v0.6 — Harness Modernization

Harness modernization for current Claude Code (CLI 2.1.x). All 6 batches implemented and validated. Plan: ~/.claude/plans/claude_extensions/harness-modernization-20260527.md.

Existing installs: re-run install.sh to pick up the hook-wiring change (re-merges settings.json — moves auto-capture to SessionEnd, drops the removed ref-scorer hook — and cleans up symlinks from removed components).

Added

• session-init.py re-injects persisted context on session start/resume and after native compaction (source: "compact") — emits a recency-ordered summary of recent session-progress.md entries via additionalContext (stdout JSON, 4 KB cap, source-gated; diagnostics stay on stderr). Closes the write-only loop so persisted status returns to the model automatically instead of only being written to disk. No UserPromptSubmit fallback needed (Batch 1 validated that SessionStart re-fires on compaction).

Removed

• Context-compression subsystem — deleted hooks/lib/compressor.py (4-tier Active→Compressed→Archived→Dropped rotation), the /compress skill, the ref-scorer.py PostToolUse hook + ref_scorer_mod.py, and hooks/lib/ref_tracker.py. An audit found the compressor had never fired in 29 sessions (no last_compression key, no archive.md, no ## Compressed Context section) and its tiers were write-only; native context compaction now summarizes the live conversation. Removed the ref-scorer entry from settings.json.reference and ~125 associated tests.
• session-init.py no longer computes active/stale counts or triggers compression — it increments session_count and logs a plain entry count.
• auto-capture.py no longer writes scores to ref-cache.json (dead data with no remaining consumer).

Changed

• README.md statusline example model label Sonnet 4.6 → Opus 4.7; README/DEVELOPER updated to drop the removed subsystem.
• auto-capture moved from the Stop hook to SessionEnd — runs once at session teardown instead of after every response. Refactored into a reusable capture(workdir) that resolves the project dir from CLAUDE_PROJECT_DIR (fallback cwd) and chdirs before running git, fixing a non-project-CWD silent no-op; captures on all SessionEnd reasons (reason is logging-only, parsed best-effort so malformed/empty stdin never suppresses a capture).
• Per-subagent model selection in the multi-agent skills — /code-review, /security-audit, /critical-review, and /implement-batch now spawn each subagent with an explicit model: opus for the higher-stakes passes (architecture, correctness, injection/auth/crypto, implementers) and sonnet for the rest. Review/audit agents never default to haiku. CLAUDE_CODE_SUBAGENT_MODEL overrides these. Documented the mapping in DEVELOPER.md.
• Shared review output-contract fragments — extracted the boilerplate that is identical verbatim between the /code-review and /security-audit subagent prompts (scope/context opening, findings-list-format line, closing instruction) into rules/review-output-contract.md; each skill inlines the fragments at <<shared:…>> markers. Audit found the field list, severity scale, rationalizations, and red flags are intentionally domain-specific (not duplicated), so they stay per-skill; /critical-review reviews plans and does not consume the fragments. Added tests/test_review_contract.py.

Fixed

• tests/test_fileutil.py — moved a module-level assignment below the deferred import to clear a long-standing ruff E402.