claude-hooks v1.8.0

claude-hooks v1.8.0 ships the /consultants v2 engine in its mature form — multi-role council with opt-in tool_executor + coder lanes, citation-verified researcher output, and a wired-by-default cross-session memory store with TTL + distillation. Plus the May-18 hardening pass that closed four bugs found during the live regression run.

73 commits since v1.7.0 across 13 task IDs (#199–#220). Full diff at v1.7.0...v1.8.0. Test count: 3939 collected / 3909 passed / 30 skipped in both claude-hooks and claude-hooks-consultants conda envs.

Headlines

/consultants v2 engine (mature form)

• Two new opt-in roles, default OFF as of 2026-05-18:
  • tool_executor — researcher PLAN-REPORT split with a dedicated tool-call lane (#103, M11c series). Default flipped back to OFF after the live A/B benchmark showed it loses on grep-shaped questions (benchmarks/consultants/results/2026-05-18/tool-executor-ab/report.md); keep enabled for slow-tool-budget questions where the synthesizer would otherwise re-issue tool calls across nine x-tier lanes.
  • coder — sandboxed write_file role (50 KB/file, 1 MB/lane, 16 files/lane) with per-language model routing (task #111). Caps enforced as tool-result errors so the LLM self-corrects.
• Three-layer CitationLinter at researcher boundary (#204 / #205 / #207): path resolution → line bounds → symbol-match against an mtime-cached code_graph.enclosing_symbol_at query. Annotations propagate via peer_findings; downstream lanes see only verified cites.
• M9 control surface (#214) — runtime_events emission, default sqlite checkpointer, node-enter INFO log; CLI now has state / inject / control / interrupt / resume / cancel / events subcommands.
• M12 parity regression suite (#98) — locks today-default behavior as the baseline; every new opt-in must add a cohort-2 test.

M14 cross-session memory store (default-on)

• LangGraph BaseStore wired by default at effort = high+ via enable_at_efforts. Four namespaces:
  • (sid, "research") — per-lane findings, TTL 30 days
  • (sid, "tool_results") — per-tool outputs, TTL 24 hours
  • ("project", project_id) — per-project durable, never expires
  • ("user", user_id) — user-global durable, never expires
• Caliber-style distillation-on-expiry — hourly daemon reaper summarizes expiring research into the durable ("project", pid) namespace BEFORE deleting originals. Primary gemma4:31b-cloud, fallback glm-5.1:cloud. Critical invariant: research originals only deleted after a successful project-namespace write — failed distillation keeps them for the next sweep.
• #215 reaper pacing — store.ttl.jitter_pct = 0.1, store.distillation.max_groups_per_sweep = 5, store.distillation.pace_seconds_between_distillations = 5.0. Stops a cohort of N sessions with the same expiry from saturating the embedder.
• #220 CLI surface — three new claude-consultants config verbs (set-store, set-store-ttl, set-store-distillation) plus an optional install.py walkthrough for TTL + distillation knobs. The TOML round-trips through the canonical dataclass renderer.

Stop hook / PreCompact

• PreCompact wrap-up (#217) — preserves AskUserQuestion Q&A across the compaction boundary via the new wrapup skill. Q&A pairs are extracted from the live transcript before the harness summarizes.
• stop_guard (#219) — 5 new commitment-stall patterns parsed from real solidPC sessions: "going to X", "now drafting / writing", verb + tail + "now:", "let me insert / patch / apply", "starting with Phase N: ...". All gated by the existing end_turn + zero-tool_use + last-paragraph stack.

May-18 hardening pass

• #212 — ProviderBackedStore._do_put no longer swallows provider failures; the M14 reaper's "delete only after successful distillation write" invariant is now load-bearing.
• #216 — tool_executor write amplification fixed; record_research uses a stable per-lane key (L{lane_idx}), _do_put deletes the prior provider row on overwrite. One row per (namespace, key) regardless of PLAN-REPORT round count.
• #218 — PgvectorProvider read-only methods (expire_before, count, _search_tables, _recall_hybrid_unlocked, _kg_search_nodes) call a new _read_only_finish() at every happy-path exit. Closes the psycopg3 implicit transaction so concurrent ALTER TABLE (M14's lazy ADD COLUMN IF NOT EXISTS expires_at) doesn't block on AccessShareLock.
• #213 — triple-fix for the live cloud regression: KIND_UNKNOWN preservation in process_tool_block, synthesizer source enumeration that exhaustively walks every researcher report instead of stopping at the first non-empty one, prompt compaction.

Upgrade notes

• Defaults flipped 2026-05-18 (commit 48f1c4e): store.enabled = True, backend = "sqlite_vec", sqlite_vec_path = "~/.claude/consultants-store.db", ttl.enabled = True, distillation.enabled = True. M12 parity holds because medium (the default effort) is not in enable_at_efforts — the store stays inactive on the default light tier and only fires at high / max / x-tiers.
• tool_executor default = OFF. Enable per session via claude-consultants config set-role tool_executor --enabled true only when the question profile matches the win condition (slow-tool-budget, multiple x-tier lanes that would otherwise re-issue grep).
• No config/claude-hooks.json migration. v1.7.0 configs are forward-compatible; new [store] knobs default to the values shown above. Operators who want to tune at install time can re-run install.py and answer yes to the "Customize TTL + distillation knobs now?" prompt.
• Authoritative docs: docs/consultants.md (runbook), docs/consultants-roles.md (role reference), docs/consultants-skill-eval-protocol.md + docs/consultants-skill-eval-baselines.md (model-pick evidence).

Full changelog

See CHANGELOG.md (51 entries across the v1.8.0 release window). The "Authoritative docs" links above cover everything user-facing.