompts, 26,293 raw bytes including line endings. Snapshot SHA-256:
+`6c16e022c970bf875b3c6eceb6a7870cb7fc87dbaf669f4a0fbbff5d028bd584`.
+
+## Import and provenance
+
+`packages/core/scripts/import-claude-history.ts` maps each selected prompt through
+the public trajectory canonicalizer with partial-transcript semantics. It neither
+invents assistant turns nor edits source artifacts. Original byte offsets anchor
+location-derived identities; an external private manifest and exact raw snapshot
+preserve lineage omitted by the canonicalizer.
+
+The supplemental slice has a namespaced ID derived from the source path, session,
+and UTC bounds. It does not replace the existing session. This is a scoped import
+utility, not default scanner/watcher integration or a general merging policy for
+overlapping windows. The normal Ledger CLI and search wrapper query its FTS data.
+
+## Results
+
+Before import, none of the frozen queries retrieved the exact June 26 prompt.
+After import, with source `claude-code` and limit 20:
+
+| Frozen query | Exact prompt rank |
+| --- | --- |
+| `"we should also make a acti-eval skill"` | 1 |
+| `eval skill` | 1 |
+| `benchflow` | 1 |
+| `eval best practice` | Not found |
+| Unique unrelated negative-control token | No hits |
+
+The initial all-positive-queries acceptance therefore **failed**; the failed
+record was retained. Source inspection explained the miss: `buildFtsExpression`
+quotes plain multiword input as one contiguous phrase. Follow-up queries, marked
+post hoc rather than replacements for the frozen workload:
+
+- `eval AND best AND practice`: exact prompt rank 1.
+- `eval`: exact prompt rank 5.
+
+A further scope check without the Claude source filter found the prompt at rank
+17 for `eval skill`; bare `eval` did not retrieve it in the top 20 across all
+sources. Coverage and ranking are separate concerns: more indexing alone does
+not establish useful top-k ranking for vague queries.
+
+All 78 stored prompt texts and timestamps matched their source. The target remains
+at `2026-06-26T11:44:16.686Z`, original history line 20775. The existing full-session
+metadata was unchanged. Database records increased from 68,223 to 68,301; session
+rows increased from 189 to 190 (the additional row is a partial slice, not a newly
+discovered complete conversation). A repeat import deduplicated; SQLite
+`quick_check` returned `ok`.
+
+## Cost and limits
+
+One measured import on Charless-MacBook-Air, Bun 1.4.0, with existing OS caches:
+
+- Child launch through exit: 0.0586 seconds.
+- Import-body CPU: 0.0471 CPU-seconds.
+- Child RSS before import: 48,611,328 bytes; lifetime maximum: 84,754,432 bytes.
+- No LLM or embedding model called; token and model-cache counters do not apply.
+- A separate repeat checked idempotence, not stable performance.
+
+Peak sampled CPU, physical I/O bytes and post-exit memory were not measured.
+The evidence validator reports structurally valid but incomplete resource evidence.
+No throughput, p95, whole-corpus recall, or full-index capacity claim follows from
+this one-message probe.
+
+The result supports broader source coverage: indexed prompt evidence is retrievable
+through Ledger. Full indexing must include Claude prompt history as well as session
+transcripts and account for phrase-search behavior. Full backfill and daemon startup
+were not performed.
+
+## Replay and artifacts
+
+See `ops/README.md` for the bounded importer command. Reuse the selected session and
+UTC bounds above, then query through `skills/ledger/scripts/ledger_search.sh` with
+the intended `--db`. To repeat RED/GREEN, use a separate copy of the pre-import
+database; never overwrite the active database with the older backup.
+
+Private local evidence is in `.local/claude-history-probe/`: `spec.json`,
+`pre-import.sqlite`, `source/selected-history.jsonl`, `source/provenance.json`,
+`attempt-1.json`, `attempt-2.json`, `verdict.json`, `followup.json`, and `run.json`.
+The report preserves the initial failed query expectation. Raw prompts/databases
+are excluded from Git; only this compact result, importer, and tests are committed.
+
+Validation: focused history-import tests (2 passed, 16 assertions), core typecheck,
+real before/after import and search, idempotent reimport, and SQLite integrity check.

package/docs/research/ledger/docs/research/2026-09-09-agentsview-native-discovery.md

@@ -0,0 +1,87 @@
+# AgentsView Native Discovery
+
+Audited **2026-09-09, 02:10 UTC**: GitHub [main][branch] resolved to
+[`e1e41fd0e31c41b7a6a0cd76c1f61a0cb2da4147`][head] (commit 2026-09-09T01:42:46Z).
+The pinned [LICENSE][license] is MIT, copyright Kenn Software LLC.
+
+## Native Readers
+
+**Cursor IDE:** [globalStorage/state.vscdb][ide] contains
+`cursorDiskKV(key,value)`: `composerData:<id>` supplies
+`fullConversationHeadersOnly[].bubbleId`, ordered references to
+`bubbleId:<composerId>:<bubbleId>`. Bubble types 1/2 mean user/assistant;
+`toolFormerData` carries tool details. Session identity is
+`cursor-ide:<composerId>`; provenance is `<db>#<composerId>`; message identity
+is the bubble UUID. Neither Cursor reader emits session parentage.
+
+**Cursor transcripts:** [discovery][cursor-discovery] reads
+`~/.cursor/projects/<project>/agent-transcripts/<id>.{txt,jsonl}` and
+`<id>/<id>.{txt,jsonl}`, preferring JSONL. The [reader][cursor-reader] accepts
+text or JSONL `role` plus `message.content`; identity is `cursor:<filename-stem>`.
+This is separate from IDE identity, not automatic cross-store deduplication.
+
+**OpenCode:** [discovery][discovery] accepts `opencode.db` and
+`opencode-<channel>.db` under the [configured root][catalog]
+(default `~/.local/share/opencode`). [SQLite assembly][opencode] reads
+`session/project/message/part`, JSON `data`, and message/part `session_id`
+plus part `message_id`. [Legacy storage][legacy] assembles
+`storage/session/<project>/<sid>.json`, `storage/message/<sid>/*.json`,
+and `storage/part/<mid>/*.json`. [Storage wins duplicate IDs][oc-discovery];
+SQLite contributes missing IDs. Identity is `opencode:<id>`; explicit
+`parent_id` or legacy `parentID` supplies parentage. SQL provenance is
+`<db>#<id>`.
+
+## Freshness And Limits
+
+Cursor IDE uses a read-only connection and one [read transaction][ide] for
+composer, bubbles, and their content digest. Missing/empty conversational bubbles
+mark truncation; malformed JSON errors rather than masquerading as missing.
+Its [container fingerprint][ide-provider] includes file identity, SQLite header,
+and WAL size/header; member digests cover raw composer/bubble bytes.
+It watches `state.vscdb` and `state.vscdb-*` shallowly; transcripts use recursive
+TXT/JSONL watches.
+
+OpenCode uses [SQLite/WAL state][sqlite-state] and [per-session fingerprints][oc-freshness]:
+session/project/child update times, child counts, and ordered child-ID/timestamp
+digests. This discovery digest is not a body-content hash. [Parsing][oc-parse]
+hashes content but uses separate autocommit queries, capturing the watermark
+before children, not a single snapshot. Legacy [stat gates][storage-state] cover
+session/message/part files; same-size, timestamp-preserving edits require watcher
+invalidation.
+
+[Watch plans][oc-watch] separate shallow DB/WAL coverage from recursive storage
+when possible; [event filtering][oc-events] ignores SHM and empty WALs.
+[Fast watermark passes][gate] deliberately defer child-only edits to full
+discovery; [digest verification expires after five minutes][verify].
+
+## Ledger Implication
+
+Beyond the [Cass format inventory](2026-09-09-native-history-discovery.md),
+useful lessons are snapshot-aligned digests, gap/error distinction,
+deletion-sensitive fingerprints, and separate live-database/archive watch budgets.
+This is complementary evidence, not an exhaustive Cass feature comparison.
+Use selected chat fields: [maintainers flag credential-bearing Cursor fields][catalog].
+
+Public-source audit only: no clone, private-store access, runtime execution,
+or upstream code copying.
+
+[branch]: https://api.github.com/repos/kenn-io/agentsview/branches/main
+[head]: https://github.com/kenn-io/agentsview/commit/e1e41fd0e31c41b7a6a0cd76c1f61a0cb2da4147
+[license]: https://github.com/kenn-io/agentsview/blob/e1e41fd0e31c41b7a6a0cd76c1f61a0cb2da4147/LICENSE
+[ide]: https://github.com/kenn-io/agentsview/blob/e1e41fd0e31c41b7a6a0cd76c1f61a0cb2da4147/internal/parser/cursor_ide.go#L17-L536
+[cursor-discovery]: https://github.com/kenn-io/agentsview/blob/e1e41fd0e31c41b7a6a0cd76c1f61a0cb2da4147/internal/parser/cursor_provider.go#L417-L552
+[cursor-reader]: https://github.com/kenn-io/agentsview/blob/e1e41fd0e31c41b7a6a0cd76c1f61a0cb2da4147/internal/parser/cursor.go#L496-L589
+[discovery]: https://github.com/kenn-io/agentsview/blob/e1e41fd0e31c41b7a6a0cd76c1f61a0cb2da4147/internal/parser/discovery.go#L96-L212
+[catalog]: https://github.com/kenn-io/agentsview/blob/e1e41fd0e31c41b7a6a0cd76c1f61a0cb2da4147/internal/parser/types.go#L257-L352
+[opencode]: https://github.com/kenn-io/agentsview/blob/e1e41fd0e31c41b7a6a0cd76c1f61a0cb2da4147/internal/parser/opencode.go#L987-L1319
+[oc-discovery]: https://github.com/kenn-io/agentsview/blob/e1e41fd0e31c41b7a6a0cd76c1f61a0cb2da4147/internal/parser/opencode_provider.go#L569-L711
+[ide-provider]: https://github.com/kenn-io/agentsview/blob/e1e41fd0e31c41b7a6a0cd76c1f61a0cb2da4147/internal/parser/cursor_ide_provider.go#L52-L192
+[sqlite-state]: https://github.com/kenn-io/agentsview/blob/e1e41fd0e31c41b7a6a0cd76c1f61a0cb2da4147/internal/parser/sqlite_container_state.go#L15-L143
+[oc-freshness]: https://github.com/kenn-io/agentsview/blob/e1e41fd0e31c41b7a6a0cd76c1f61a0cb2da4147/internal/parser/opencode.go#L764-L847
+[oc-parse]: https://github.com/kenn-io/agentsview/blob/e1e41fd0e31c41b7a6a0cd76c1f61a0cb2da4147/internal/parser/opencode.go#L1151-L1207
+[storage-state]: https://github.com/kenn-io/agentsview/blob/e1e41fd0e31c41b7a6a0cd76c1f61a0cb2da4147/internal/parser/opencode_storage_state.go#L15-L127
+[oc-watch]: https://github.com/kenn-io/agentsview/blob/e1e41fd0e31c41b7a6a0cd76c1f61a0cb2da4147/internal/parser/opencode_provider.go#L714-L778
+[oc-events]: https://github.com/kenn-io/agentsview/blob/e1e41fd0e31c41b7a6a0cd76c1f61a0cb2da4147/internal/parser/opencode_provider.go#L2056-L2112
+[gate]: https://github.com/kenn-io/agentsview/blob/e1e41fd0e31c41b7a6a0cd76c1f61a0cb2da4147/internal/sync/opencode_container_gate.go#L257-L378
+[legacy]: https://github.com/kenn-io/agentsview/blob/e1e41fd0e31c41b7a6a0cd76c1f61a0cb2da4147/internal/parser/opencode.go#L1663-L1918
+[verify]: https://github.com/kenn-io/agentsview/blob/e1e41fd0e31c41b7a6a0cd76c1f61a0cb2da4147/internal/sync/opencode_container_gate.go#L17-L43