[joslat]

What

LocalMemoryStore now guards its registered memory providers:

• every provider call has a timeout;
• a failing provider can't break providerStatus();
• recalled results go through governance on the way back;
• replication audits say which provider received the copy.

Why

Closes #2332

Governance ran before anything reached a provider, but nothing after that. A hung provider stalled memory.write / memory.search with no limit. One throwing or hung status() broke providerStatus() for everyone. And whatever a provider returned from search() went straight to the agent, including classes it was never sent and secret-looking text that would have been rejected on write.

How

All in packages/squad-sdk/src/memory/index.ts:

• Timeouts. withProviderTimeout() bounds provider write, search and status calls. The limit is configurable through the new LocalMemoryStoreOptions.providerTimeoutMs, with a default of 5000 ms; non-positive or non-finite values fall back to the default. A timeout throws ProviderTimeoutError, which goes through the existing safe provider-error audit path. The call isn't cancelled, so a slow write may still land in the provider later; the option's doc comment says so.
• providerStatus(). Each provider's status() is wrapped, so a throw or timeout becomes { available: false, reason: '<name> provider status failed (<ErrorName>); raw provider error text omitted' }, and the other providers are still reported.
• Recall checks.
  • A result is kept only if its class is one the provider declares and one write() routes to providers (LOCAL, DECISION, POLICY), and its title and snippet don't classify as FORBIDDEN.
  • Dropped results produce one reject audit record per provider with a count, and no content.
  • loadGuidance goes through the same normalizeLoadGuidance helper local entries use, falling back to the class default.
• Audit. The write record for a replicated copy now carries provider: <id>, the same as the error records.
• The MemoryProvider doc comment describes the return-path contract.

No interface signatures change, and nothing in the CLI registers a provider today, so shipped Squad behaves the same. The changes take effect for SDK users who pass registeredProviders.

I left the scope, limit, ToolRegistry wiring and delete propagation out on purpose. Those need a design decision first (listed in the issue).

Heads-up: this changes the same line as #2319: the safeProviderErrorReason signature, which there accepts the host adapter and here adds 'status'. Whichever merges second needs a one-line rebase, and I'll do it.

Testing

New tests in test/memory-providers.test.ts:

• a hung search times out, local results still come back, and the timeout is audited;
• a hung write times out without failing the local write;
• the replication audit carries the provider id;
• throwing and hung status() providers are reported as unavailable next to a healthy one;
• recalled results with an undeclared class, a never-routed class (COPILOT_MEMORY) or a credential-like snippet are dropped and counted, and the audit doesn't contain the snippet;
• an invalid recalled loadGuidance falls back to the class default.

Results:

• With the memory/index.ts change reverted, all 6 new tests fail; the two timeout tests hang until the test timeout.
• test/memory-providers.test.ts, memory-governance, memory-value-benchmark, tools, cli/state-mcp and cli/state-mcp-lifecycle: 174 passed.
• npm run lint: clean.

---

⚠️ Quick Check

• If SDK/CLI source files changed: changeset added (.changeset/memory-provider-guards.md, sdk patch)

PR Readiness Checklist

Branch & Commit

• Branch created from dev (not main)
• Branch is up to date with dev (9d7dff8)
• Verified diff contains only intended changes (3 files)
• PR is not in draft mode
• Commit history is clean

Build & Test

• npm run lint passes
• Relevant tests pass
• Conflicts with other open PRs: one line with fix(sdk): keep memory search working when the host-injected adapter has no client #2319 (noted above), none with the rest

Docs

• No docs describe registeredProviders; the option and the provider contract are documented in the source

---

Breaking Changes

None. The interface is unchanged. A provider that returned results outside its declared classes, or FORBIDDEN-looking text, will now have those results dropped, which is the intended behavior.

Waivers

None.

[github-actions]

🛫 PR Readiness Check

ℹ️ This comment updates on each push. Last checked: commit b40430c

PR Scope: 📦🔧 Mixed (product + infrastructure)

⚠️ 3 item(s) to address before review

Status	Check	Details
❌	Single commit	2 commits — consider squashing before review
✅	Not in draft	Ready for review
✅	Branch up to date	Up to date with dev
❌	Copilot review	No Copilot review yet — it may still be processing
✅	Changeset present	Changeset file found
✅	Scope clean	No .squad/ or docs/proposals/ files
✅	No merge conflicts	No merge conflicts
✅	Copilot threads resolved	No Copilot review threads
❌	CI passing	13 check(s) still running

Files Changed (3 files, +208 −7)

File	+/−
.changeset/memory-provider-guards.md	+5 −0
packages/squad-sdk/src/memory/index.ts	+79 −7
test/memory-providers.test.ts	+124 −0

Total: +208 −7

---

This check runs automatically on every push. Fix any ❌ items and push again.
See CONTRIBUTING.md and PR Requirements for details.

[github-actions]

🟡 Impact Analysis — PR #2333

Risk tier: 🟡 MEDIUM

📊 Summary

Metric	Count
Files changed	3
Files added	1
Files modified	2
Files deleted	0
Modules touched	3
Critical files	1

🎯 Risk Factors

• 3 files changed (≤5 → LOW)
• 3 modules touched (2-4 → MEDIUM)
• Critical files touched: packages/squad-sdk/src/memory/index.ts

📦 Modules Affected

root (1 file)

• .changeset/memory-provider-guards.md

squad-sdk (1 file)

• packages/squad-sdk/src/memory/index.ts

tests (1 file)

• test/memory-providers.test.ts

⚠️ Critical Files

• packages/squad-sdk/src/memory/index.ts

---

This report is generated automatically for every PR. See #733 for details.

[tamirdresher]

Thanks Jose, another sharp catch on the memory governance boundary. I verified on dev that packages/squad-sdk/src/memory/index.ts had no timeout on provider.write (:838) or provider.search (:936), ran an unguarded Promise.all in providerStatus() (:1112), and trusted recalled provider.search() hits without re-checking supportedClasses, FORBIDDEN patterns, or loadGuidance (:937). Bounding write, search, and status with withProviderTimeout and filtering recalled hits through his.classify and
ormalizeLoadGuidance with a content-free
eject audit count closes that return-path trust gap cleanly.

I verified the merge-tree against dev is clean and approved the CI workflow runs. As you noted, line 411 of packages/squad-sdk/src/memory/index.ts (safeProviderErrorReason) touches the same signature as #2319. We are in a release freeze for 2.0.0-preview.2 right now and will merge your memory series in order (#2315, #2317, #2319, then #2333) right after the release ships, with that one-line combine on safeProviderErrorReason (provider: Pick<MemoryProvider, 'name'> and operation: 'write' | 'search' | 'status').

[joslat]

Rebased on dev now that #2319 is in, so you don't need to do the combine. safeProviderErrorReason takes provider: Pick<MemoryProvider, 'name'> and operation: 'write' | 'search' | 'status', and everything else applied cleanly. Still one commit. The memory and tools suites (memory-providers, memory-governance, memory-value-benchmark, memory-class-normalization, memory-store-resolution, tools, cli/state-mcp, cli/state-mcp-lifecycle) give 190 passed, and lint is clean.