v3.71.3
v3.71.3 — Claude Code 2.1.318 adoption, narrow band, AutoMemory + daemon fixes
Added 3
- Headless workers read credential_expires_at and emit a warning 10 days before a sign-in ends
- The agentModels option sets a model per agent type for spawns that chose none
- AutoMemory adds a read-only read <namespace> <key> subcommand
Changed 8
- Headless workers pass --permission-mode auto so they no longer stall in Claude Code 2.1.302's manual default
- The pre-bash gate fails closed with onFailure set to block behavior changed · inferred, marked in the notes below
- The nested-leaf and nested-queen-leaf agents run at effort low with omitClaudeMd, cutting about 64k tokens per spawn
- The guard uses tool.check fields isReadOnly, agentId and decider person, reducing read-only check time from about 0.3–0.5 ms to 0.13–0.25 ms
- Hot console tools use isDeferred
- Text uses weight, with a fallback for Claude Code before 2.1.299
- The README shows the one-command claude plugin install form with --marketplace
- The browser 3.0.1 package uses agent-browser 0.41.2
Fixed 12
- The console band above the prompt fits narrow panes, measuring rows in terminal columns and counting overflow parts with +N instead of cutting words
- The AutoMemory JSON store no longer swallows write failures or overwrites an unreadable store with []
- AutoMemory imports verify what was written and exit non-zero on failure
- An explicit daemon stop is respected by every auto-start path until the next daemon start
- hooks_intelligence_learn refuses a trajectoryIds selection instead of learning over everything behavior changed · inferred, marked in the notes below
- Guidance in five plugins no longer routes research documents or trace spans to agentdb_context-synthesize
- claims_handoff removes the claim from the stealable list
- Completed claims leave claims_load and claims_rebalance
- task_create with assignTo marks the assigned agents busy
- The performance 3.0.1 attention benchmark range check requires both min and max, and the comparison is de-biased
- The cost-tracker counts each Claude Code response once
- hive-mind spawn --claude --model works
2 changes to check, marked in the notes below
From ruflo
Patch release: @claude-flow/cli, claude-flow, ruflo 3.71.3. Standalone packages: @claude-flow/performance 3.0.1 and @claude-flow/browser 3.0.1.
Claude Code 2.1.290–2.1.318 adoption
- Headless workers: they pass
--permission-mode auto, so they no longer stall in 2.1.302's manual default. They readcredential_expires_atand emit a warning 10 days before a sign-in ends (#4699). - pre-bash gate: fails closed with
onFailure: "block"(#4699). behavior changed · inferred - Leaf agents:
nested-leafandnested-queen-leafrun ateffort: lowwithomitClaudeMd. That measured about 64k fewer tokens per spawn, and a one-leaf run cost $0.081 instead of $0.153 (#4700). - Mods (#4702):
- The guard uses the
tool.checkfieldsisReadOnly,agentIdanddecider: "person". Read-only checks went from about 0.3–0.5 ms to 0.13–0.25 ms. - Hot console tools use
isDeferred. - Text uses
weight, with a fallback for Claude Code before 2.1.299. - Kit tests pass on 2.1.318.
- The guard uses the
agentModelsoption: sets a model per agent type for spawns that chose none (#4703).- Install: the README has the one-command
claude plugin install … --marketplaceform.
Console
- The band above the prompt fits a narrow pane. Rows are measured in terminal columns, parts that don't fit are counted with "+N" instead of being cut mid-word, and the open button sizes to its label (#4701).
Fixes
- AutoMemory (#4685, #4708, thanks @sparkling):
- The JSON store no longer swallows write failures or overwrites an unreadable store with
[]. - Imports verify what was written and exit non-zero on failure.
- New read-only
read <namespace> <key>subcommand.
- The JSON store no longer swallows write failures or overwrites an unreadable store with
- daemon stop (#4654, #4706, thanks @junseokpark-png): an explicit
daemon stopis respected by every auto-start path until the nextdaemon start. - hooks_intelligence_learn (#4693, thanks @sparkling): refuses a
trajectoryIdsselection instead of learning over everything. behavior changed · inferred - agentdb_context-synthesize guidance (#4691, #4705, thanks @sparkling): guidance in five plugins no longer routes research documents or trace spans to it.
- claims (thanks @kevin9327):
claims_handoffremoves the claim from the stealable list (#4694).- Completed claims leave
claims_loadandclaims_rebalance(#4695).
- tasks:
task_createwithassignTomarks the agents busy (#4698, thanks @kevin9327). - performance 3.0.1: the attention benchmark's range check now requires both min and max, and the comparison is de-biased (#4696, thanks @cmf235).
- cost-tracker: each Claude Code response is counted once (#4688, thanks @ibucketbranch).
- hive-mind:
spawn --claude --model(#4672, thanks @ibucketbranch). - browser 3.0.1: agent-browser 0.41.2 (#4662).
Known
- cost-tracker-smoke Case 2 is flaky on main (#4711).
Release PR: #4713