Engram v0.2.0-beta.1

Engram v0.2.0-beta.1 Release Notes

Release date: TBD
Status: Pre-release candidate

Supported Beta Path

This beta is scoped to the local/Codex Brain Loop path:

• local engram serve MCP operation,
• lean orient with current-plan retrieval, trace/cursor fields, used-memory candidate IDs, and
  obligation summary,
• generated Memory OS vault inspection,
• review-gated M6 inventory/export/status paths,
• scoped obligations doctor and advisory harness lifecycle guidance,
• preserved approval boundaries for destructive or broad writes.

Deferred From This Beta

The following remain production-hardening or host-parity gates, not blockers for this beta:

• native Claude prompt-bearing proof,
• effective-hook visibility proof,
• live Claude host-label proof,
• full multi-host parity,
• direct legacy deprecation/deletion,
• broad lifecycle cleanup or broad lint apply_safe,
• exhaustive telemetry completeness,
• OIDC/Vault/native-Claude auth/debugging edge cases,
• new feature work.

T306 resolved the current Rustdoc warning set for this candidate. Future Rustdoc polish remains a
production-hardening activity, not an initial-beta blocker.

Release Gate

Before tagging this beta, the candidate commit must have normal exact-head hosted CI proof, or an
explicit release-owner decision accepting local validation as a fallback while hosted Actions is
externally account-blocked. The expected gate remains:

• exact-head CI green for Format, Docs, Check, Clippy, and Test,
• cargo fmt --all --check passing locally,
• a focused local/Codex smoke confirming current source-rendered harness guidance,
• canonical generated vault status count-aligned,
• obligations(action=doctor, project=engram, cwd=/Users/yuval.meiri/projects/engram) clean,
• a refreshed installed runtime/adapters check.

The local fallback command is ./scripts/local-ci.sh. It mirrors the exact-head local validation
sequence used for this candidate: whitespace diff check, rustfmt, check, clippy, tests with CI-like
incremental/debug settings, and docs.

The local pre-publish packaging command is ./scripts/package-release.sh. It builds the release
binary, verifies that engram --version matches the workspace package version, and writes a tarball
plus SHA-256 checksum under ignored dist/. The archive includes MANIFEST.json with the package
version, host triple, git head, tracked-change state, Cargo.lock hash, and payload file hashes.

The local install-smoke command is ./scripts/package-install-smoke.sh. It builds the package,
verifies the checksum, extracts the archive, verifies MANIFEST.json, installs the packaged binary
into a temporary prefix, confirms PATH resolution and engram --version, starts the packaged
binary with engram serve --http --memory, and verifies /health. The smoke starts the server
from the temporary install workspace and sets ENGRAM_EMBED_CACHE_DIR explicitly, so package
validation no longer relies on the repository root as the process cwd for embedding model cache
discovery.

The post-publish install verifier is ./scripts/verify-published-release-install.sh. After a
release is published, it downloads the expected archive and checksum from GitHub, then runs the
same package install smoke against those downloaded assets with SKIP_PACKAGE_BUILD=1. It also
supports --asset-dir <path> for pre-publish rehearsal against a local release-asset mirror. The
verifier is evidence-only: it does not create a release, upload assets, mark PR #3 ready, merge,
tag, publish, accept the hosted-CI fallback, or mutate release state.

The hosted-CI pre-step blocker verifier is
./scripts/verify-hosted-ci-prestep-blocker.sh <run-id>. It is a read-only GitHub CLI check that
fails unless the run targets the expected head, is the CI pull-request workflow, contains exactly
the expected release-gate jobs, and every job completed with conclusion failure and steps=[].
It is evidence for release-owner waiver review only; it does not accept the waiver or change PR
state.
Add --json to emit machine-readable success evidence for the verified run. Failed checks remain
stderr-only and non-zero.

The beta release-gate evidence command is
./scripts/beta-release-gate-report.sh --hosted-run <run-id>. It verifies the branch is synced,
the tracked tree is clean, PR #3 points at the local head, hosted CI is either green or the
pre-step blocker is verified, and, unless --quick or skip flags are used, the local CI and
package/install smoke commands pass on the same head. It is evidence for release-owner review only;
it does not accept the waiver, mark the PR ready, merge, tag, publish, or change release scope.
Add --json to emit the final evidence object on stdout for automation; validation logs are routed
to stderr in JSON mode so stdout remains parseable. When hosted CI is accepted only as a verified
pre-step blocker candidate, the JSON object embeds the verifier's structured proof under
hosted_ci.verifier; a green hosted-CI path leaves that field null. The same report also emits
release_gate_state, ready_for_release_owner_review,
hosted_ci_fallback_decision_required, and remaining_release_actions so automation can
distinguish full exact-head evidence from quick/incomplete evidence without performing release
actions.

Release-Owner Signoff Checklist

The remaining beta decision is explicit release-owner signoff, not additional product scope. Before
tagging v0.2.0-beta.1, the release owner should confirm:

1. ./scripts/local-ci.sh on the signed-off head is accepted as the local CI-equivalent beta gate
   while hosted GitHub Actions is externally blocked before workflow-step execution.
2. ./scripts/package-install-smoke.sh on the same head is accepted as package/checksum/install
   and packaged HTTP /health proof.
3. The current hosted run can be waived for this beta only if it failed before workflow-step
   execution with steps=[], matching the external billing/spending-limit blocker rather than a
   source or packaging failure. Verify the run with
   ./scripts/verify-hosted-ci-prestep-blocker.sh <run-id> before signoff.
4. ./scripts/beta-release-gate-report.sh --hosted-run <run-id> has been run on the signed-off
   head, or the equivalent branch/PR/CI/local/package evidence has been reviewed manually.
5. The known beta limitations listed in this file are accepted: native Claude prompt-bearing
   behavior, effective-hook visibility, live Claude host labels, full multi-host parity, direct
   legacy deprecation/deletion, broad lifecycle cleanup or broad lint apply_safe, M6 write-apply
   expansion, and exhaustive telemetry/auth/ops/performance/cross-platform hardening remain
   deferred.
6. After signoff, PR #3 may be marked ready, merged, tagged as v0.2.0-beta.1, published with the
   release archive and checksum, and verified with
   ./scripts/verify-published-release-install.sh.

Beta Install Quickstart

For source installs, build and place the binary on PATH before running engram init or
engram serve:

git clone https://github.com/ymeiri/engram.git
cd engram
cargo build --release

mkdir -p "$HOME/.local/bin"
install -m 755 ./target/release/engram "$HOME/.local/bin/engram"
export PATH="$HOME/.local/bin:$PATH"
engram --version

After v0.2.0-beta.1 is published, install the macOS arm64 release artifact with checksum
verification:

version=0.2.0-beta.1
archive="engram-${version}-aarch64-apple-darwin.tar.gz"

curl -LO "https://github.com/ymeiri/engram/releases/download/v${version}/${archive}"
curl -LO "https://github.com/ymeiri/engram/releases/download/v${version}/${archive}.sha256"
shasum -a 256 -c "${archive}.sha256"
tar -xzf "${archive}"

mkdir -p "$HOME/.local/bin"
install -m 755 "engram-${version}-aarch64-apple-darwin/engram" "$HOME/.local/bin/engram"
export PATH="$HOME/.local/bin:$PATH"
engram --version

The expected version output for this beta is:

engram 0.2.0-beta.1

Recent phase-1 local evidence is strong: T317 validated PR #3 head
78f14d0bebd980070a4fcb8d1f259be47517c704 with cargo fmt --all --check,
git diff --check, cargo check --all-targets,
cargo clippy --all-targets -- -D warnings, cargo test --all-targets --jobs 1, and
cargo doc --no-deps. T318 reran hosted GitHub Actions run 27091138284, creating attempt 2 on
that same head, but all five jobs failed before runner assignment with zero steps, runner_id=0,
and billing/spending-limit annotations. That external account gate does not contradict the local
validation, but the normal exact-head hosted-CI release proof is still missing until Actions can run
on the head intended for release or the release owner explicitly accepts the local fallback.

T329 advanced the draft PR #3 head to fe46d0a73d39e3309b149703dda4c108da91fc02 through
docs-only release evidence plus exact lifecycle archive records. Local validation for that head
passed git diff --check, cargo fmt --all --check, cargo check --all-targets, canonical vault
compile with zero skipped files, and cached diff checks. Hosted GitHub Actions run 27096981016
on the same head again failed before workflow steps ran with the same billing/spending-limit
annotations. Treat this as the current hosted-CI blocker.

Fresh AI Council review after T329 places the initial local/Codex beta at about 88-92% ready while
hosted CI is externally blocked and local fallback evidence is accepted, or about 95% ready once
GitHub Actions billing is fixed and exact-head checks pass or the release owner explicitly accepts
local validation as the beta fallback. T330 also records a current-head local/Codex smoke with lean
orient, obligations doctor, vault status/compile, lint-sample evidence, and bounded M6
inventory/temp-export/status/dry-run-apply evidence. This is not a production/GA readiness claim;
production readiness remains materially lower because native Claude proof, effective hooks,
host-label proof, host parity, telemetry completeness, and operational hardening remain open.

T331 closes the next exact superseded rolling-handoff lifecycle batch:
019e7cf7-560c-70e2-bbeb-3448f4637055,
019e7d27-32d6-7200-944c-ef5945436f8c,
019e7d28-add4-70e3-a55c-453f8fe8695d,
019e7d29-0f3c-7961-9588-c1adbe4628af, and
019e7da0-d384-7b12-b43a-d7188b1a8c38. Post-archive lint advances to
019e7db8-de1e-7251-87ba-fea21bed17f7, so broad lifecycle cleanup remains deferred and
exact-target-gated rather than part of the beta release gate.

T332 closes the next single exact superseded rolling-handoff target,
019e7db8-de1e-7251-87ba-fea21bed17f7, after successor review showed it is directly superseded
by active handoff 019e844c-6a05-7a10-858b-5212d117a4bb. Post-archive lint no longer returns the
T332 target in the first ten sampled findings; the bounded sample now reports stale-feedback review
signals rather than a superseded-active warning.

Current Installation Status

T305 refreshed the installed local binary, generated Codex adapter, and global daemon from the
0.2.0-beta.1 candidate. The installed binary now reports:

engram 0.2.0-beta.1

The installed Codex harness is Ready: true, and both source-rendered and installed Codex harness
guidance include scoped final-response obligation checks:

obligations(action=doctor, project=..., cwd=...)

Already-open agent UI sessions may still need a fresh session or tool reload before they ingest the
updated skill text. This does not change the beta deferrals for native Claude, effective hooks,
host labels, or full multi-host parity.

Claude Code Adapter Safety Follow-Up

T315 adds source-level coverage for the T314 repair path later executed by T333. The new harness
test proves that HarnessSettingsTarget::SnippetOnly can repair generated Claude Code adapters
without rewriting an existing settings.json, settings.local.json, or
engram-settings-snippet.json.

T333 executes the prepared T314 command with the installed CLI:

/Users/yuval.meiri/.local/bin/engram harness install --harness claude-code --settings-target snippet-only --write --json

The write updated exactly the three generated Claude Code adapters, left settings.json,
settings.local.json, and engram-settings-snippet.json unchanged, and made Claude Code
harness status and harness doctor report ready=true.

This closes generated-adapter drift only. It still does not prove native Claude prompt-bearing
execution, effective-hook visibility, live host labels, or full production parity.

Native Claude Post-Repair Preflight

T334 reruns the native-Claude/effective-hook/host-label preflight after the T333 adapter repair.
The Claude 2.1.168 path, target, SHA-256, installed daemon, obligations doctor, canonical vault
status, generated adapter hashes, snippet-only dry-run, and Claude Code harness readiness now
match the expected state. The preflight still hard-stops before launching native Claude because
existing native Claude CLI sessions are live on ttys001 and ttys005, making attribution
ambiguous for a new prompt-bearing run.

This narrows the remaining production-hardening gap: adapter drift is closed, but native Claude
prompt-bearing proof, effective-hook visibility, and live host labels remain deferred from the
initial beta.

Effective-Hook Successor Packet

T335 records a docs-only successor for the T269 effective-hook visibility packet under the observed
Claude Code 2.1.168 runtime. It updates the future /hooks preflight baseline from the stale
2.1.161 target to the T334-observed path, version, and SHA-256, while preserving the strict
one-/hooks transcript observation contract, bounded cleanup, T312/T270 separation, and T334
attribution hard-stop.

T335 does not launch native Claude, run /hooks, mutate settings or adapters, prove
effective-hook visibility, or change the initial beta scope.

Project-Scoped Lint

T336 adds optional project filtering to Memory OS lint. MCP callers can pass project, and the CLI
can run engram lint run --scope-project <name> to focus memory, stale-session, and open-obligation
findings on the current project while preserving global/user memory checks and unchanged unscoped
lint behavior.

This improves project health visibility for beta closeout. It does not run lint apply_safe,
archive memory, or close lifecycle cleanup.

Project-Scoped Lint Installed Runtime Refresh

T337 installs the current engram-cli into /Users/yuval.meiri/.local/bin/engram and restarts the
daemon so the T336 project-scoped lint surface is live in the installed local/Codex path. The
installed binary hash changed from
01b171ec654da95ea5b1f8363bc109e3069c0ff78bdb38581a202e472f9fd09b to
b775efa0946862eba8d4d8993bb946f0926372d8a3fe9bbfea98ea38e786e7c2, daemon status now reports
PID 57356 spawned by /Users/yuval.meiri/.local/bin/engram, and fresh installed MCP smoke
confirmed lint.project appears in tools/list and tools/call accepts project=engram.

T337 is installed-runtime adoption for T336 only. It does not run lint apply_safe, mutate memory,
or change the beta deferrals.

Rolling Handoff Evidence

T338 makes rolling handoff updates evidence-backed. After T337, project-scoped lint showed the
current rolling handoff was active but missing evidence. HandoffService::update now adds
tool-call evidence for global, project, and session handoff writes while preserving existing
session-event evidence for session-scoped handoffs.

The T338 runtime was installed to /Users/yuval.meiri/.local/bin/engram with hash
e53765568a2232c55c2d17a8a48480e745b2c2fda044a8d087681c20534e3dc5, daemon PID 92750, and the
new active handoff 019ea34a-c3ac-74d0-ae42-52cd6adcb610 carries
handoff(action=update,project=engram) evidence. Installed project-scoped lint no longer flags
that active handoff as missing evidence.

This improves Memory OS hygiene for future handoffs. It does not run lint apply_safe, archive
historical handoffs, or change the beta deferrals.

Canonical Vault Resync

T339 refreshes the durable generated Memory OS vault after the latest T338/T339 memory writes. The
preflight status showed the vault was initialized and generated-only, but count-stale:
generated_file_count=2566 while expected_generated_file_count=2568.

vault(action=compile, vault_path="/Users/yuval.meiri/.engram/vault") completed with
files_skipped=[], and postflight status returned generated_file_count=2568,
user_file_count=0, and expected_generated_file_count=2568.

This restores the canonical vault validation gate for the current Memory OS state. It does not
change source behavior, run lint apply_safe, close hosted CI, or change the beta deferrals.

Native Claude Preflight Refresh

T340 reruns the read-only native-Claude/effective-hook/host-label preflight at PR #3 head
6d0467e933a880f9039fd943b34848c2ca93f069. The Claude 2.1.168 path, target, SHA-256, Engram
daemon, obligations doctor, canonical vault status, Claude Code harness readiness, and snippet-only
install dry-run all match the expected state. The canonical vault is currently count-aligned at
2573 generated files, zero user files, and 2573 expected generated files.

The preflight still hard-stops before launching native Claude because native Claude CLI sessions
remain live on ttys001 and ttys005, making new-session attribution ambiguous. T340 does not
launch native Claude, run /hooks, signal processes, mutate settings or adapters, prove
prompt-bearing behavior, prove effective-hook visibility, prove live host labels, close hosted CI,
or change the initial beta scope.

Exact-Head Local CI Fallback

T341 validates PR #3 head 2fa5b577bda8ab6141e0f7272736044d441a7e88 with the full
CI-equivalent local workflow after hosted run 27101388242 again failed all five jobs before
executing any workflow steps. gh run view reports steps: [] for Check, Test, Format, Clippy,
and Docs, matching the prior external account/billing/spending-limit pattern.

That T341 head passed:

• git diff --check
• cargo check --all-targets
• CARGO_INCREMENTAL=0 CARGO_PROFILE_DEV_DEBUG=0 cargo test --all-targets --jobs 1
• cargo fmt --all --check
• cargo clippy --all-targets -- -D warnings
• cargo doc --no-deps

Fresh AI Council consensus after T341 puts the scoped local/Codex MVP beta at about 95%
complete and shippable if the release owner accepts local validation as the fallback. Hosted CI
passing is still not achieved and remains either an ops/GA hygiene item or a beta gate only if the
release owner requires it. T341 does not mark PR #3 ready, merge, tag, publish, close hosted CI,
or change the production/GA deferrals.

Beta Scope Consensus Refresh

T343 refreshes the release-scope decision on the current PR #3 head
966dc00d5248ac342b156974b5392700706f3139. The PR body records that this exact head passed the
full local CI-equivalent workflow before push, while hosted run 27101972733 failed Check, Test,
Format, Clippy, and Docs before workflow-step execution with steps: [].

Fresh AI Council consensus keeps the scoped local/Codex MVP beta at about 95-98% ready before
release mechanics if the release owner accepts local validation as fallback. The remaining beta gate
is either explicit release-owner fallback acceptance or restored exact-head hosted CI. Production/GA
readiness remains separate and materially lower because native Claude prompt-bearing proof,
effective-hook visibility, live host labels, full multi-host parity, broad lifecycle cleanup,
direct legacy deprecation/deletion, exhaustive telemetry, auth edge hardening, performance, and
cross-platform polish remain open.

T343 does not mark PR #3 ready, merge, tag, publish, close hosted CI, run native Claude, prove
hooks or host labels, or change the supported beta scope.

One-Command Local CI and Package Proof

T345 adds ./scripts/local-ci.sh as the single exact-head local CI-equivalent fallback command for
release validation. It runs whitespace diff checks, rustfmt, cargo check, clippy with warnings as
errors, the full test suite with CI-like incremental/debug settings, and rustdoc generation.

T346 adds ./scripts/package-release.sh as the local pre-publish packaging command. It builds the
release binary, checks that engram --version matches the workspace package version, packages the
binary with README, LICENSE, changelog, and these release notes, and writes a SHA-256 checksum under
ignored dist/.

T347 validates PR #3 head b0d8e075a04b5e35f4ed7c4d60654231ed5c1324 with both commands. The
exact head passed ./scripts/local-ci.sh, and ./scripts/package-release.sh produced
dist/engram-0.2.0-beta.1-aarch64-apple-darwin.tar.gz plus its .sha256 file. The archive
contents were inspected, the checksum verified, and the extracted packaged binary reported
engram 0.2.0-beta.1.

Hosted GitHub Actions run 27102953577 on that same head still failed before workflow-step
execution with steps: [] and account billing/spending-limit annotations. The remaining beta gate
is therefore explicit release-owner fallback acceptance or restored exact-head hosted CI. T347 does
not mark PR #3 ready, merge, tag, publish, close hosted CI, run native Claude, prove hooks or host
labels, or change the supported beta scope.

T350 closes the first-user onboarding gap on PR #3 head
54d01eb71e2020960fa62c0d6b72a05b5c00aee4. README and these release notes now show source
install steps that put engram on PATH before engram init or engram serve, plus published
tarball download, checksum, install, and expected-version commands. That head passed
./scripts/local-ci.sh, ./scripts/package-release.sh, checksum verification, packaged release
note inspection, and a manual isolated package install/health smoke.

T351 adds ./scripts/package-install-smoke.sh so the install proof is repeatable instead of
manual. T351 package-smoke candidate head 4d05f6fc2f4fb4c6309c431c083ea55540c32380 passed
./scripts/local-ci.sh and ./scripts/package-install-smoke.sh. The smoke builds the package,
verifies the checksum, extracts the archive, checks required packaged files, installs the packaged
binary into a temporary prefix, confirms PATH resolution and engram 0.2.0-beta.1, starts the
packaged binary with engram serve --http --memory, and verifies /health returns
{"status":"ok","service":"engram","version":"0.2.0-beta.1"}.

Hosted GitHub Actions run 27114219090 on 4d05f6f still failed before workflow-step execution:
Check, Test, Format, Clippy, and Docs all report steps: []; the sampled Check annotation says
recent account payments failed or the spending limit must be increased. Fresh AI Council consensus
on 2026-06-08 gives a conservative release-management estimate of about 92% complete until
release-owner approval, ready/merge/tag/publish, and a practical local/Codex beta readiness of
98-99% if the release owner accepts exact-head local CI plus package/install validation as the
hosted-CI fallback.

T351 does not mark PR #3 ready, merge, tag, publish, close hosted CI, run native Claude, prove
hooks or host labels, or change the supported beta scope.

T353 makes embedding model cache discovery deterministic for beta installs. EmbedConfig::default
now uses ~/.engram/cache/fastembed by default, accepts ENGRAM_EMBED_CACHE_DIR as the
Engram-specific override, and preserves upstream FASTEMBED_CACHE_DIR compatibility. HF_HOME
still remains an upstream Hugging Face override when set. The package install smoke now starts the
packaged binary from the temporary install workspace with ENGRAM_EMBED_CACHE_DIR set explicitly,
so /health validation no longer depends on repository-root cwd.

T353 validation passed cargo test -p engram-embed config::tests,
cargo test -p engram-tests --test multi_session_tests, ./scripts/package-install-smoke.sh, and
./scripts/local-ci.sh. The first full local CI attempt exposed that isolated multi-session test
daemons also needed an explicit stable cache path; the test daemon helper now sets
ENGRAM_EMBED_CACHE_DIR to the configured cache override or package-local .fastembed_cache.

T353 does not mark PR #3 ready, merge, tag, publish, close hosted CI, run native Claude, prove
hooks or host labels, or change the supported beta scope.

T354 refreshes the installed local binary after T353. cargo install --path engram-cli --force --root /Users/yuval.meiri/.local replaced the previous installed hash
e53765568a2232c55c2d17a8a48480e745b2c2fda044a8d087681c20534e3dc5 with
a47edffa8c8ed955a311adac85033ce8a28235c37007b5149ea81a8ffeb456de, while
/Users/yuval.meiri/.local/bin/engram --version still reports engram 0.2.0-beta.1.

An isolated installed-runtime smoke started /Users/yuval.meiri/.local/bin/engram serve --http --memory from a temporary workspace with temp HOME, temp ENGRAM_DATA_DIR, and explicit
ENGRAM_EMBED_CACHE_DIR pointing at the warmed repo cache. GET /health returned
{"status":"ok","service":"engram","version":"0.2.0-beta.1"}. The existing global daemon was not
restarted during this validation.

T354 does not mark PR #3 ready, merge, tag, publish, close hosted CI, run native Claude, prove
hooks or host labels, or change the supported beta scope.

T355 restarts the live global daemon onto the refreshed T354 installed binary. Before restart,
engram daemon status reported PID 92750, spawned by /Users/yuval.meiri/.local/bin/engram,
with spawn version 0.2.0-beta.1. After engram daemon stop, the first engram daemon start
attempt exited before readiness because the daemon log reported a transient RocksDB
~/.engram/data/LOCK conflict. A follow-up status check reported no running daemon, and the
retry succeeded.

The refreshed live daemon now reports PID 64693 on port 8765, spawned by
/Users/yuval.meiri/.local/bin/engram, with spawn version 0.2.0-beta.1 and installed hash
a47edffa8c8ed955a311adac85033ce8a28235c37007b5149ea81a8ffeb456de. GET /health
returned {"status":"ok","service":"engram","version":"0.2.0-beta.1"}, and a live MCP
orient(response_shape="lean") smoke for the current MVP-status prompt returned trace
019ea620-b1be-7dc1-a022-15f49badddf6.

T355 does not mark PR #3 ready, merge, tag, publish, close hosted CI, run native Claude, prove
hooks or host labels, or change the supported beta scope.

T356 reruns the read-only native-Claude/effective-hook/host-label preflight after the T355 live
daemon refresh. The branch is synchronized with its upstream, and origin/main...HEAD is 0 48,
so the earlier divergent-branches pull hint is not current for yuval.meiri/memory-os-phase1.
PR #3 remains draft at 6a561863154d7b3fabc0e176d3105ed1f0dedd7a; hosted run
27122932006 still fails all five jobs before workflow-step execution with steps: [].

The Claude binary preflight still matches the T312/T335/T340 baseline:
/Users/yuval.meiri/.local/bin/claude, version 2.1.168 (Claude Code), SHA-256
377f0ecedba8246bdabdf312ce8b7cc8ae1160997b26f5edca352a4a8d61dc78. Claude Code
harness status and harness doctor report ready=true; snippet-only install dry-run reports
planned=[]; the global daemon remains PID 64693 from /Users/yuval.meiri/.local/bin/engram
with spawn version 0.2.0-beta.1; /health returns
{"status":"ok","service":"engram","version":"0.2.0-beta.1"}; canonical vault status is
count-aligned at 2628 generated files; and obligations doctor is clean.

The preflight still hard-stops before launching native Claude because native Claude CLI sessions
remain live as PID 60453 on ttys001 and PID 311 on ttys005, making new-session attribution
ambiguous. T356 does not launch native Claude, run /hooks, signal processes, mutate settings or
adapters, prove prompt-bearing behavior, prove effective-hook visibility, prove live host labels,
close hosted CI, or change the supported beta scope.

CLI Daemon Admin Routing

T357 hardens local beta administration while the daemon owns the store. engram lint,
engram obligations, and engram vault now prefer the healthy global or project daemon when no
explicit --data-dir is supplied, using the existing one-shot MCP proxy path. Explicit data-dir
commands still use direct RocksDB access, and the direct path remains the fallback when no matching
daemon is healthy.

This fixes the practical lock failure observed when running installed CLI health checks while the
global daemon held ~/.engram/data/LOCK. Focused validation passed cargo fmt --all --check,
git diff --check, cargo check -p engram-cli, and cargo test -p engram-cli. Live source-binary
and installed-path smokes proved lint run --scope-project engram, vault status, and
obligations doctor --limit 3 return daemon-backed output instead of the lock error while daemon
PID 64693 remains running.

The same T357 head also passed ./scripts/local-ci.sh and
./scripts/package-install-smoke.sh. The package smoke rebuilt
engram-0.2.0-beta.1-aarch64-apple-darwin.tar.gz, verified its checksum, installed the packaged
binary into a temporary prefix, confirmed engram 0.2.0-beta.1, started packaged
engram serve --http --memory, and verified /health returned
{"status":"ok","service":"engram","version":"0.2.0-beta.1"}.

cargo install --path engram-cli --force --root /Users/yuval.meiri/.local refreshed the installed
CLI hash from a47edffa8c8ed955a311adac85033ce8a28235c37007b5149ea81a8ffeb456de to
fa91efbd228683dae608881f5828bdc1ffe55b67376e414653f8ac8eb92ba8c9; engram --version
continues to report engram 0.2.0-beta.1.

T357 does not mark PR #3 ready, merge, tag, publish, close hosted CI, run native Claude, prove
hooks or host labels, mutate lifecycle state, run broad lint apply_safe, or change the supported
beta scope.

Lint Project Schema Exposure

T358 hardens the public MCP metadata for project-scoped lint. LintRequest.project now has an
explicit schema description, and the test suite guards both the generated request schema and the
HTTP daemon tools/list response so lint.inputSchema.properties.project remains visible to agent
callers.

Focused validation passed cargo fmt --all --check, git diff --check,
cargo test -p engram-mcp lint_request_schema_exposes_project_filter, cargo test -p engram-mcp,
cargo test -p engram-tests --test lint_tests test_mcp_lint_project_filter_excludes_unrelated_project_memory,
and
cargo test -p engram-tests --test multi_session_tests test_mcp_tools_list_lint_schema_exposes_project_filter.

cargo install --path engram-cli --force --root /Users/yuval.meiri/.local refreshed the installed
CLI hash from fa91efbd228683dae608881f5828bdc1ffe55b67376e414653f8ac8eb92ba8c9 to
62c9955925f74fba706ad466416033cc0bdbc211cf0443a373d4e5925760589a; engram --version
continues to report engram 0.2.0-beta.1. The live global daemon was restarted onto that binary
as PID 36562 on port 8765, and an installed-daemon tools/list smoke returned true for
lint.inputSchema.properties.project.description == "Optional project scope to lint.".

T358 does not mark PR #3 ready, merge, tag, publish, close hosted CI, run native Claude, prove
hooks or host labels, mutate lifecycle state, run broad lint apply_safe, or change the supported
beta scope.

T359 hardens scoped obligation health checks in the CLI. engram obligations list and
engram obligations doctor now accept --scope-project and --cwd, matching the existing MCP
project/cwd filters for obligation list and doctor actions. The daemon-backed CLI path serializes
those filters into tools/call; the direct-store fallback path passes them to
ObligationService::list and ObligationService::doctor. Unscoped list/doctor behavior remains
unchanged, and top-level engram obligations --project <name> continues to act as the default
project scope.

Focused validation passed cargo test -p engram-cli obligation_daemon_arguments,
cargo check -p engram-cli, cargo fmt --all --check, git diff --check, and
cargo test -p engram-tests --test obligation_tests. Source CLI help exposes the new flags for
both commands. Source daemon-backed smokes for scoped doctor and list passed, and isolated
--data-dir smokes covered the direct-store fallback path.

cargo install --path engram-cli --force --root /Users/yuval.meiri/.local refreshed the installed
CLI hash from 62c9955925f74fba706ad466416033cc0bdbc211cf0443a373d4e5925760589a to
ae45c01ab2a4c5046508e916a7c381655a71611f223fd8fc7989392cd3879f79; engram --version
continues to report engram 0.2.0-beta.1. Installed daemon-backed smokes passed:
engram obligations doctor --scope-project engram --cwd /Users/yuval.meiri/projects/engram --limit 3 --json
returned {"open":[],"warnings":[]}, and
engram obligations list --scope-project engram --cwd /Users/yuval.meiri/projects/engram --status open --limit 3 --json
returned [].

Final exact-head validation passed the local CI-equivalent steps (git diff --check,
cargo fmt --all --check, cargo check --all-targets,
cargo clippy --all-targets -- -D warnings, serialized cargo test --all-targets --jobs 1, and
cargo doc --no-deps) plus ./scripts/package-install-smoke.sh. The first ./scripts/local-ci.sh
run completed through check and clippy before its test step stalled in rustc with no CPU activity;
the stalled process was terminated, and the same serialized test step plus docs step were rerun
directly and passed. The package smoke verified the release tarball checksum, installed the package
into a temporary prefix, confirmed engram 0.2.0-beta.1, and verified packaged HTTP /health
returned {"status":"ok","service":"engram","version":"0.2.0-beta.1"}.

T359 does not mark PR #3 ready, merge, tag, publish, close hosted CI, run native Claude, prove
hooks or host labels, mutate lifecycle state, run broad lint apply_safe, or change the supported
beta scope.

T360 hardens public MCP metadata for scoped obligation health checks. ObligationRequest.project
now explicitly describes its supported detect/add/list/open/doctor scope, and
ObligationRequest.cwd now explicitly describes detect/list/open/doctor scoping. This matches the
runtime MCP behavior and the T359 CLI flags, so agents reading tools/list can discover the scoped
obligation list/doctor path without guessing.

Focused validation passed cargo test -p engram-mcp obligation_request_schema_exposes_scope_filters,
cargo test -p engram-tests --test multi_session_tests test_mcp_tools_list_obligations_schema_exposes_scope_filters,
cargo fmt --all --check, git diff --check, cargo test -p engram-mcp, and
cargo test -p engram-tests --test multi_session_tests test_mcp_tools_list.

cargo install --path engram-cli --force --root /Users/yuval.meiri/.local refreshed the installed
CLI hash from ae45c01ab2a4c5046508e916a7c381655a71611f223fd8fc7989392cd3879f79 to
ff16b90be46e54d089ce66e5b360630449bffc9f874da031beb10884f994756b; engram --version
continues to report engram 0.2.0-beta.1. The live global daemon was restarted onto that binary
as PID 48118 on port 8765, /health returned
{"status":"ok","service":"engram","version":"0.2.0-beta.1"}, and a live tools/list smoke
confirmed the obligations schema exposes:
project = "Optional project scope for detect, add, list, open, and doctor." and
cwd = "Current working directory for detect/list/open/doctor scoping.".

Final exact-head validation passed ./scripts/local-ci.sh and ./scripts/package-install-smoke.sh.
The package smoke verified the release tarball checksum, installed the package into a temporary
prefix, confirmed engram 0.2.0-beta.1, and verified packaged HTTP /health returned
{"status":"ok","service":"engram","version":"0.2.0-beta.1"}.

T360 does not mark PR #3 ready, merge, tag, publish, close hosted CI, run native Claude, prove
hooks or host labels, mutate lifecycle state, run broad lint apply_safe, or change the supported
beta scope.

T361 hardens public MCP metadata for project-scoped telemetry reports. TelemetryRequest.project
now explicitly describes its support for record_trace, list_traces, list_feedback,
stats_by_intent, and real_session_eval. This matches the runtime telemetry behavior, so agents
reading tools/list can discover the project-filtered telemetry/eval path without guessing.

Focused validation passed
cargo test -p engram-mcp telemetry_request_schema_exposes_project_filter and
cargo test -p engram-tests --test multi_session_tests test_mcp_tools_list_telemetry_schema_exposes_project_filter.
Broader validation passed cargo fmt --all --check, git diff --check, cargo test -p engram-mcp, cargo test -p engram-tests --test multi_session_tests test_mcp_tools_list,
./scripts/local-ci.sh, and ./scripts/package-install-smoke.sh.

cargo install --path engram-cli --force --root /Users/yuval.meiri/.local refreshed the installed
CLI hash from ff16b90be46e54d089ce66e5b360630449bffc9f874da031beb10884f994756b to
6c278872d2f71a5ce96fba3e1777b3cc2f4690e6d6c9caf74df093fb4fd7e49a; engram --version
continues to report engram 0.2.0-beta.1. The live global daemon was restarted onto that binary
as PID 2865 on port 8765, /health returned
{"status":"ok","service":"engram","version":"0.2.0-beta.1"}, and a live tools/list smoke
confirmed the telemetry schema exposes
project = "Optional project scope for record_trace, list_traces, list_feedback, stats_by_intent, and real_session_eval.".

T361 does not mark PR #3 ready, merge, tag, publish, close hosted CI, run native Claude, prove
hooks or host labels, mutate lifecycle state, change telemetry filtering semantics, or change the
supported beta scope.

T362 hardens public MCP metadata for the core orient Brain Loop entrypoint. OrientRequest.cwd
now explicitly describes repository/project resolution and scoped memory selection,
OrientRequest.project describes project resolution and project-scoped memory selection, and
OrientRequest.response_shape describes the full/lean response shape contract for compact
trace/cursor/Brain Loop guidance.

Focused validation passed
cargo test -p engram-mcp orient_request_schema_exposes_context_contract and
cargo test -p engram-tests --test multi_session_tests test_mcp_tools_list_orient_schema_exposes_context_contract.
Broader validation passed cargo fmt --all --check, git diff --check, cargo test -p engram-mcp,
cargo test -p engram-tests --test multi_session_tests test_mcp_tools_list, and
./scripts/local-ci.sh. ./scripts/package-install-smoke.sh rebuilt the release package, verified
the checksum, installed the packaged binary in a temporary prefix, reported engram 0.2.0-beta.1,
and verified packaged HTTP /health. The installed runtime hash is
77a08e895614bea3b02816e67bafd64087ea0634f4b0ca58b8199a9ef7855633; the live daemon was
restarted as PID 47577; an installed-daemon tools/list smoke confirmed the three orient
schema descriptions.

T362 does not mark PR #3 ready, merge, tag, publish, close hosted CI, run native Claude, prove
hooks or host labels, mutate lifecycle state, change orientation ranking, or change the supported
beta scope.

T363 refreshes the native Claude prompt-bearing, effective-hook, and live host-label preflight after
T362. The read-only assertions still match: Claude Code resolves to
/Users/yuval.meiri/.local/share/claude/versions/2.1.168, reports 2.1.168 (Claude Code), and
has SHA-256 377f0ecedba8246bdabdf312ce8b7cc8ae1160997b26f5edca352a4a8d61dc78; Claude Code
harness status/doctor report ready=true; snippet-only install dry-run plans no writes; daemon PID
47577 is healthy; canonical vault status is count-aligned at 2651 generated files with zero
user files; scoped obligations doctor is clean.

T363 hard-stops before native execution because process inventory shows an active native claude
process on ttys004 (PID 34797). That process makes a new single-session native transcript
ambiguous for T312 prompt-bearing proof, T335 /hooks effective-hook proof, and T270 live
host-label proof.

T363 does not launch native Claude, send prompts, run /hooks, signal processes, mutate settings
or adapters, mark PR #3 ready, merge, tag, publish, close hosted CI, or change the supported beta
scope.

T364 performs exact lifecycle cleanup for two stale, evidence-less Claude-harness MemoryItems:
019dd4e3-bcec-7c02-9174-ba0ac0380d45 (Claude Code native hook harness implemented) and
019dd509-46f2-71c0-aff7-ebe777810825 (Claude Code Engram-native harness activated). Fresh
memory(get) checks showed both were active with empty evidence; graph(around, depth=1) showed
only project scope; and project-scoped lint flagged them as stale-feedback/missing-evidence active
memory. Both were archived with explicit reasons because newer evidenced T333/T340/T363 records now
carry the current Claude-harness state: generated adapter drift is closed, while native Claude
prompt-bearing behavior, effective-hook visibility, and live host labels remain unproved.

Post-archive validation confirmed both IDs are status=archived, live daemon project-scoped
lint(action=run, project=engram, limit=30) returned archived_targets_present=[], and direct
search for the old activation titles now surfaces active evidenced successor/limitation records
first.

T364 does not change source code, launch native Claude, run /hooks, signal processes, mutate
settings or adapters, run broad lint apply_safe, delete data, mark PR #3 ready, merge, tag,
publish, close hosted CI, or change the supported beta scope.

T365 performs exact lifecycle cleanup for two stale, evidence-less May 6 orient-contract and
architecture checkpoint MemoryItems: 019dfed3-519d-7f01-8c46-c9245ba0045b
(AI Council and Claude next-step synthesis after orient contract) and
019dfed5-1875-7110-b355-8d1060e6d04a
(Brain Harness Architecture synced after orient contract checkpoint). Fresh memory(get) checks
showed both were active with empty evidence; graph(around, depth=1) showed only project scope; and
project-scoped lint flagged both as stale-feedback/missing-evidence active memory. Both records were
archived with explicit reasons because their May 6 next-step/checkpoint guidance is historical, not
current.

Post-archive validation confirmed both IDs are status=archived, live daemon project-scoped
lint(action=run, project=engram, limit=40) returned archived_targets_present=[] and reduced the
returned finding count from 20 to 16, and direct search for the old titles still returns active
evidenced current-plan/beta-scope records first.

T365 does not change source code, run broad lint apply_safe, delete data, mark PR #3 ready,
merge, tag, publish, close hosted CI, run native Claude, prove hooks or host labels, or change the
supported beta scope.

T366 refreshes local beta acceptance evidence for PR #3 while hosted GitHub Actions remains
externally blocked. The local CI-equivalent command set passed: git diff --check,
cargo fmt --all --check, cargo check --all-targets,
CARGO_INCREMENTAL=0 CARGO_PROFILE_DEV_DEBUG=0 cargo test --all-targets --jobs 1,
cargo clippy --all-targets -- -D warnings, and cargo doc --no-deps.
./scripts/package-install-smoke.sh also passed, rebuilding the release package, verifying the
tarball checksum, installing the packaged binary into a temporary prefix, confirming
engram 0.2.0-beta.1, and verifying packaged HTTP /health returned
{"status":"ok","service":"engram","version":"0.2.0-beta.1"}.

T366 narrows the initial beta gate to either restored exact-head hosted CI or explicit
release-owner acceptance of this local CI plus package/install fallback, followed by marking PR #3
ready, merging, tagging, and publishing. It does not mark PR #3 ready, merge, tag, publish, close
hosted CI, run native Claude, prove hooks or host labels, or change production/GA readiness.

T367 refreshes the native Claude prompt-bearing, effective-hook, and live host-label preflight
after T366. Fresh read-only process inventory no longer shows an active native claude CLI
process, so the T363 attribution hard stop is cleared for the native CLI process class. Claude
Code 2.1.168, the resolved SHA-256, installed Engram 0.2.0-beta.1, daemon health, harness
status/doctor ready=true, snippet-only no-write dry-run, canonical vault status, and scoped
obligations doctor all match the preflight contract. Ambient Claude-family helper processes still
exist and must be listed in any future execution report. T367 does not launch native Claude or run
/hooks; T312, T335, and T270 remain default-deny until exact approval names the intended packet.

T368 reruns the same native-Claude/effective-hook/live-host-label preflight and finds that the
T367 attribution-clear state is no longer current. A live native claude CLI process is present
again on ttys004 as PID 34797, started on 2026-06-08 at 11:49:15 local time. That process
hard-stops T312 prompt-bearing proof, T335 /hooks effective-hook proof, and T270 live
host-label proof until it exits naturally or the user gives exact approval for a process action
and a fresh preflight passes. T368 is read-only and does not launch native Claude, run /hooks,
signal the process, mark PR #3 ready, merge, tag, publish, or change the beta scope.

T369 refreshes exact-head local beta validation after T368. ./scripts/local-ci.sh passed on the
current PR #3 candidate, covering git diff --check, cargo fmt --all --check,
cargo check --all-targets, cargo clippy --all-targets -- -D warnings,
CI-like cargo test --all-targets --jobs 1, and cargo doc --no-deps.
./scripts/package-install-smoke.sh also passed, rebuilding the release package, verifying the
tarball checksum, installing the packaged binary into a temporary prefix, confirming
engram 0.2.0-beta.1, starting packaged engram serve --http --memory, and verifying packaged
HTTP /health returned {"status":"ok","service":"engram","version":"0.2.0-beta.1"}.
Hosted run 27138579667 on T368 head 8dad84e37b6bc033904319c675683345c7b60972 still fails all
jobs before workflow-step execution with steps=[], so the remaining beta gate is explicit
release-owner acceptance of the local fallback or restored exact-head hosted CI green, followed by
ready/merge/tag/publish mechanics.

T370 refreshes real-session telemetry confidence for the Brain Harness evidence trail. The
50-trace project-scoped report first failed at 13/50 feedback records (26% coverage), then
improved to 19/50 (38%) after the initial feedback catch-up, and finally passed at 26/50
(52%) after feedback was added only for judgeable recent traces. The 20-trace window now passes
at 18/20 (90%). This strengthens sampled retrieval-feedback evidence, but it is not beta
release approval and does not authorize M6 write-apply, lifecycle cleanup, native Claude,
effective hooks, live host labels, or production/GA readiness.

T371 refreshes local validation after T370 moved the PR head. ./scripts/local-ci.sh passed on the
current candidate tree, covering whitespace diff check, rustfmt, check, clippy, serialized tests,
and docs. ./scripts/package-install-smoke.sh also passed, rebuilding the release package,
verifying the tarball checksum, installing the packaged binary into a temporary prefix, confirming
engram 0.2.0-beta.1, starting packaged engram serve --http --memory, and verifying packaged
HTTP /health returned {"status":"ok","service":"engram","version":"0.2.0-beta.1"}.
Hosted run 27141590404 on T370 head 4249855bee0fe4b33a9bd343d7750ce7a8da368f still fails all
jobs before workflow-step execution with steps=[], so the remaining beta gate stays explicit
release-owner local-fallback acceptance or restored exact-head hosted CI green.

T372 refreshes current PR #3 and native-Claude production-gate evidence after T371. Current hosted
run 27142919365 targets head 3cf0e3d453fe4f02a0e1019bcf79fe8779e72cde and still fails Docs,
Clippy, Test, Format, and Check before workflow-step execution with steps=[]. Claude Code
2.1.168, SHA-256 377f0ecedba8246bdabdf312ce8b7cc8ae1160997b26f5edca352a4a8d61dc78,
installed Engram daemon 0.2.0-beta.1, harness status/doctor ready=true, clean obligations, and
canonical vault counts match the expected preflight state. A live native claude PID 34797 on
ttys004 still blocks unambiguous T312 prompt-bearing proof, T335 /hooks effective-hook proof,
and T270 live host-label proof. T372 is read-only for runtime state and does not launch native
Claude, run /hooks, signal processes, mutate settings or adapters, mark PR #3 ready, merge, tag,
publish, or change the beta scope.

T373 makes stale active-session lint findings more actionable for production lifecycle review.
engram-index/src/lint.rs now includes project, agent, RFC3339 started_at, and age_hours in
each stale_active_session finding while preserving safe_action=none. This improves cleanup
review safety without ending sessions, archiving memory, running lint apply_safe, mutating M6,
changing ranking/orient, launching native Claude, or changing release scope. Focused lint tests,
public lint integration tests, fmt, check, clippy, local CI, and package/install smoke pass for the
T373 candidate.

T374 installs the T373 source behavior into the live local/Codex runtime. Before install,
/Users/yuval.meiri/.local/bin/engram and ./target/release/engram both reported
0.2.0-beta.1 but had different SHA-256 hashes. After cargo build --release and installing the
fresh binary, both paths hash to
2446fe249b0b24745f47fafd356eec62fde2ca585b16c4f865f56d5e7c4c6a6c; the daemon was restarted on
PID 30394, /health returned {"status":"ok","service":"engram","version":"0.2.0-beta.1"}, and
installed engram lint run --scope-project engram --limit 20 --json showed stale active-session
messages with project, agent, started_at, and age_hours. T374 does not run lint apply_safe,
archive sessions or memory, mutate M6, change orient, launch native Claude, mark PR #3 ready,
merge, tag, publish, or change the beta scope.

T375 makes ready harness-doctor lifecycle evidence more actionable. When
engram harness doctor reports a ready harness, the soft lifecycle warning now names the exact
advisory trigger set from the policy: task_start_orient,
before_major_decision_changes_since, after_discovery_record,
before_final_changes_since, before_final_obligations,
before_context_compaction_save, session_end_handoff, and
commit_workflow_consult_memory. Focused harness unit tests, MCP harness integration tests, fmt,
check, clippy, release build, live installed runtime validation, daemon restart, /health, and the
existing dist checksum pass. T375 improves production-gate evidence quality only; it does not
install or mutate hooks/adapters, enforce lifecycle behavior, run lifecycle cleanup or M6, launch
native Claude, mark PR #3 ready, merge, tag, publish, or change the beta scope.

T376 makes lifecycle state machine-readable in harness reports. HarnessStatusReport now includes
a lifecycle object with soft_contract=true, enforced=false, the advisory trigger list, and a
summary message; the CLI text output also prints the lifecycle contract before adapter checks. This
lets agents and release checks inspect adapter readiness separately from advisory lifecycle
compliance without parsing warning text. Focused service and MCP harness tests, fmt, check, clippy,
release build, live installed runtime validation, daemon restart, /health, live JSON/text
harness doctor, live JSON harness status, and the existing dist checksum pass. T376 improves
structured evidence only; it does not install or mutate hooks/adapters, enforce lifecycle behavior,
run lifecycle cleanup or M6, launch native Claude, mark PR #3 ready, merge, tag, publish, or change
the beta scope.

T377 makes MCP tool availability state machine-readable in harness reports. HarnessStatusReport
now includes an mcp_tools object with checked, required_tools, observed_tools,
missing_tools, and message, while preserving the compatibility missing_mcp_tools field. This
removes the ambiguity where missing_mcp_tools=[] could mean either "all required tools were
observed" or "no observed tool list was supplied." The CLI text output now states whether MCP tools
were checked, checked complete, or checked with missing tool names. Focused service and MCP harness
tests, fmt, check, clippy, release build, daemon restart, /health, live installed JSON/text
harness doctor, live MCP checked missing-tool status, and dist checksum validation passed. The
installed CLI does not expose an observed-tool flag, so the checked path was not validated through
CLI text flags. T377 improves structured evidence only; it does not install or mutate
hooks/adapters, enforce lifecycle behavior, run lifecycle cleanup or M6, launch native Claude, mark
PR #3 ready, merge, tag, publish, or change the beta scope.

T378 exposes the T377 checked MCP tool path through the CLI. engram harness status and
engram harness doctor now accept repeatable --observed-mcp-tool <TOOL> flags and pass those
names into the same status/doctor service used by MCP. Omitting the flag preserves the unchecked
behavior. Focused CLI parse/check/clippy validation passed; the rebuilt installed binary hash is
d7e17ae33bdfd48c84fd24070b1d10b17a284c0e31993e7a9b190c7450180b34, the daemon restarted on PID
39185, /health returned ok, and installed CLI checked-complete plus checked-missing smokes
passed. T378 improves release/operator evidence only; it does not install or mutate hooks/adapters,
enforce lifecycle behavior, run lifecycle cleanup or M6, launch native Claude, mark PR #3 ready,
merge, tag, publish, or change the beta scope.

T379 refreshes the implementation-plan completion matrix after T378 closeout. The current matrix
now names PR #3 head c876374db987252f4ad7ed88885ab55f30860b8a, exact-head local CI plus
package-install smoke evidence, hosted run 27153053722 with steps=0 on every job, and the
current native-Claude attribution blocker. T379 is a docs-only freshness correction; it does not
mark PR #3 ready, merge, tag, publish, accept hosted-CI fallback, close hosted CI, run native
Claude, prove effective hooks or live host labels, run lifecycle cleanup, mutate M6, or change the
beta scope.

T380 archives exactly the stale active M6 checkpoint MemoryItem
019dd35d-1a48-7103-b0e2-390225f8b418 after T278 made its active
Memory OS completion is paused at migration review gate guidance stale for the current review
batch. Post-archive memory(get) reports status=archived, current T278 search no longer returns
that checkpoint in active memory results, project-scoped lint no longer reports its stale-feedback
finding, and obligations doctor is clean. T380 is exact lifecycle hygiene only; it does not delete
data, run broad lint apply_safe, run M6 write-apply, change source behavior, change ranking or
orient, accept hosted-CI fallback, mark PR #3 ready, merge, tag, publish, launch native Claude,
prove effective hooks, prove live host labels, or change the beta scope.

T383 restores the current sampled telemetry confidence signal after the 50-trace project window
regressed to 21/50 feedback records (42%). Four judgeable recent traces were scored:
current continuation orient 019ea8a5-f60b-76c2-9e53-016980300077, T382 orient
019ea884-8d3b-7f92-9433-2c3b35159208, T382 changes_since
019ea886-d4a6-7890-ac40-b1f3f52baf2d, and T380 post-archive search
019ea83c-150b-7011-998f-54f61ba618d4. Fresh
telemetry(action="real_session_eval", project="engram", limit=50) now passes at 25/50
feedback records (50%) with zero task failures, zero bad-memory-used, zero stale-memory, and
zero wrong-scope-memory outcomes. The 20-trace window also passes at 10/20 feedback records
(50%) with the same clean outcome counters. T383 changes no source behavior, telemetry formula,
or release scope; it does not accept hosted-CI fallback, mark PR #3 ready, merge, tag, publish,
run M6, run lifecycle cleanup, launch native Claude, prove hooks or host labels, or make the system
production/GA ready.

T384 archives exactly five stale active MemoryItems that project-scoped lint reported as
feedback_stale_active_memory: 019dd080-612a-7540-a028-42991c20ef1b,
019dd083-e014-74f1-95e5-b1eef478e894, 019dd3a8-138d-7453-9991-d724f96a128f,
019dd3e4-9143-7721-9bff-b3fb505c8859, and
019e68a7-3375-7943-8ef0-dc0dde64c8bd. The targets were April/May runtime,
implementation-plan, readiness, gap, and direct-search live-smoke snapshots now superseded by
current T337-T383 evidence. Post-archive project-scoped lint no longer reports those five stale
active-memory findings; the remaining sample is missing-evidence and stale-session debt with
safe_action="none". Vault status remains aligned at 2749/2749, and scoped obligations doctor
is clean. T384 preserves history and changes lifecycle state only; it does not delete data, run
broad lint apply_safe, end sessions, mutate M6, change ranking or orient, accept hosted-CI
fallback, mark PR #3 ready, merge, tag, publish, launch native Claude, prove hooks or host labels,
or make the system production/GA ready.

T385 replaces or archives the next five project-scoped missing-evidence MemoryItems. Three useful
implementation facts now have evidenced active successors: digest extraction apply
019ea8e5-d8ba-7623-abb0-b4151504ad14, orient contract checkpoint
019ea8e6-19b0-7353-bc93-7124bfea5b61, and real-session telemetry eval
019ea8e6-59e6-7980-9a97-e08ab77073be. Their old evidence-less records
019dcaa6-0223-73a2-9fe4-76b61ff14faa,
019dfecb-16ca-71c2-b391-e3c216601590, and
019dfee0-90f8-7a61-a548-c2bddefcf897 are superseded. The stale global rolling handoff
019dddbe-d369-7523-ac91-9bfeb016463b and stale May 6 installed-runtime snapshot
019dfee6-c83a-7c02-a3ed-78fc9e80329b are archived. Post-change project-scoped lint starts at
stale active-session warnings with safe_action="none", canonical vault status is aligned at
2756/2756 after KnowledgeCommit 019ea8e8-f0ee-7f91-b78b-5b96060b9f0a, and scoped obligations
doctor is clean. T385 changes Memory OS lifecycle/evidence state only; it does not run broad
cleanup, end sessions, mutate M6, change ranking or orient, accept hosted-CI fallback, mark PR #3
ready, merge, tag, publish, launch native Claude, prove hooks or host labels, or make the system
production/GA ready.

T386 closes exactly five project-scoped stale active-session lint findings while preserving durable
session history. The completed session records are
019dd063-ee4f-7943-8409-0450bac3a724,
019e7d38-68d7-7652-b62f-3e8e635253ae,
019e7e6e-73e2-7e72-9351-da62e69686af,
019e8470-5d37-7db0-ac21-1725184849e7, and
019e990f-e4fb-7a02-a840-77a38dceab3e. Matching stale coordination rows with empty
components/current_file were unregistered for all but 019e7d38, which had no coordination row.
Post-cleanup project-scoped lint returns no findings, and the targeted coordination rows no
longer appear in coord(list). T386 changes session lifecycle/coordination state only; it does
not delete session events, run broad lint apply_safe, mutate M6, change source behavior, accept
hosted-CI fallback, mark PR #3 ready, merge, tag, publish, launch native Claude, prove hooks or
host labels, clean unrelated coordination-only rows, or make the system production/GA ready.

T387 removes the remaining Engram-project coordination-only orphan rows after T386. The exact
rows were 019e683b-1560-7361-b535-53b012e04aa5,
d238eb81-870e-4da7-8e5d-381014f151b0, and
d96b3edb-f8ac-4d45-b644-92be32d0eae4: all had empty components, null current_file, stale
June 1 heartbeats, and no durable session record from session(get). Source inspection confirmed
coord.unregister deletes only the ephemeral active_session row, and the row contents are
preserved in docs/BRAIN_HARNESS_T387_EXACT_COORD_ORPHAN_CLEANUP_2026-06-08.md. Post-cleanup
coord(list, project="engram") returns count=0, project-scoped lint remains empty, and scoped
obligations doctor remains clean. T387 does not delete session events, run broad lint apply_safe,
mutate M6, change source behavior, accept hosted-CI fallback, mark PR #3 ready, merge, tag,
publish, launch native Claude, prove hooks or host labels, or make the system production/GA ready.

T388 refreshes exact-head release validation after T384-T387 lifecycle/evidence/session/coordination
cleanup. ./scripts/local-ci.sh passed on the T388 head, covering git diff --check,
cargo fmt --all --check, cargo check --all-targets,
cargo clippy --all-targets -- -D warnings, CI-like cargo test --all-targets --jobs 1, and
cargo doc --no-deps. ./scripts/package-install-smoke.sh also passed, rebuilding the release
package, verifying the checksum, installing the packaged binary into a temporary prefix,
confirming engram 0.2.0-beta.1, and verifying packaged HTTP /health returned
{"status":"ok","service":"engram","version":"0.2.0-beta.1"}. T388 refreshes local release
evidence only; it does not accept hosted-CI fallback, mark PR #3 ready, merge, tag, publish,
launch native Claude, prove hooks or host labels, mutate M6, or make the system production/GA
ready.

T390 hardens deterministic memory ranking for ungated, un-gated, and not gated
continuation wording. These variants now share the same continuation gate-negation normalization as
non-gated / non gated, so an M6 continuation prompt does not accidentally match the gated
substring and promote migration-gate context. Explicit modal gate prompts such as "should we
proceed with migration apply" still trigger gate guidance. Focused ranker and search regression
tests, formatter check, cargo check -p engram-index, focused clippy, and git diff --check
passed. Exact-worktree ./scripts/local-ci.sh and ./scripts/package-install-smoke.sh also
passed; the package smoke rebuilt the release archive, verified the checksum, installed the
packaged binary into a temporary prefix, confirmed engram 0.2.0-beta.1, and verified packaged
HTTP /health returned {"status":"ok","service":"engram","version":"0.2.0-beta.1"}. T390
changes only deterministic ranking query classification; it does not accept the hosted-CI fallback,
mark PR #3 ready, merge, tag, publish, launch native Claude, prove hooks or host labels, mutate M6,
or make the system production/GA ready.

T391 tightens deterministic gate-language matching after T390. Bare gate terms such as gate,
gated, must, blocked, cannot, and never now require ASCII word boundaries, while
multi-word or hyphenated phrases such as approval gate, review-gated, do not, should not,
and requires approval remain substring checks. This prevents unrelated words such as gateway,
gatekeeper, and gatedness from creating contextual M6 gate promotion. Focused ranker tests,
existing M6/current-plan search regressions, formatter check, cargo check -p engram-index,
focused clippy, and git diff --check passed. Exact-worktree ./scripts/local-ci.sh and
./scripts/package-install-smoke.sh also passed; the package smoke rebuilt the release archive,
verified the checksum, installed the packaged binary into a temporary prefix, confirmed
engram 0.2.0-beta.1, and verified packaged HTTP /health returned
{"status":"ok","service":"engram","version":"0.2.0-beta.1"}. T391 changes only deterministic
ranking query classification; it does not accept the hosted-CI fallback, mark PR #3 ready, merge,
tag, publish, launch native Claude, prove hooks or host labels, mutate M6, or make the system
production/GA ready.

T392 tightens deterministic decision-gate action matching. Fallback action terms in
asks_for_decision_gate and migration-apply action terms in
asks_for_explicit_migration_apply_gate now require ASCII word boundaries, so incidental substrings
such as mustache, blockchain, unblocked, allowance, and safetybelt do not force
decision-gate mode. Explicit modal prompts and real boundary-delimited gate terms such as must,
blocked, allowed, safety, and write-apply still trigger gate classification. Focused ranker
tests, existing M6/current-plan search regressions, formatter check, cargo check -p engram-index,
focused clippy, and git diff --check passed. Exact-worktree ./scripts/local-ci.sh and
./scripts/package-install-smoke.sh also passed; the package smoke rebuilt the release archive,
verified the checksum, installed the packaged binary into a temporary prefix, confirmed
engram 0.2.0-beta.1, and verified packaged HTTP /health returned
{"status":"ok","service":"engram","version":"0.2.0-beta.1"}. T392 changes only deterministic
ranking query classification; it does not accept the hosted-CI fallback, mark PR #3 ready, merge,
tag, publish, launch native Claude, prove hooks or host labels, mutate M6, or make the system
production/GA ready.

T393 adds an explicit beta release-owner signoff checklist in
docs/BRAIN_HARNESS_T393_BETA_RELEASE_SIGNOFF_CHECKLIST_2026-06-09.md and this release note. The
checklist names the exact local evidence, hosted-CI waiver condition, known beta limitations, and
manual release actions required after signoff. It does not accept the hosted-CI fallback, mark PR #3
ready, merge, tag, publish, launch native Claude, prove hooks or host labels, mutate M6, or make the
system production/GA ready.

T394 hardens ./scripts/package-install-smoke.sh for release-artifact validation. Before extracting
the tarball, the smoke now rejects empty archive listings, absolute or parent-directory paths,
members outside the expected engram-<version>-<host-triple>/ root, and archives missing required
package members. The /health checks now use fixed-string matching. The valid package smoke passed,
including checksum verification, archive-path inspection, temp-prefix install,
engram 0.2.0-beta.1, and packaged HTTP /health; a synthetic wrong-root archive failed closed
before extraction. T394 does not accept the hosted-CI fallback, mark PR #3 ready, merge, tag,
publish, launch native Claude, prove hooks or host labels, mutate M6, or make the system
production/GA ready.

T395 commits Cargo.lock and makes the local/hosted CI plus release-package paths use the locked
dependency graph. .github/workflows/ci.yml, ./scripts/local-ci.sh, ./scripts/package-release.sh,
and ./scripts/package-install-smoke.sh now use --locked for dependency-resolving Cargo commands,
and README development commands reflect the locked workflow. This makes dependency drift fail
explicitly instead of silently changing release inputs between local validation, hosted CI, and
archive creation. Validation passed locked metadata/pkgid checks, script syntax checks, locked
check/clippy, package/install smoke, and the full locked ./scripts/local-ci.sh. T395 does not
accept the hosted-CI fallback, mark PR #3 ready, merge, tag, publish, launch native Claude, prove
hooks or host labels, mutate M6, or make the system production/GA ready.

T396 hardens Claude Code harness readiness reporting against native-hook overclaims. A statically
ready Claude Code status report still returns ready=true when generated adapter files and
settings entries are present, but it now also warns that static readiness does not prove live
effective-hook visibility and that Claude Code /hooks verification is still required before
claiming native Claude hook behavior. Focused service and MCP tests prove the warning is present
without running native Claude, executing /hooks, or mutating real project .claude settings.
Full local CI and package-install smoke passed for the T396 head. T396 does not accept the
hosted-CI fallback, mark PR #3 ready, merge, tag, publish, launch native Claude, prove hooks or
host labels, mutate M6, or make the system production/GA ready.

T397 adds ./scripts/verify-hosted-ci-prestep-blocker.sh as a repeatable hosted-CI waiver
condition check. The script uses gh run view and jq to assert that a run is for the exact
expected head, workflow CI, event pull_request, conclusion failure, and exactly the expected
jobs Check, Test, Format, Clippy, and Docs, with every job completed as failure and
steps=[]. On final T397 PR #3 head 3d01a6ff2e539b23fffcd3666d11a4b5a137df98, it verified
hosted run 27181816572 as a pre-step blocker. T397 does not accept the hosted-CI fallback, mark
PR #3 ready, merge, tag, publish, launch native Claude, prove hooks or host labels, mutate M6, or
make the system production/GA ready.

T398 adds release-package provenance to the tarball itself. ./scripts/package-release.sh now
writes MANIFEST.json with package/version/host metadata, git head, tracked-change state,
Cargo.lock SHA-256, and payload hashes for engram, README.md, LICENSE, CHANGELOG.md, and
RELEASE_NOTES.md. ./scripts/package-install-smoke.sh requires and validates the manifest before
installing the binary, and a corrupted-manifest temporary tarball failed closed on the engram
hash mismatch. T398 does not accept the hosted-CI fallback, mark PR #3 ready, merge, tag, publish,
launch native Claude, prove hooks or host labels, mutate M6, or make the system production/GA
ready.

T399 adds ./scripts/beta-release-gate-report.sh as a release-owner evidence collector. The script
fails closed unless the branch is synced to its upstream, tracked source changes are absent, PR #3
points at the local head, and hosted CI is either green or the hosted pre-step blocker verifier
passes for the exact head. By default it also runs ./scripts/local-ci.sh and
./scripts/package-install-smoke.sh; --quick and skip flags are available for read-only status
checks. T399 does not accept the hosted-CI fallback, mark PR #3 ready, merge, tag, publish, launch
native Claude, prove hooks or host labels, mutate M6, run lifecycle cleanup, or make Engram
production/GA ready.

T400 adds ./scripts/native-claude-gate-preflight.sh as a repeatable read-only production-gate
preflight for native Claude prompt-bearing proof, effective-hook visibility, and live host labels.
The current observed Claude Code baseline is 2.1.169 at
/Users/yuval.meiri/.local/share/claude/versions/2.1.169 with SHA-256
86d8b820ad7eed50e50a130706d3dc5ef70696f91194de1b3897a842182afe3a. The script also verifies
branch sync, installed Engram daemon state, harness status/doctor, snippet-only dry-run drift,
obligations doctor, canonical vault alignment, and live native Claude CLI processes. The current
report is still blocked by live native Claude PID 34797 on ttys004; strict mode exits 2 until
that blocker clears. T400 also recompiled the canonical vault back to 2814/2814 generated files
after a new MemoryItem write. T400 does not launch native Claude, run /hooks, signal processes,
accept the hosted-CI fallback, mark PR #3 ready, merge, tag, publish, mutate M6, run lifecycle
cleanup, or make Engram production/GA ready.

T401 adds machine-readable --json output to ./scripts/native-claude-gate-preflight.sh and
corrects the T400 tracked evidence head to the pushed commit
197ddc67c4bf65baad838919e7b992af31593539. JSON mode exposes gate_state, branch/upstream/head
state, tracked/untracked source state, Claude path/version/hash, daemon state, harness status and
doctor summaries, snippet-only dry-run state, obligations, vault state, native Claude processes,
blockers, and explicit non-action flags. Strict JSON mode still exits 2 while blocked. T401 does
not launch native Claude, run /hooks, signal processes, accept the hosted-CI fallback, mark PR #3
ready, merge, tag, publish, mutate M6, run lifecycle cleanup, or make Engram production/GA ready.

T402 adds machine-readable --json output to ./scripts/beta-release-gate-report.sh. JSON mode
emits branch/upstream/head state, tracked source state, PR draft/merge/check state, hosted CI state,
local CI state, package/install smoke state, release_owner_decision_required, and
release_actions_performed. Step logs go to stderr in JSON mode so stdout remains parseable. T402
does not accept the hosted-CI fallback, mark PR #3 ready, merge, tag, publish, mutate M6, run
lifecycle cleanup, or make Engram production/GA ready.

T403 adds machine-readable --json output to ./scripts/verify-hosted-ci-prestep-blocker.sh.
JSON mode emits the verified run, expected/current head, workflow, event, expected jobs, per-job
status/conclusion/step counts, hosted_ci_fallback_accepted=false, and
release_actions_performed=false after all existing checks pass. Failed checks remain non-zero and
do not emit a success JSON object. T403 does not accept the hosted-CI fallback, mark PR #3 ready,
merge, tag, publish, mutate M6, run lifecycle cleanup, or make Engram production/GA ready.

T404 embeds the T403 hosted verifier JSON inside the T402 beta release-gate report JSON. When
./scripts/beta-release-gate-report.sh --json verifies hosted CI via the pre-step blocker path,
hosted_ci.verifier now carries the exact verifier success object with run/head/workflow/job
evidence and non-action flags. The text report and green-hosted-CI path remain unchanged, and a
green hosted-CI path reports hosted_ci.verifier=null. T404 does not accept the hosted-CI
fallback, mark PR #3 ready, merge, tag, publish, mutate M6, run lifecycle cleanup, or make Engram
production/GA ready.

T405 adds explicit release-review state to ./scripts/beta-release-gate-report.sh. Text and JSON
reports now include release_gate_state, ready_for_release_owner_review, and
hosted_ci_fallback_decision_required; JSON reports also include remaining_release_actions.
Full evidence with verified hosted pre-step blocker, local CI, and package/install smoke reports
fallback_release_owner_decision_required with review readiness true. Quick or skipped local
evidence reports evidence_incomplete. T405 does not accept the hosted-CI fallback, mark PR #3
ready, merge, tag, publish, mutate M6, run lifecycle cleanup, or make Engram production/GA ready.

T406 adds ./scripts/verify-published-release-install.sh for post-publish install verification.
The script downloads the expected GitHub release archive and checksum, or validates a local
--asset-dir mirror, then delegates to ./scripts/package-install-smoke.sh with
SKIP_PACKAGE_BUILD=1 and manifest expectations for the signed-off head. T406 also renames the
beta report's final remaining action to verify_published_release_install. T406 does not create a
release, upload assets, accept the hosted-CI fallback, mark PR #3 ready, merge, tag, publish,
mutate M6, run lifecycle cleanup, or make Engram production/GA ready.