v5.3.3
Fixed
-
mergeFullAgentsnow honors lean mode.rulebook updatewas still emitting a 6300-line AGENTS.md becausemergeFullAgents(used by the update command) calledgenerateModularAgentsdirectly without checkingagentsMode. Added an early-return that routes togenerateLeanAgentswhenagentsMode === 'lean'. Verified end-to-end:rulebook updatenow produces a 211-line AGENTS.md. -
Hooks now resolve project root from stdin
cwd, not$(pwd)/Get-Location. The previous behavior caused.rulebook/handoff/to be created inside whatever sub-directory the user was editing instead of the actual project root. Fixed in both.shand.ps1variants ofcheck-context-and-handoffandresume-from-handoff, in.claude/hooks/andtemplates/hooks/. Falls back toCLAUDE_PROJECT_DIRthenpwdif stdin doesn't includecwd. -
Lean AGENTS.md is now unconditional.
generateFullAgentsandmergeFullAgentsalways emit the condensed ~211-line template regardless ofagentsModeormodularconfig values. Users withagentsMode: 'full'from older versions will get the lean output on nextrulebook update. Dead legacy codepath (procedural 6k-line generator) removed. -
rulebook updateno longer regenerates a 6k-line bloated AGENTS.md. Lean mode is now the default foragentsMode(config-manager.ts), andupdate.tstreats existing configs withoutagentsModeas lean. Existing projects upgrading to 5.3.2 will get the condensed 224-line AGENTS.md on nextrulebook update. -
templates/core/AGENTS_LEAN.mdreplaced with the full 224-line condensed AGENTS.md (was a 25-line index that didn't match the rootAGENTS.md). Includes Tier 1 prohibitions (rule #8: full-task-no-questions), Critical Rules, Task Management, Agent Automation, Memory, Knowledge Base, DAG, Quality Enforcement, Token Optimization, Ralph, Multi-Agent Teams, Workspace, Agent Delegation, Plans, Decision Records, withLANGUAGE_REFS/MODULE_REFSplaceholders.
Changed
MAX_CONTEXT_CHARSdoubled from 800k → 1.6M (~200k → ~400k tokens) in.claude/hooks/check-context-and-handoff.shandtemplates/hooks/check-context-and-handoff.sh. Handoff thresholds now match the larger context windows of current models.