Maintainer triage summary for YOMXXX open PR queue

[YOMXXX]

Maintainer triage summary for my open PR queue

I reviewed my currently open PRs and added per-PR status / merge-order notes where useful.

Current status

• All 36 open PRs are mergeable.
• All checks are passing.
• I did not find any currently open PR that should be closed as a duplicate.
• Some PRs touch the same test/config files, so the later PR in each group may need a small rebase. These are mechanical conflicts, not duplicate functionality.

Suggested merge priority

1. Security / data-boundary fixes

These are narrow, high-value, and mostly low product-risk:

• fix(gateway): harden CORS / Host header / startup host binding #39 Gateway CORS / Host header / startup host binding hardening
• fix(tcvdb): escape session filter literals #232 TCVDB session filter literal escaping
• fix(gateway): reject invalid and oversized JSON bodies #237 Gateway malformed / oversized JSON request handling
• fix(offload): keep backend TLS verification enabled #242 Backend TLS verification + timeout propagation
• fix(conversation): scope L0 search candidates by session #244 Conversation search session scoping
• fix(offload): sanitize session storage paths #287 Offload session storage path sanitization
• fix(offload): keep MMD files inside storage dir #288 MMD filename/path boundary hardening
• fix(offload): sanitize ref and skill paths #289 Ref and L4 skill path sanitization
• fix(profile): sanitize pulled scene filenames #368 Pulled profile filename sanitization
• fix(scene): normalize indexed filenames on read #369 Scene index filename normalization

Merge-order notes:

• fix(offload): sanitize session storage paths #287 / fix(offload): keep MMD files inside storage dir #288 / fix(offload): sanitize ref and skill paths #289 overlap in src/offload/storage.ts; whichever lands later will need a small rebase.
• fix(profile): sanitize pulled scene filenames #368 and fix(scene): normalize indexed filenames on read #369 are adjacent but distinct profile/scene-index boundaries and do not appear to conflict with each other.

2. Embedding / recall correctness

These combine cleanly and can be reviewed as an independent cleanup set:

• fix(embedding): cap remote batch size at 10 #238 Cap remote embedding batch size at 10
• fix: reindex embeddings after config drift #270 Reindex embeddings after embedding config drift
• fix(embedding): forward proxy and timeout config #271 Forward embedding proxy + timeout config
• fix(search): honor recall embedding timeout #272 Honor recall embedding timeout in agent-callable search tools
• fix(recall): report effective fallback strategy #281 Report effective fallback search strategy

I tested these together locally; they combine cleanly and focused tests pass.

Merge-order note:

• fix(search): honor recall embedding timeout #272 and fix(core): honor layer model config for runners #280 both add coverage in src/core/tdai-core.test.ts; whichever lands later will need a small test-file rebase.

3. Package / Hermes / runner integration

These combine cleanly:

• fix(runner): avoid duplicate system prompt #241 Avoid duplicate system prompt for OpenClaw versions with systemPromptOverride
• fix(package): exclude python test artifacts #243 Exclude Python test/cache artifacts from npm package
• fix(hermes): install durable user plugin link #245 Install a durable Hermes user plugin link

I tested these together locally; focused tests and shell syntax checks pass.

4. Config / timeout consistency

These are complementary timeout/config propagation fixes:

• fix(config): align offload timeout default #275 Align offload backend timeout manifest default with runtime default
• fix(offload): honor local L2 timeout config #276 Honor local L2 timeout config
• fix(l1): honor runner timeout config #277 Honor L1 extraction/dedup runner timeout config
• fix(config): expose offload manifest fields #278 Expose supported offload manifest fields
• fix(pipeline): honor L2/L3 runner timeout config #279 Honor L2/L3 runner timeout config
• fix(core): honor layer model config for runners #280 Honor layer-specific model config for runner creation

Merge-order notes:

• fix(config): align offload timeout default #275 and fix(config): expose offload manifest fields #278 both edit openclaw.plugin.json; whichever lands later will need a small manifest rebase.
• fix(search): honor recall embedding timeout #272 and fix(core): honor layer model config for runners #280 both add coverage in src/core/tdai-core.test.ts.

5. Cleaner / timezone behavior

These are complementary cleanup semantics:

• fix(cleaner): allow one-day retention cutoff #284 Allow one-day retention cutoff
• fix(cleaner): preserve local day in negative timezones #285 Preserve local day in negative timezones
• fix(cleaner): schedule cleanup in configured timezone #286 Schedule cleanup in configured timezone

Merge-order note:

• fix(cleaner): allow one-day retention cutoff #284 / fix(cleaner): preserve local day in negative timezones #285 / fix(cleaner): schedule cleanup in configured timezone #286 all touch src/utils/memory-cleaner.test.ts and related cleaner/time helpers, so later PRs will need small rebases.

6. L1 memory behavior requiring product judgment

These are all mergeable and tested, but they intentionally affect memory behavior, so I would prefer explicit maintainer agreement before merging:

• fix(record): harden L1 dedup JSON parsing #251 Harden L1 dedup JSON parsing
  • Low product risk; parser robustness only.
• fix(prompts): preserve scoped instruction context #258 Preserve scoped instruction context
  • Prompt-only guardrail for scoped instructions; lowest-risk [Bug] L1 指令提取缺少场景限定机制，用户的场景受限偏好被泛化为全局规则 #48 fix.
• fix(dedup): cap merged memory priority #265 Cap merged memory priority
  • Deterministic guardrail; trades priority amplification for stability.
• fix(l1): drop ungrounded extracted memories #266 Drop ungrounded extracted memories
  • Precision-over-recall trade-off; may reject legitimate abstractions.
• feat(l1): pre-extract direct memory rules #269 Pre-extract direct memory rules
  • Deterministic fast path; introduces hardcoded rule patterns and fixed phrasing/priority.
• fix(l1): treat content-filtered extraction as failed #430 Treat content-filtered L1 extraction as failed
  • Prevents cursor advancement on blocked/unparseable extraction.
• fix(hermes): require explicit memory indexing #434 Require explicit Hermes memory indexing
  • Preferred/canonical [Bug] Hermes memory_tencentdb does not mirror explicit memory() writes into TencentDB memory search #417 fix; strict only for explicit memory writes.

Merge-order notes:

• fix(record): harden L1 dedup JSON parsing #251 and fix(dedup): cap merged memory priority #265 both add src/core/record/l1-dedup.test.ts; later PR needs test-file rebase.
• fix(l1): drop ungrounded extracted memories #266 and feat(l1): pre-extract direct memory rules #269 both add src/core/record/l1-extractor.test.ts; later PR needs test-file rebase.
• fix(prompts): preserve scoped instruction context #258 and fix(dedup): cap merged memory priority #265 combine cleanly.

Local verification notes

For the grouped checks I ran locally with COREPACK_ENABLE_AUTO_PIN=0:

• Embedding/search group focused tests passed.
• Package/Hermes/runner group focused tests passed.
• Pipeline lifecycle focused test passed.
• The L1 product-judgment PRs each passed their focused tests and git diff --check.

Local environment note: my Node version is v22.14.0, below the project engine >=22.16.0, so pnpm/tsdown may emit engine warnings locally, but the targeted tests completed successfully.