Remove memory_search and memory_get from SUBAGENT_TOOL_DENY_ALWAYS.
These are read-only tools with no side effects that are essential for
multi-agent setups relying on shared memory for context retrieval.
Rationale:
- Read-only tools (memory_search, memory_get) have no side effects and
cannot modify state, send messages, or affect external systems
- Other read-only tools (read, web_search, web_fetch) are already
available to sub-agents by default
- Multi-agent deployments with shared knowledge depend on memory tools
for context retrieval
- The workaround (tools.subagents.tools.alsoAllow) works but requires
manual configuration that contradicts memorySearch.enabled: true
Fixes#55385
* gateway: prefer transcript model in sessions list
* gateway: keep live subagent model in session rows
* gateway: prefer selected model until runtime refresh
* gateway: simplify session model identity selection
* gateway: avoid transcript model fallback on cost-only reads
loadChannelOutboundAdapter (via createChannelRegistryLoader) was reading
from getActivePluginRegistry() — the unpinned active registry that gets
replaced whenever loadOpenClawPlugins() runs (config schema reads, plugin
status queries, tool listings, etc.).
After replacement, the active registry may omit channel entries or carry
them in setup mode without outbound adapters, causing:
Outbound not configured for channel: telegram
The channel inbound path already uses the pinned registry
(getActivePluginChannelRegistry) which is frozen at gateway startup and
survives all subsequent registry replacements. This commit aligns the
outbound path to use the same pinned surface.
Adds a regression test that pins a registry with a telegram outbound
adapter, replaces the active registry with an empty one, then asserts
loadChannelOutboundAdapter still resolves the adapter.
Fixes#54745Fixes#54013
When approvals.exec.targets routes to a Telegram DM, the recipient
receives inline approval buttons but may not have explicit
channels.telegram.execApprovals configured. This adds a fallback
isTelegramExecApprovalTargetRecipient check so those DM recipients
can act on the buttons they were sent.
Includes accountId scoping for multi-bot deployments and 9 new tests.
* fix(plugins): reuse active registry for sub-agent tool resolution
* test(plugins): harden resolveRuntimePluginRegistry with per-field, caller-shape, and cold-start tests
Add 11 regression tests covering:
- R1: Per-field isolation (coreGatewayHandlers, includeSetupOnlyChannelPlugins,
preferSetupRuntimeForChannelPlugins each independently prevent fallback;
empty onlyPluginIds[] treated as non-gateway-scoped)
- R2: Caller-shape regression (tools.ts, memory-runtime.ts,
channel-resolution.ts shapes fall back; web-search-providers.runtime.ts
with onlyPluginIds does not)
- R3: Cold-start path (null active registry falls through to loadOpenClawPlugins)
Add debug logging to resolveRuntimePluginRegistry recording which exit path
was taken (no-options, cache-key-match, non-gateway-scoped fallback, fresh load).
* refactor: simplify plugin registry resolution tests and trim happy-path debug logs
* fix(plugins): address review comments on registry fallback
- Fix cold-start test assertion: loadOpenClawPlugins always activates
the registry (shouldActivate defaults to true), so getActivePluginRegistry()
is not null after the call. Updated assertion to match actual behavior.
- Add safety comment documenting why the non-gateway-scoped fallback is
safe despite cache-key mismatch: single-gateway-per-process model means
sub-agents share workspaceDir, config, and env with the gateway.
* test(plugins): restructure per-field isolation tests to avoid load timeouts
Test isGatewayScopedLoad directly instead of going through the full
resolveRuntimePluginRegistry path which triggers expensive plugin
discovery. This fixes the includeSetupOnlyChannelPlugins test timing
out in CI while providing more precise coverage of the predicate.
* fix(plugins): expand safety comment to address startup-scoped registry concern
* fix(plugins): scope subagent registry reuse to tool loading
---------
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
Fixes#46185.
Verified:
- pnpm install --frozen-lockfile
- pnpm build
- pnpm test -- extensions/line/src/markdown-to-line.test.ts src/tts/prepare-text.test.ts
Note: `pnpm check` currently fails on unchanged `extensions/microsoft/speech-provider.test.ts` lines 108 and 139 on the rebased base, outside this PR diff.
- Two-pass line splitting: first slice at maxChars (unchanged for Latin),
then re-split only CJK-heavy segments at chunking.tokens. This preserves
the original ~800-char segments for ASCII lines while keeping CJK chunks
within the token budget.
- Narrow surrogate-pair adjustment to CJK Extension B+ range (D840–D87E)
only, so emoji surrogate pairs are not affected. Mixed CJK+emoji text
is now handled consistently regardless of composition.
- Add tests: emoji handling (2), Latin backward-compat long-line (1).
Addresses Codex P1 (oversized CJK segments) and P2s (Latin over-splitting,
emoji surrogate inconsistency).
- Use code-point length instead of UTF-16 length in estimateStringChars()
so that CJK Extension B+ surrogate pairs (U+20000+) are counted as 1
character, not 2 (fixes ~25% overestimate for rare characters).
- Change long-line split step from maxChars to chunking.tokens so that
CJK lines are sliced into token-budget-sized segments instead of
char-budget-sized segments that produce ~4x oversized chunks.
- Add tests for both fixes: surrogate-pair handling and long CJK line
splitting.
Addresses review feedback from Greptile and Codex bots.
The QMD memory system uses a fixed 4:1 chars-to-tokens ratio for chunk
sizing, which severely underestimates CJK (Chinese/Japanese/Korean) text
where each character is roughly 1 token. This causes oversized chunks for
CJK users, degrading vector search quality and wasting context window space.
Changes:
- Add shared src/utils/cjk-chars.ts module with CJK-aware character
counting (estimateStringChars) and token estimation helpers
- Update chunkMarkdown() in src/memory/internal.ts to use weighted
character lengths for chunk boundary decisions and overlap calculation
- Replace hardcoded estimateTokensFromChars in the context report
command with the shared utility
- Add 13 unit tests for the CJK estimation module and 5 new tests for
CJK-aware memory chunking behavior
Backward compatible: pure ASCII/Latin text behavior is unchanged.
Closes#39965
Related: #40216
