Eidetic Engine 0.18.0

Eidetic Engine 0.18.0

[0.18.0] - 2026-10-14

Covers v0.17.0..35b0bc06b: 931 commits (924 non-merge: 60 features, 205
fixes, 19 performance changes; 455 tracker bookkeeping), plus three fixes made
in the 2026-10-14 release wave (GH #70, #67, #68). Imported CASS history
becomes readable evidence that teaches ee how past failures were fixed.
Writes publish incremental index updates again, and packs rank all sources.

Upgrade notes

• Forward-only migrations V127–V131: evidence_admission_verdicts,
  pack_rule_items, pack_guard_triggers_without_joins,
  evidence_reader_projections and evidence_admission_counts. V127, V130 and
  V131 backfill existing evidence. There is no down-migration: once a 0.18.0
  command has migrated a store, ee 0.17.0 reports it as storage_unavailable
  ("could not be opened or needs migration"). Run ee backup create first if
  you may need 0.17.0 again.
• Rebuild indexes: run ee migrate run --workspace . --json, then
  ee index rebuild --workspace . --json. Frankensearch 0.7.1 changes the
  producer revision, so even default-model semantic indexes need re-embedding,
  and projection revision 4 marks indexes stale. Sources stored only as
  external_ingestion_oversized markers need a CASS re-import.
• The error-recall derivation is now v9, so the next ee import cass
  re-processes old sessions. Pack cache keys and policy hashes change, so new
  packs can differ from 0.17.0 output; historical packs still replay.
• ee.import.cass.v1 adds a required evidenceAdmission object.
  status can be completed_with_refusals. sessions[].status can be
  refused, with refusalCode and repair. Refused sessions count toward
  sessionsSkipped and raise cass_import_history_refused (high).
• Pack items can be rules: entityKind: "rule", section
  procedural_rules, provenance ee://rule/<id>. A rule with a source memory
  now carries its own id instead of that memory's, so
  ee outcome --pack <id> --item <n> credits the rule.
• Config paths: the user layer is $XDG_CONFIG_HOME/ee/config.toml when
  that root is absolute (HOME is then ignored). Storage defaults prefer an
  absolute $XDG_DATA_HOME/ee. Empty or relative XDG roots are ignored.
• ee doctor reports malformed .ee/config.toml as EE-E101 (blocked), and
  --fix quarantines it by renaming. ee.daemon.write rejects unknown
  parameters and a confidence outside 0–1. It accepts null optional strings
  and the new dryRun, validFrom and validTo.
• ee model fetch <model> now selects the model in initialized workspaces;
  reembed and restart the daemon after switching. On Unix, new stores are
  created 0600. Nothing in this cut is specific to Windows.

CASS import and refresh

• [cass] binary in the user config is honored by import, status,
  capabilities and doctor; EE_CASS_BINARY still wins (873ed8b4f). A
  workspace's own .ee/config.toml cannot set it, because ee import cass
  executes that path and a cloned repository must not choose it (GH #67). On
  Unix, ee also accepts a cass in its own bin directory, or in the paired
  .cargo/bin or .local/bin. The path must be owned by root or ee's owner
  and not be group- or world-writable (e100b74e7, 5cd1e2dd7).
• Each import reports admitted and quarantined span counts, with each span's
  first policy reason (e518c5d3f). Refresh checks retained history against
  full source-line commitments. Changed, missing, unverifiable or still-growing
  sessions are refused while the rest of the batch imports (GH #65, #66, #77;
  d40699c43, b645122c9).
• Sessions imported by 0.17.0 keep importing after the upgrade (GH #70).
  Claude Code and Codex transcript lines are compact JSON, and a compact-JSON
  excerpt stored by 0.17.0 has no source commitment, so every such session was
  refused as cass_refresh_history_unverifiable on each run and never received
  its new lines. Such an excerpt is now kept when today's cass view line is
  byte-for-byte the stored excerpt (same BLAKE3 hash and length): nothing
  outside it is left to authenticate. An excerpt that differs from today's
  source is still refused.

Failure-to-fix recall and session learning

• ee import cass pairs each failed run with a later successful retry of the
  same invocation, and records error fingerprints linked to the repair and
  proof (55aa95673). ee diagnose-error returns data.repairEvidence.
  ee pack --error-log adds up to 4 fix excerpts (4096 bytes) to retrieval
  (359b78c21).
• Each resolved failure also produces an extractive incident card of at
  most 120 tokens: symptom, fix and verifying command. Search and packs use it
  instead of the raw turns (ADR 0091, 439a47667).
• Learnable output: rustc/Cargo (text and JSON), libtest, Pyright, ESLint,
  Jest/Vitest, go test -json, pytest, unittest, Python tracebacks, Node TAP,
  ee/RCH structured failures and long-running Codex exec output.
• A repair counts only when the same command, cwd, environment and filters
  succeed later, and every failed test passes in the same test list. Narration
  is never linked as the fix (81719ab5e, 353b7b156, f70c68464).
• ee review session --propose keeps labeled lessons, drops task requests and
  names topics by subject; recall on the internal test corpus rose from 0.67 to
  0.83 (4667de09b). Session-arc confidence is 0.6/0.7/0.8 for one, two or
  three independent sessions, replacing a fixed 0.82 (9236f211b).
• Pairing reads test counters and exit codes (21 passed, 0 failed is a
  success) and checks the windows that follow, so a failed retry never becomes
  a lesson (296ea999d, 505fa9fb6, a721ca31a).

Readable CASS evidence

• CASS readers share a bounded, typed projection of Claude and Codex records.
  Search, packs and extractive answers use decoded message text with each
  record's own role, and summaries are labeled. Reasoning, analysis-channel
  records and envelope metadata cost no reader tokens. Records without safe,
  readable text are excluded rather than shown as raw JSON (bc90a61f6).
• Migration V130 stores reader projections and egress-screen results, bound to
  exact source, version and content. Writes, recovery, migration and rebuild
  fill them in. Missing or stale values are computed in memory, with no writes
  at read time. Source identity, reader bytes and pack hashes are unchanged.
• Admission revision 3 invalidates older cached verdicts, including unreadable
  records that were admitted before. Projection revision 4 marks indexes stale
  so a rebuild applies the new text. Stored excerpts, hashes, ids and line
  locators are unchanged
  (migration notes).
• Escaped-Unicode JSONL windows are screened as complete, ordered records.
  Clean windows keep their bytes, and redaction keeps each record's role and
  kind. Bootstrap and linked-session learning use the stricter projection, so
  summaries and reasoning-bearing records cannot supply lessons.
• Safe multi-record windows share the 64 KiB import budget across message
  bodies; short later replies stay whole and every record keeps its role,
  envelope and text. Mixed Claude messages shrink only visible text, and
  screened thinking blocks stay out of search, packs and learning. The full
  source is screened before shortening. Windows that cannot be bounded safely
  stay withheld; a valid prefix never stands in for the whole (a63e7c469).

Search and retrieval

• Queries that quote flags or paths (cargo fmt --check,
  std::fs::read_to_string, src/core/search.rs:120) no longer become
  exclusions or field filters that match nothing (e518c5d3f, 681dd4b8d).
• Lexical hits that share only common words with the query are dropped and
  counted in lexical_stopword_only_filtered (313e3fc9d). The check splits
  words the way the index does, so request_timeout counts as request and
  timeout and non-ASCII letters are lowercased: a memory containing
  request_timeout=30 is kept for the query request timeout (GH #68). Recency has a 0.5
  floor, so older memories rank lower but stay in packs (6f64aa179).
• Corrupt calibration data yields an unknown [0,1] interval with null
  coverage, not a 95% claim (c01fe8207). Weak semantic searches are recorded
  as misses for ee learn gaps (77bc3b095). An explicit remote embedding
  backend overrides the workspace registry (d53ca2fe2).

Context packs

• Result caps and pagination rank admitted memories, rules and native evidence
  together, so a relevant transcript hit can take the first slot. A
  higher-ranked rule or evidence item replaces weaker memories only when the
  complete replacement fits the token budget. A memory's LODs move together,
  and the reserved anti-pattern stays. Displaced memories keep persisted
  omissions, audit and replay (3bbec38e9).
• Rules with no source memory now enter packs (V128, cf8c95030). Without an
  explicit floor, evidence below 25% of the best evidence score is left out
  (context_evidence_below_relative_floor). Exact transcript repeats collapse,
  while distinct incident cards keep occurrence counts (b6bcabf9f).
• --relevance-floor applies to every candidate type, and an explicit
  --as-of covers its whole millisecond (d23129050, fdff9016a). Packs that
  use a memory from another shard are saved; 0.17.0 reported success without
  saving them. Storage errors suggest a retry rather than ee init
  (159178b4b).

ee ask, ee resume, ee focus and ee why

• Imported Claude Code transcripts now reach all four commands. Private-path
  checks screen the message rather than the envelope (which always contains an
  absolute cwd), and cass-session:// citations are no longer mistaken for
  secrets (10116d00f, 2620011e5). In ee ask, two sources no longer agree
  when their option arguments or assignment values differ (978572317).

Index publication

• Ordinary writes publish incremental index updates again; since 2026-08-06
  every write had re-embedded the whole corpus. A full build is used when
  digests are missing, the embedder changed, or more than 25% (at most 256) of
  inputs changed. Metadata edits and removals skip re-embedding (95d981a63).
• Each publication keeps only the two newest valid index copies. ee index vacuum shows a retention plan, and --apply removes the rest and records
  an audit row. Publication no longer fails after 1000 copies (e20782cd4,
  bc9d81d45).
• Large stale stores: rebuild planning and live fallback read in bounded pages,
  and hook-time reads wait at most 250 ms for a writer (GH #78, partly
  addressed; 60d7d1511). Index files are staged with filesystem copy-on-write
  where supported (feb0da79f).

Models, doctor and configuration

• ee model fetch potion-base-8M (256d, about 31 MB) or potion-base-32M
  (512d, about 131 MB) selects a smaller English model (GH #51, 8119c39dc).
  After a failed automatic download, ee waits before retrying, from 10 minutes
  doubling up to 24 h (c162289f5).
• ee config explain <key> shows what a setting does, its effective value and
  its source (569acaf06). ee doctor --fix gives WAL checkpoint guidance for
  EE-E205 (fa27e0529). ee diag store-integrity reads omitted generations
  from the store (3222a386c). ee curate doctor --trend totals are no longer
  capped at 1.0 (c4cc8575e).

Daemon, mesh, backup and security

• Remember retries with an idempotency key are atomic (5a0e047f1). Mesh
  responders shut down cleanly, route changes reach running listeners
  (a0260f996), and backups taken before V128 restore (112300fbb).
• Unicode-escaped JSON is decoded before credential screening (c077ef030).
  curl | sh, recursive rm and chmod 777 are flagged only in dangerous
  positions, so rm -rf target stays searchable (95d981a63).

Performance

• Debug-build timings from commit messages, for one 3,000-span session: search
  7.1 s -> 3.4 s and status 5.4 s -> 2.0 s (e013e2afc). A pack over 5k spans:
  41 s -> 8.2 s (2ffd8c917). Offline search: 9.2 s -> 1.4 s (c162289f5).

Fixed (0.17.0 known issues)

• Daemon write null strings (GH #62, 1810bb783), settings with HOME unset
  (GH #63, b3e08c1bc), and repair-test embedder identity (GH #64).
• ee import cass no longer aborts the whole batch when a live session grows
  between its reads; that session is refused for this run and picked up later
  (GH #77, b645122c9).

Dependencies

• FrankenSQLite 0.4.6 -> 0.4.10 (fixes the 0.4.7 strict read-only open
  regression and bulk-insert corruption). Frankensearch 0.6.1 -> 0.7.1,
  franken-agent-detection 0.3.3 -> 0.3.8. fnx stays =0.3.0. See
  UPGRADE_LOG.md.

Security

• rustls 0.23.46 (RUSTSEC-2026-0346..0349): four low-severity TLS advisories
  published 2026-10-14, covering Debug output of secrets, ECH, certificate
  decompression memory use and a ServerHello downgrade check.

Security note (GH #71)

• For each imported CASS line, ee stores an unsalted BLAKE3 hash of the
  raw line next to the redacted excerpt (cassSource.contentHash, and the
  reconciliation audit's sourceContentHash). The raw line includes the part
  that redaction removed.
• Someone holding the database or a backup can therefore test guesses of a
  short redacted value offline, such as a phone number, a short password or an
  email address.
• Until a keyed commitment lands (tracked in #71), treat the ee data directory
  and every ee backup as sensitive:
  • chmod 700 ~/.local/share/ee and your backup directories;
  • don't share or upload ee.db, its -wal/-shm files, or backups.
• New store files are created owner-only (0600) and backups are staged in
  0700 directories. The -wal/-shm sidecars are not tightened.

Known issues

• Index rebuild on large stores (GH #78). A large ee import cass can leave
  the search index behind the database. Commands that read the index can then
  be slow or wait on it, and ee index rebuild can be cancelled by its own
  deadline (cancelled, exit 130). This release bounds rebuild planning and
  hook-time waits (see Index publication). A resumable rebuild, and reads that
  degrade rather than wait, are still open. Import in smaller batches
  (--limit) and check ee index status.
• fnx 0.3.1 (GH #73). ee stays on fnx-* =0.3.0. The 0.3.1 Louvain API
  change is not integrated yet.
• CLI contract spawn timeouts (GH #72). On heavily loaded hosts, some
  real-binary contract tests exceed their 420 s spawn bound.

Release build and verification

Built with dsr from tag v0.18.0, release profile defaults (opt-level 3):

• the default x86-64 Linux archive is musl, statically linked;
• the supplemental x86-64 GNU and ARM64 GNU archives need glibc 2.28 or newer;
• both macOS archives;
• Windows x86-64 (MSVC).

Each archive has a SHA-256 sidecar and an epoch-2 Minisign signature (key 1BBD79B28BF718D0, signing/minisign-epoch2.pub). The signed SHA256SUMS also covers both installers, which are unchanged since 0.17.0. No Sigstore or SLSA attestation is claimed.

Gates:

• Lint:
  • all-target Clippy with warnings denied passed on macOS;
  • library and binary Clippy passed for x86-64 musl, x86-64 and ARM64 GNU, and x86-64 macOS;
  • Windows compiles cleanly. Its Clippy run reports 46 lint-level unused or dead-code items, Unix-only helpers; 45 of them were already present at 0.17.0.
• Tests (macOS): the full library, binary, integration and contract suites ran on this release and on its base commit 35b0bc06b in the same environment. No test fails in this release that passes at the base. The suites still carry pre-existing and environment-sensitive failures, which fail at the base too; they include three backup-recovery content-fence tests.
• Tests (Linux): the library suite gave 12,590 passed and 36 failed. 33 of those also fail at the base, and three are timing and contention failures that pass on macOS. The binary tests passed 11 of 11.
• #70: verified end to end. A session imported by ee 0.17.0 grows by a line, and this release imports the new line, where the base refused it on every run.
• The rustls update: release build, Clippy, cargo audit (no RUSTSEC-2026-0346..0349), and a smoke run of the release binary (init, remember, search).

Release source: 969f0cbf19a5fdaa367328350c937cf566063efc.