# @qiuxu/memex offline harness · Node v24.21.0 · 9 source packages + bundle + 9 /api/memory.* routes + web client + sticky L2/L3 + L1 eviction + cold tier/compression + hybrid recall/cold abstracts + sensitive evict + backup · stubs: cordis, schemastery, dsh-home-paths, dsh-llm, dsh-tools, dsh-compaction
ok A1 write then get returns the stored card fields
ok A2 record file exists on disk with YAML frontmatter
ok A3 get returns the exact body that was written
ok B1 search finds a 4-char CJK term in the title
ok B2 search finds a 4-char CJK term that only exists in the body
ok B3 search falls back to LIKE for the 2-char term AI
ok B4 search returns no hits for nonsense
ok C0a search only returns records of the requested workspace (A)
ok C0b search only returns records of the requested workspace (B)
ok D1 list filters by kind
ok D1b search applies the kind filter to every hit path
ok D2 list filters by status
ok D2b search applies the status filter to every hit path
ok D3 search filters by tag
ok D4 grep finds literal lines inside stored files
ok D5 facets reports counts, kinds, statuses and capacity
ok D6 catalog renders a bounded markdown catalog
ok D7 forget removes the row, the FTS entry and the markdown file
ok D8 setStatus updates lifecycle state and catalog pinning
ok D9 markRead records read usage
ok E1 archiveRegion persists the card, file and idempotency marker
ok E2 archiveRegion is idempotent for the same session and compaction
ok F1 checkpoint archive creates one record for the workspace key
ok F2 memory_open-style tool read returns the archived body
ok F3 pre-step injects a checkpoint receipt for the archived record
ok F4 pre-step injects the memory catalog message
ok F5 pre-step catalog is deduped by digest on the second run
ok F6 reconcile archives a persisted compaction discovered through sessionQuery
ok G1 pre-step auto-recalls a relevant record for the current user text
ok G2 pre-step does not repeat the recall for identical user text
ok H1 llm refinement replaces the deterministic card
ok H2 cardModel off keeps the deterministic card
ok I1 a second LocalStoreCore on the same root still searches and gets
ok J1 unknown record kind throws
ok J2 non-string body is coerced to a string
ok J2b body size guard is published as 8 MiB (oversize allocation intentionally skipped)
ok J3 grep rejects invalid patterns with a clear message
ok J4 corrupt record input throws
ok K1 backend row alone owns ctx.memory and registers no surfaces
ok K2 checkpoint row alone registers the policy listeners and no surfaces
ok K3 tools row registers exactly six memory_* tools and one tool:memory section
ok K4 commands row registers /memory with sticky-level control and is inert without one
ok K5 consumer modules use ctx.memory only, and the loader stubs only @deepseek-ai/*
ok K6 generated bundle modules mount as four rows end-to-end
ok K7 client module registers composer trigger + shell overlay panel slots
ok K8 bundle patch declares eight rows including memory-compress
ok L1a memory API registers eight exact Fetch routes through ctx.effect
ok L1b index route returns levels, both pools, evictions and the active fields
ok L1c record route returns records visible through the enabled levels only
ok L1d search route returns ranked snippets over the visible-level union
ok L1e status route pins and unpins records visible in the enabled levels
ok L1f API returns coded JSON errors for unknown sessions
ok L1g index/search limit fallbacks apply when ?limit is absent
ok M1 default levels none archive nothing, inject nothing and gate tools/commands
ok M2 L2 archives privately and stays invisible to another session in the same workspace
ok M2b record route exposes L2 records only to their own session
ok M3 L3 shares archives across sessions and keeps L2 records private
ok M4 facets, /memory status and the index API expose both pools and L1 evictions
ok M5 level API enables sticky levels and immediately gates the visible pool
ok M6a client ring bridge hides the arc and labels extra context off
ok M6b client ring bridge colors and labels the L2 vs L3 capture target
ok M6c client panel source exposes the sticky level switches and L1 eviction footer
ok M7 grep applies the enabled-level visibility union
ok M8 search body fallback respects the enabled-level scopes
ok M9 search works with L2 and L3 both enabled
ok N1 fresh backend defaults to no levels and L2/L3 are sticky once enabled
ok N2 level API enables idempotently and rejects disabling
ok N3 L2 stays private end-to-end (archive, index, record, tools)
ok N4 L3 wins the capture scope while L2 records stay session-private
ok N5 memory-evict moves the oldest balanced L1 span into the enabled level
ok N6 bundle artifact ships the memory-evict/compress modules and eight-row patch
ok C1 cold tier releases hot capacity while keeping records retrievable
ok C1b catalog footer reports the cold tier
ok C2 memory_compress cold strategy releases hot capacity losslessly
ok C2b memory_compress gates off sessions and disabled levels
ok C2c memory_compress respects minAgeMs for very recent records
ok C3a memory_compress digest runs bounded llm batches and cold-moves originals
ok C3b memory_compress digest is bounded and leaves originals hot on failure
ok C4a capacity notice fires above the hot ratio and stays silent below it
ok C4b capacity notice lists both hot levels and respects the cooldown
ok C5a L1 eviction never crosses an image-bearing event
ok C5b a single oversized event stops the eviction batch
ok C5c captureLevel l2 keeps L1 evictions session-private with L3 enabled
ok C6 bundle/module/client/manual cover the compress + cold-tier wave
ok R1 FTS indexes card fields only; body search uses the hot-file fallback and grep
ok R2 restoreCold moves files back hot and gcCold deletes the oldest cold records
ok R3 compressCold cold-spills the oldest eligible records to the target ratio
ok R4a recordCompression/compressionStats persist per-session history
ok R4b history surfaces in /memory status, memory_index and the index API
ok R5 new API routes: compress, restore and gc
ok R6a client source carries the compression / restore / capacity-warning UI
ok R6b recallTerms builds the search query and the capacity notice is bilingual
ok R7 record route paginates the body and reports truncation flags
ok S1 memoryQueryTerms extracts identifiers, quoted spans and CJK n-grams
ok S2 hybrid search recalls cold abstracts, hot bodies and ranks field matches
ok S3 GET /api/memory.metrics reports pool/cold/eviction/compression metrics
ok S4 sensitive eviction forces L2, skips under L3-only and honors sensitivePolicy:off
ok S5 memory-backup backup/verify/restore round-trips a store with a tamper check
ok U1 agent-agnostic core has no host dependencies and the DSH seam re-exports it
ok U2 model profiles resolve per family and tune injection, budgets and digest prompts
ok U3 generic MCP server speaks JSON-RPC: tools list/call, search, open, index
ok U4 MCP cold compression frees capacity and cold records remain openable
ok U5 Claude Code init writes .mcp.json, hooks, and a model-tailored CLAUDE.md
ok U6 Claude Code hooks inject catalog/recall and archive transcripts before compaction
ok U7 Codex init writes AGENTS.md and prints a working MCP config snippet
ok U8 DSH seam and core stay one vocabulary: engine writes are visible to DSH search
ok U9 per-model rerank weights tune ranking and merge over defaults
ok U10 engine applies the model profile rerank and accepts per-integration overrides
ok U11 PreCompact archives transcripts larger than the store body limit instead of failing
ok U12 injected catalog/recall blocks stay inside each profile budget (wrapper included)
ok U13 body-term rescue recovers facts that live only in hot bodies
ok U14 MCP parity tools: list, set-status and forget operate on the shared store
ok U16 dedicated GPT-6.1 / Opus 5.5 / Sonnet 5.5 / DeepSeek Flash profiles win over family bases
ok U17 adapter CLIs default to recommended dedicated profiles and advertise them
ok U18 storage root resolution: MEMEX_HOME wins, legacy env/path fallbacks work
ok U19 Claude init --session aligns hooks and MCP L2 namespace; approval hint is printed
ok U20 cold natural-language retrieval uses card/bare-number anchors (turn 15)
ok U21 hot body rescue: a bare turn number anchors a body-only fact
ok U22 anchor-free natural-language recall scans bounded hot bodies
ok U23 cold anchor-free recall pages a fully covered cold abstract
ok U24 cold morphology-lite and partial coverage recall (decide/decision + extra term)
ok U25 live A/B tail follows insertion order when timestamps tie
ok U26 context materializer: chronology, stable prefix, budget, specialist slot
ok U27 canonical context events round-trip with checksums and store persistence
ok U28 context pager replays canonical events through the materializer
ok U29 context encoder baseline ranks overlapping pages and trains offline
ok U30 KV host capability negotiation fails closed and pages when fully capable
ok U31 trained transformer context encoder matches Python reference and ranks pages
ok U32 reference paged-KV host pages blocks, caches prefixes and evicts by capacity
ok U33 checkpoint runtime injects a materialized window when enabled
ok U34 GPU-trained 2-layer transformer encoder matches Python logits in JS

# 131/131 passed