v0.48.0

Hook-overhead minimization. Surfaced by greenfield-api's self-assessment: every Edit, every UserPromptSubmit, every Read-after-touched-file pays a token tax from the broader memory/preflight integration. Deep audit identified five hook-message and gating improvements that preserve quality protection while cutting per-fire token cost. No functionality removed; every hook keeps its quality role, the messaging just becomes right-sized. Estimated net saving: ~4,500 tokens per active dev session. Smoke: 509 passed, 0 failed (+5 new gates, +1 updated gate to match new compact format).

Changed

• hooks/pre-flight-guard.sh — two improvements. (1) Exit silently when workflow.yaml exists but active=false. Prior implementation only checked file existence, so completed workflows left the guard firing on every Edit indefinitely (state.cjs never deletes workflow.yaml on completion — it sets active=false). (2) Compacted the deny/warn message from ~150 tokens to ~35 tokens. The compact message preserves the load-bearing recovery cue (literal PREFLIGHT <ts> edit <path> :: <ADR-ids|ungoverned> format hint + ungoverned escape keyword) so agents without the memory-pre-flight skill loaded — e.g., raw-dispatched agents that bypassed the workflow — can still recover from the message alone. Verbose protocol re-explanation removed; agents with the skill loaded already know the protocol from their system prompt.
• hooks/workflow-context-injector.sh — two improvements. (1) Compacted the active-workflow context line: [devt] STANDARD · implement (iter 2) [autonomous, tdd] · "task..." (~80 tokens) → [devt] STANDARD/implement·i2·auto+tdd · "task..." (~50 tokens). The line is human-facing only (no programmatic consumer — validated via codebase grep), so compactness wins without breaking contracts. (2) Removed the idle-state context line. Prior implementation emitted [devt] idle · last: <phase> · "<task>" on every UserPromptSubmit when active=false but phase was set — pinning idle context into every prompt long after workflow completion. /devt:status and /devt:next provide explicit resume paths; the implicit per-prompt reminder is pure overhead.
• hooks/read-before-edit-guard.sh — compacted the reminder from ~80 tokens to ~25 tokens. The hook duplicates the CC harness's own read-before-edit enforcement; the advisory is a preemptive reminder, not a guard. Compact form: Reminder: if "X" has not been Read in this session, Read it first — the runtime requires it before Edit. Verbose form removed.

Added

• Five new smoke gates under == Hook-overhead minimization ==: read-before-edit compact reminder (byte budget), workflow-context-injector compact active line format, workflow-context-injector silent on idle, pre-flight-guard silent on active=false, pre-flight-guard compact warn message with format hint + ungoverned escape preserved.

Context

• What's NOT in this release (intentionally deferred):
  • Directory-level PREFLIGHT coverage (would let one PREFLIGHT line cover edits to multiple files under a directory). Investigated, but skills/memory-pre-flight/SKILL.md documents the protocol as PER-FILE — loosening it would be a spec change, not just an optimization. Requires a /devt:council decision before encoding.
  • PostToolUse:Read recent-Read tracking (would eliminate read-before-edit false-positives when agent just-Read the file). Race-condition risk with the existing async PostToolUse hooks; the simpler compact-message fix in this release captures most of the win without the new hook.
  • Memory-auto-index incremental updates (full FTS5 rebuild replaced with single-doc index). Already debounced at 5s; the marginal complexity isn't worth the wall-time saving for this release.
• Where the actual token cost goes (full attribution, audited from greenfield-api field data): claude-mem's PreToolUse:Read injection (~300-700 tokens per Read on touched files) is the dominant per-edit cost and not under devt's control. devt's per-edit contribution after this release is ~25 tokens (read-before-edit reminder) + 0-35 tokens (compact preflight warn when triggered). The "Security Guidance" UserPromptSubmit injection (~470 tokens/turn observed in field) comes from an external/global hook, not devt.