* fix(memory): enforce canonical SecretRef resolution
Make Gateway runtime snapshots the exclusive owner of memory SecretRef materialization. Bind embedding credentials and headers to provider-owned destinations, and fence per-agent stale reuse by the provider destination/auth contract.
Release note: Memory search resolves secret references through configured provider policy and keeps embedding credentials scoped to their intended destination.
* fix(lmstudio): preserve resolved memory headers
memory remote headers are already materialized by the Gateway snapshot and now bypass SecretRef re-resolution; provider-owned headers retain canonical resolution; final loopback request proof covers literal preservation and precedence.
* fix(memory): bind stale credentials to auth owners
Resolve memory adapter credential owners from snapshot manifest metadata, conservatively fail cold when metadata is absent, and prove Gemini/Google destination changes plus zero-egress unresolved refs.
* fix(memory): scope compatible embedding credentials
Apply destination ownership to the core compatible adapter while preserving destination-owned credentials and intentionally unauthenticated endpoints. Distinguish loopback principals, consolidate duplicate security tests, and verify the final credential boundary through a live isolated Gateway request.
* test(memory): align destination auth precedence
* fix(memory): bind credentials to query identity
Include URL query parameters in embedding destination ownership so provider credentials and headers never cross tenant boundaries.
* fix(memory): preserve query-bound embedding destinations
* test(memory-core): release cached state handles before removing fixture dirs
Six memory-core fixtures remove their temporary state directory while the
per-agent or shared SQLite handles opened underneath it are still cached, so
Windows fails the removal with EBUSY while Linux unlinks the open files and
stays green.
Release both handles before each removal, in the ordering merged for this
class: the agent close releases its leases through shared state and reopens
it, so the shared store has to be released after it.
* test(memory-core): release cached handles in the shared fixture harness and cli teardown
* fix(memory-core): read dreaming narrative from the terminal reply instead of racing the session store (#123360)
A completed run's terminal reply is captured by the run registry at
completion time and survives the window where the session store lags the
run. Prefer it over readSettledNarrativeText, whose 1,250ms settle budget
expires while a first-finishing phase's cleanup keeps the store empty,
discarding every other phase's diary entry on multi-phase nights.
The settle loop remains the bounded fallback when the run result carries
no visible terminal reply.
* fix(memory-core): respect explicit non-visible terminal replies (#123360)
Review feedback: absent, silent, and empty terminal results were all
falling back to transcript polling. An explicit silent/empty terminal
reply is authoritative no-text — polling could resurface an older
narrative over it. Classify the terminal reply into visible /
non-visible / absent and only poll the store for absent (legacy
runtime) results.
Also compresses the selection to net +13 production lines.
* test(memory-core): lock the non-visible terminal reply contract (#123360)
Adversarial review hardening: cover the message-tool-not-called
variant, assert that text beside a non-visible disposition is ignored,
cover whitespace-only visible text, and use fake timers so a
regression fails fast instead of after the settle budget.
Doctor migrations and auth-profile writes open per-agent and shared state
databases under the fixture's temporary directory. Clearing the plugin
state store or the runtime auth snapshots does not release those handles,
so Windows fails the directory removal with EBUSY while Linux unlinks the
open files and stays green.
Close the cached databases before each removal, matching the ordering the
zalouser and zalo fixtures already use.
The memory CLI resolved --agent by returning the caller's string verbatim,
so an id that is not configured produced a confident empty result:
`memory status` rendered a panel for it, `memory index` fabricated a
workspace-<id> path, and `memory search` reported No matches. A typo read
as an empty memory rather than a nonexistent agent, while hooks, status
--usage, capability, migrate, and session targets already rejected unknown
ids.
Consolidate that duplicated check into resolveConfiguredAgentId beside the
agent roster owner, reuse it at the matching core sites, and route memory
to it through the existing memory-core host-runtime facade so no new
plugin SDK surface is added.
The canonical hint uses formatCliCommand rather than a literal: under a
profile or container the bare command is wrong, so consolidating on a
literal would have regressed the hooks and migrate hints and left the
status, capability, and session-target hints unrunnable.
* test(control-ui): add --operator-scopes flag to the mock dev server
* feat(control-ui): simplified settings experience for non-admin operators
Non-admin browsers previously saw every settings page, many of which
dead-ended or rendered enabled controls whose RPCs fail with
'missing scope: operator.admin'.
- config.schema drops from operator.admin to operator.read: the schema is a
static document describing options whose values are already readable via
read-scoped config.get; admin-only schema only broke read-only settings
rendering (Automation/Infrastructure/AI Agents/Communications showed
'Schema unavailable. Use Raw.').
- Settings sidebar and settings search hide admin-only routes (custodian,
labs, updates, automation, infrastructure, mcp, security, secrets,
cloud-workers, communications, ai-agents, model-setup) for non-admin
viewers; legacy gateways without advertised scopes keep the full UI.
- Channels, Devices, Worktrees, Memory Import, Profile gate their mutation
controls on actual scopes with 'Browsing only…' notices instead of
enabled-but-failing buttons; Devices no longer fires device.pair.list /
exec.approvals.get without the scopes to call them (kills the two red
error callouts on page load).
- Scope-upgrade banner: dismissing it in the guidance phase (no in-app
upgrade path) now hides it fully instead of leaving a permanent chip.
- Config write coordinator surfaces scope refusals as a visible
admin-required error instead of silently resolving false.
* test(control-ui): advertise config.schema in the mock dev gateway
ensureSchemaLoaded now checks method advertisement + scope before loading
the schema; the mock harness must advertise config.schema like a real
gateway does or schema-driven settings pages render empty in the mock.
* fix(control-ui): close the worktree create draft on scope downgrade
* perf(doctor): isolate memory health artifact
Doctor lint loaded the broad Memory Core API barrel only to register health checks and read isolated check IDs. That synchronously pulled the full memory public graph into the first lint run, consuming most of the 120-second test budget.
Load a dedicated doctor-health public artifact instead and verify it is packaged. The bisect boundary was 9de3ca5fc9 (#125571); because that commit only adds upgrade-test assets, it exposed a pre-existing runner-sensitive cost rather than introducing the expensive import path.
* test(control-ui): restore device lifecycle test boundary
* perf(control-ui): lazy-load settings sidebar
* fix(ui): recheck access after confirmations
* fix(control-ui): gate presence-driven device reloads on pairing access
The presence connectivity-change path still called device.pair.list without
operator.pairing, the same invariant the pair-event and poller paths already
guard; a limited browser got a doomed RPC on every connectivity change.
* fix(control-ui): fail open on schema loads for legacy scope-less gateways
canCallGatewayMethod hardened to strict advertisement+scope checks (#125478),
which made the new ensureSchemaLoaded gate silently skip config.schema for
legacy hellos without advertised scopes or a method list. Schema loads now
skip only on a definitive denial (method advertised absent, or advertised
scopes without operator.read), reusing the fail-open hasOperatorReadAccess
semantics the rest of the non-admin UI uses; regression test pins the
legacy snapshot path.
* test(control-ui): split schema-access coverage into its own file
runtime-config-capability.test.ts crossed the max-lines cap; the legacy
fail-open regression and its denial counterpart move to a colocated
schema-access test file.
* fix(scripts): keep mapped Vitest lanes at their measured no-output floor
The codex extension shard legitimately works in silence beyond 300s under
the default reporter (measured 61s import + 293s testing at ~95% CPU); the
CI-wide OPENCLAW_VITEST_NO_OUTPUT_TIMEOUT_MS=300000 env override shrank the
lane below that and the watchdog killed healthy runs, flipping with
incidental flake output (#125825). Per-config entries in
VITEST_CONFIG_NO_OUTPUT_TIMEOUT_MS now act as measured silence floors: a
global env value may widen a mapped lane's window but no longer shrinks it;
unmapped configs and the explicit '0' disable keep env verbatim. Adds the
codex extension lane to the map at the extra-long tier (same class as the
discord entry from #123025).
Linux signal-zero probes succeed for zombie processes. Reclaim memory promotion and session usage locks only when their exact zombie owner is still current, and route shared test waits through the canonical zombie-aware PID helper.
memory_search corpus=all silently swallowed a returned memory-manager
error (missing sqlite, embedding misconfiguration, corrupted index):
the error branch was gated on !shouldQuerySupplements, so the tool
served wiki-only results in a payload with no error/warning field and
recorded no cooldown — a degraded search read as complete, contradicting
the tool description's own disabled-flag contract. The thrown-error path
already surfaced this; only the returned-error path was silent.
Record the cooldown in both cases and, when supplements still serve,
carry a warning naming the unavailable memory corpus in the payload.
Move llama.cpp chat and local embeddings onto a verified externally managed llama-server runtime. Remove the in-process native runtime, forked embedding workers, and node-llama-cpp dependency while preserving guided setup, local GGUF models, tool-capable agent runs, diagnostics, and operator docs.
Replace assertion-driven cleanup with real runtime narrowing and typed owner contracts across memory-core dreaming.
Reject the unsound one-file lint ratchet and document repository-wide rules against checker-gaming and baseline-driven enforcement.
Co-authored-by: Amp <amp@ampcode.com>
* refactor(memory-core): split index manager along the inheritance chain
* fix(memory): restore the manager lazy boundary
* refactor(memory-core): delete dead manager plumbing
* chore(lint): ratchet max-lines baseline after memory manager split
* fix(memory): keep the manager out of the runtime facade
* test(memory): cover transitive manager facade imports
Josh Lehman