kairos-chain 3.30.0 → 3.31.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 3ec94fadb095a49a89c0235010c80e37913aaf6781aa16d22a2e5855b1eb154d
-  data.tar.gz: 38c3b631326b87f07285f9a5f4ecdd64fc22d086b6b08de76d5b2aaf2d7d7040
+  metadata.gz: 227d024f36839c595295ed0e3b4415c6764750597283b59a21ea1f5e16112210
+  data.tar.gz: 8fd8767580dbe3db617cbaf4ee0dcfafdb38f5823315f0321bae0c4893b2f45d
 SHA512:
-  metadata.gz: 7d3d7e9d5f52b58109a795ce4f67243551ad353d18d273d4062034e09bcafe8201559ceabcef16c5ad0d0fc0bca4b8b8629c27cb49e224091ae05ffd5aa364ac
-  data.tar.gz: a20e8e7a2cb18d8cbcbd76f39c7fa69a9bbd55d4447c16fa289ca8227fd2537325693748e711e46e57693c6572b41f15096dc358368f89911d95eaea3c714ca8
+  metadata.gz: 7486fef959f54577c0a13a301112f746f8c440100f11f5d424f27771759ad3024d6d30988b7a9e7c58a4544db1ffe6c2705b3271226d9445f1537dc591f96ef7
+  data.tar.gz: 8d670bb2a2f5d8d848072101527fcb5b3153c23ea4c5ae1359c1d46c30187011911b32512086e1e1734e72ba5d0d3ae20913499d5d1d8c4b4d1d359b79963281

data/CHANGELOG.md

@@ -4,6 +4,41 @@ All notable changes to the `kairos-chain` gem will be documented in this file.
 
 This project follows [Semantic Versioning](https://semver.org/).
 
+## [3.31.0] - 2026-06-11
+
+### Changed — multi_llm_review roster: Fable 5 + Opus 4.6/4.8 (6 reviewers)
+
+Default reviewer roster updated for the Fable 5 / Opus 4.8 model generation:
+
+- Orchestrator/team slot: `claude-fable-5` (`claude_team_fable5`) replaces
+  Opus 4.7. Opus 4.8 added as a second subprocess CLI reviewer
+  (`claude_cli_opus4.8`) alongside Opus 4.6, which is retained for its
+  documented complementary bias (ambiguity-preserving, self-reference-friendly).
+  4.7 retired: its register is covered by 4.8 and Fable 5.
+- Convergence rules: `4/6 APPROVE` full roster, `3/5 APPROVE` after
+  orchestrator exclusion ("exclude" strategy only — "subprocess" keeps the
+  full roster; "delegate" re-adds the slot at collect, so 4/6 governs there).
+- `timeout_seconds` raised 300 → 600 (live 6-roster run measured 381s
+  wall-clock with `max_concurrent: 2`).
+- Validated by a 2-round self-referential review of the workflow L1 with the
+  new roster itself (R1 REVISE → fixes → R2 with 4.6/4.8/codex-5.4 APPROVE,
+  including a code-grounded Cursor correction of the exclusion semantics).
+
+### Changed — `multi_llm_review_workflow` L1 v3.5
+
+- All roster-dependent sections updated (pre-flight checklist, CLI tool
+  matrix, convergence rules, orchestrator self-identification, orchestration
+  template, LLM identifiers). Claude CLI 4.6/4.8 rows verified live 2026-06-10.
+- Effort escalation paragraph scoped to coding/design sub-agents and the
+  revision phase (reviewers stay at default per the 2026-04-29 policy).
+
+### Fixed — `knowledge_update` size guidance
+
+- Removed the "~2 KB MCP stdio limit" warning from the tool description and
+  nil-content diagnostic: two ~40 KB updates succeeded over stdio on
+  2026-06-10/11. The old figure generalized a single unreproduced
+  nil-content incident; the nil-content detection itself is retained.
+
 ## [3.30.0] - 2026-06-08
 
 ### Added — `dream_digest`: derived narrative view over L2/L1 fragments (dream SkillSet v0.3.0)

data/lib/kairos_mcp/drift_detection/correspondence_checker.rb

@@ -0,0 +1,136 @@
+# frozen_string_literal: true
+
+require 'digest'
+require 'json'
+require_relative '../kairos_chain/chain'
+
+module KairosMcp
+  module DriftDetection
+    # CorrespondenceChecker — INV-A detection floor (Cycle 1, toward by-construction).
+    #
+    # Checks whether a live L0/L1 artifact still corresponds to its *current
+    # recorded provenance*: the content digest stored for that artifact at the
+    # head of the constitutive record (the hash chain). This is detection only —
+    # it surfaces divergence; it does not prevent edits or gate writes. Those are
+    # later cycles (single-source enforcement, record-as-gate).
+    #
+    # Provenance is rooted in the hash chain, not in the SQLite knowledge_meta
+    # cache: INV-A names the chain head as the non-editable anchor. The chain is
+    # therefore the single source consulted here; the meta table (when present)
+    # is a derived view and is intentionally not used for the comparison.
+    #
+    # The digest is computed over the *raw file content* (frontmatter included),
+    # matching exactly how it was recorded on create/update (a verbatim write,
+    # no normalization). Comparing the parsed/stripped body would never match.
+    class CorrespondenceChecker
+      # Result of a single correspondence check.
+      #
+      # status:
+      #   :match            live artifact corresponds to recorded provenance
+      #   :mismatch         live content diverged from recorded provenance (silent edit)
+      #   :missing_record   live artifact relied upon, but no recorded provenance exists
+      #   :missing_artifact recorded/expected artifact is absent at the reliance point
+      #   :error            the check itself could not complete (not a correspondence claim)
+      Result = Struct.new(
+        :status, :name, :active_digest, :recorded_digest, :message,
+        keyword_init: true
+      ) do
+        def corresponds?
+          status == :match
+        end
+
+        # A surfaced non-correspondence per INV-A (divergence, not an internal error).
+        def divergence?
+          %i[mismatch missing_record missing_artifact].include?(status)
+        end
+      end
+
+      class << self
+        # Check an L1 knowledge artifact against its recorded provenance.
+        #
+        # @param name [String] knowledge id (knowledge_id on the chain record)
+        # @param md_file_path [String, nil] path to the live .md file relied upon
+        # @param storage_backend [Storage::Backend, nil] backend for chain access
+        # @return [Result]
+        def check_l1(name:, md_file_path:, storage_backend: nil)
+          unless md_file_path && File.file?(md_file_path)
+            # Relied upon but absent — a missing artifact is itself a
+            # non-correspondence under INV-A (the expected set is recorded).
+            return Result.new(
+              status: :missing_artifact, name: name,
+              active_digest: nil, recorded_digest: nil,
+              message: "L1 '#{name}': artifact missing at the point of reliance"
+            )
+          end
+
+          active = Digest::SHA256.hexdigest(File.read(md_file_path))
+          recorded = recorded_digest_for(name, storage_backend)
+
+          if recorded.nil?
+            return Result.new(
+              status: :missing_record, name: name,
+              active_digest: active, recorded_digest: nil,
+              message: "L1 '#{name}': live artifact has no recorded provenance on the chain"
+            )
+          end
+
+          if active == recorded
+            Result.new(
+              status: :match, name: name,
+              active_digest: active, recorded_digest: recorded, message: nil
+            )
+          else
+            Result.new(
+              status: :mismatch, name: name,
+              active_digest: active, recorded_digest: recorded,
+              message: "L1 '#{name}': live content diverged from recorded provenance " \
+                       "(active #{short(active)} ≠ recorded #{short(recorded)})"
+            )
+          end
+        rescue StandardError => e
+          Result.new(
+            status: :error, name: name,
+            active_digest: nil, recorded_digest: nil,
+            message: "L1 '#{name}': correspondence check could not complete: #{e.message}"
+          )
+        end
+
+        private
+
+        # The current recorded content digest for a knowledge_id: the next_hash of
+        # the most recent knowledge_update record, scanning the chain from head
+        # backward. Returns nil when the most recent relevant record removed the
+        # artifact (next_hash nil — delete/archive) or when none exists.
+        def recorded_digest_for(name, storage_backend)
+          chain = KairosChain::Chain.new(storage_backend: storage_backend)
+          chain.chain.reverse_each do |block|
+            Array(block.data).each do |entry|
+              record = parse_entry(entry)
+              next unless record.is_a?(Hash)
+              next unless record['type'] == 'knowledge_update'
+              next unless record['knowledge_id'] == name
+
+              # First match from the head is the current provenance (may be nil
+              # if the artifact was removed — caller treats nil as no record).
+              return record['next_hash']
+            end
+          end
+          nil
+        end
+
+        def parse_entry(entry)
+          return entry if entry.is_a?(Hash)
+          return JSON.parse(entry) if entry.is_a?(String)
+
+          nil
+        rescue JSON::ParserError
+          nil
+        end
+
+        def short(digest)
+          digest ? digest[0, 12] : '-'
+        end
+      end
+    end
+  end
+end

data/lib/kairos_mcp/knowledge_provider.rb

@@ -23,6 +23,11 @@ module KairosMcp
   # - Blockchain: Uses the configured storage backend
   #
   class KnowledgeProvider
+    # Main knowledge directory (constitutively-recorded L1). Exposed so callers
+    # can distinguish main-dir knowledge from read-only external SkillSet
+    # knowledge, e.g. to scope INV-A correspondence checks to recorded artifacts.
+    attr_reader :knowledge_dir
+
     ARCHIVED_DIR = '.archived'
     ARCHIVE_META_FILE = '.archive_meta.yml'
     # Backup directories created by upgrade flow (`.bak.<timestamp>`).

data/lib/kairos_mcp/tools/knowledge_get.rb

@@ -2,6 +2,7 @@
 
 require_relative 'base_tool'
 require_relative '../knowledge_provider'
+require_relative '../drift_detection/correspondence_checker'
 
 module KairosMcp
   module Tools
@@ -77,11 +78,59 @@ module KairosMcp
         end
 
         output = build_output(skill, arguments, provider)
-        text_content(output)
+        # INV-A detection floor: reading L1 knowledge "in order to act upon" is a
+        # point of reliance. Surface any divergence from the recorded provenance
+        # here — never silently. Scoped to main-dir L1 (external SkillSet
+        # knowledge has no chain provenance and would false-positive).
+        banner = correspondence_banner(skill, provider)
+        text_content(banner ? banner + output : output)
       end
 
       private
 
+      # Returns a surfacing prefix if the live artifact does not correspond to its
+      # recorded provenance, or nil when it corresponds (stay silent on match).
+      #
+      # Surfacing is graded by signal strength (the invariant requires only
+      # non-silence; grading is a Cycle-1 backlog policy):
+      #   :mismatch       a recorded artifact whose content silently changed —
+      #                   high signal, rare → an alarm banner.
+      #   :missing_record a live artifact with no chain provenance — overwhelmingly
+      #                   template-provisioned knowledge whose provenance root is the
+      #                   gem/template, not a per-instance record. Chain-rooting the
+      #                   expected set is explicit Cycle-1 backlog, so this is a muted
+      #                   one-line note, not an alarm — bannering every bundled read
+      #                   would train the reader to ignore the banner.
+      def correspondence_banner(skill, provider)
+        return nil unless main_dir_l1?(skill, provider)
+
+        result = DriftDetection::CorrespondenceChecker.check_l1(
+          name: skill.name,
+          md_file_path: skill.md_file_path
+        )
+
+        case result.status
+        when :mismatch, :missing_artifact
+          "> ⚠️ **Drift detected (INV-A)** — #{result.message}.\n" \
+            "> This content was modified outside the recorded change path; treat it as unverified.\n\n"
+        when :missing_record
+          "> ℹ️ No recorded provenance for this entry (provisioning not yet chain-rooted — Cycle-1 backlog).\n\n"
+        end
+      rescue StandardError => e
+        # A failed check must not break the read; report it without claiming correspondence.
+        warn "[knowledge_get] correspondence check failed: #{e.message}"
+        nil
+      end
+
+      # True only for knowledge living under the provider's main knowledge dir —
+      # i.e. constitutively-recorded L1, not read-only external SkillSet knowledge.
+      def main_dir_l1?(skill, provider)
+        return false unless skill.md_file_path && provider.respond_to?(:knowledge_dir)
+
+        root = File.expand_path(provider.knowledge_dir) + File::SEPARATOR
+        File.expand_path(skill.md_file_path).start_with?(root)
+      end
+
       def build_output(skill, arguments, provider)
         output = "## [#{skill.name}] #{skill.description || 'No description'}\n\n"
         output += "**Layer:** L1 (Knowledge)\n"

data/lib/kairos_mcp/tools/knowledge_update.rb

@@ -13,8 +13,9 @@ module KairosMcp
 
       def description
         'Create, update, or delete L1 knowledge skills. Changes are recorded with hash references to the blockchain. ' \
-        'NOTE: MCP stdio transport may silently drop large arguments. Keep combined content + reason under ~2 KB. ' \
-        'For larger L1 entries, trim prose or split into multiple entries.'
+        'Large content is supported (a ~40 KB update was verified over stdio on 2026-06-10; the earlier ~2 KB guidance ' \
+        'generalized a single unreproduced nil-content incident). If content ever arrives as nil, the error message ' \
+        'below explains recovery; no preemptive size limit applies.'
       end
 
       def category
@@ -128,9 +129,10 @@ module KairosMcp
       def content_missing_error(command, content)
         if content.nil?
           "Error: content is required for #{command}. " \
-          "The content argument arrived as nil — this often means the MCP transport silently dropped it " \
-          "because the combined argument size exceeded the client limit (~2 KB for stdio). " \
-          "Try trimming content and reason, or split into smaller entries."
+          "The content argument arrived as nil — the MCP transport or the calling LLM dropped it. " \
+          "(A ~2 KB stdio limit was once suspected, but a ~40 KB update succeeded on 2026-06-10, " \
+          "so size alone is unlikely to be the cause.) " \
+          "Retry the call; if nil persists, write the content to a file and report the incident."
         else
           "Error: content is required for #{command} (received empty string)"
         end

data/lib/kairos_mcp/version.rb

@@ -1,4 +1,4 @@
 module KairosMcp
-  VERSION = "3.30.0"
+  VERSION = "3.31.0"
   CHANGELOG_URL = "https://github.com/masaomi/KairosChain_2026/blob/main/CHANGELOG.md"
 end

data/templates/knowledge/llm_cross_evaluation/scripts/run_cross_eval.rb

@@ -45,6 +45,14 @@ MODELS = {
     input_mode: :stdin,
     thinking_effort: "medium",
   },
+  "claude_fable5" => {
+    tool: :claude,
+    cmd: "claude --print --model claude-fable-5 --effort medium",
+    label: "Claude Fable 5",
+    provider: "anthropic",
+    input_mode: :stdin,
+    thinking_effort: "medium",
+  },
   "claude_opus46" => {
     tool: :claude,
     cmd: "claude --print --model claude-opus-4-6 --effort medium",

data/templates/knowledge/multi_llm_review_workflow/multi_llm_review_workflow.md

@@ -1,7 +1,7 @@
 ---
 name: multi_llm_review_workflow
 description: "Multi-LLM review methodology and execution — workflow pattern, CLI tooling, consensus analysis, Persona Assembly. Applicable to design, implementation, documentation, or any artifact."
-version: "3.4"
+version: "3.5"
 tags:
   - workflow
   - review
@@ -192,11 +192,11 @@ starting** and verify each against `config/multi_llm_review.yml`:
 ```
 - [ ] Your model (orchestrator): ___
 - [ ] Agent Team Personas model: = orchestrator model (NOT a different model)
-- [ ] Subprocess CLI model: opposite Opus (4.6 if you are 4.7, vice versa)
+- [ ] Subprocess CLI models: Opus 4.6 AND Opus 4.8 (both, not either/or)
 - [ ] Codex models: gpt-5.5 (default) AND gpt-5.4 (both, not either/or)
 - [ ] Cursor model: default (composer-2.5, no --model flag)
-- [ ] Total reviewer count: 5 (or 4 after orchestrator exclusion from subprocess)
-- [ ] Convergence rule: 3/5 APPROVE (full) or 3/4 APPROVE (after exclusion)
+- [ ] Total reviewer count: 6 (or 5 after orchestrator exclusion from subprocess)
+- [ ] Convergence rule: 4/6 APPROVE (full) or 3/5 APPROVE (after exclusion)
 ```
 
 ### Common mistakes (Path A)
@@ -206,7 +206,7 @@ starting** and verify each against `config/multi_llm_review.yml`:
 | Exclude orchestrator model from Agent Team Personas | Agent Team uses orchestrator model — they provide persona diversity, not epistemic diversity | LLM misreads "do not assign yourself as a reviewer" as applying to Agent Team; it applies only to subprocess CLI |
 | Run only Codex GPT-5.4, skip 5.5 | Run both — they catch different things (5.5 found §5 schema contradiction in Phase 2 Case A that no other reviewer caught) | Cost-saving heuristic; roster has both for a reason |
 | Use a smaller/cheaper model as Agent Team substitute | Use the orchestrator's own model with different personas | Confusing "model diversity" with "persona diversity" — Agent Team is the latter |
-| Run 3 reviewers instead of 5 (or 4 after exclusion) | Use the full roster from config | Ad-hoc "3 is enough" reasoning; config specifies 5 for empirical reasons |
+| Run 3 reviewers instead of 6 (or 5 after exclusion) | Use the full roster from config | Ad-hoc "3 is enough" reasoning; config specifies 6 for empirical reasons |
 
 ## Roles
 
@@ -304,10 +304,10 @@ The rule applies **after** orchestrator classifies each finding as (a)/(b)/(c) p
 findings count toward the thresholds below; (c) findings are recorded as advisory
 and never block.
 
-- **3/4 APPROVE** (no (a)/(b) REJECT) = proceed to next step
+- **4/6 APPROVE** full roster, or **3/5 APPROVE** after orchestrator exclusion ("exclude" strategy only — the default "delegate" strategy keeps 6 voters via collect) (no (a)/(b) REJECT) = proceed to next step
 - **Any (a) or (b) REJECT or FAIL** = revise and re-review
 - **(c)-only REJECT** = record as advisory, non-blocking
-- **4/4 APPROVE** (no (a)/(b)) = highest confidence, proceed
+- **Unanimous APPROVE** (no (a)/(b)) = highest confidence, proceed
 - Legacy 3-reviewer mode: 2/3 APPROVE (no (a)/(b)) = proceed
 - Codex REJECT with (a)/(b) findings + others APPROVE = likely real issue, investigate before overriding
 - Codex REJECT with only (c) findings = expected per Codex value-system divergence; non-blocking
@@ -319,11 +319,11 @@ For normative detail and the underlying classification, see
 
 | Agreement | Meaning | Action |
 |-----------|---------|--------|
-| **4/4** (or **3/3**) | Architectural-level gap | Must fix |
-| **3/4** (or **2/3**) | Implementation-level issue | Should fix |
-| **1/4 only** | Specialty-specific insight | Do NOT ignore — often the most novel finding |
+| **N/N** (unanimous) | Architectural-level gap | Must fix |
+| **Majority** (e.g. 4/6, 3/5) | Implementation-level issue | Should fix |
+| **1/N only** | Specialty-specific insight | Do NOT ignore — often the most novel finding |
 
-1/4 (or 1/N) findings are not "minority opinions to discard." They represent unique expertise.
+1/N findings are not "minority opinions to discard." They represent unique expertise.
 
 ### Majority Rule — Reference Only
 
@@ -401,19 +401,20 @@ which agent 2>/dev/null && echo "agent: available" || echo "agent: NOT FOUND"
 which claude 2>/dev/null && echo "claude: available" || echo "claude: NOT FOUND"
 ```
 
-- All three available → Auto mode (4 reviewers, default)
-- Codex + Agent only → Auto mode (3 reviewers, legacy)
+- All three available → Auto mode (6 reviewers, default)
+- Codex + Agent only → Auto mode (legacy, reduced roster — apply "Legacy 3-reviewer mode: 2/3 APPROVE" from Convergence Rules)
 - Any of codex/agent missing → Manual mode
 - User override: `mode: manual` or `mode: auto`
 
-### CLI Tool Matrix (Tested 2026-03-28)
+### CLI Tool Matrix (tested 2026-03-28; Claude CLI 4.6/4.8 rows verified live 2026-06-10)
 
 | Tool | Command | Prompt Input | Output Collection | Model |
 |------|---------|-------------|-------------------|-------|
-| **Codex** | `codex exec` | stdin pipe: `cat prompt.md \| codex exec -` | `-o /path/output.md` | GPT-5.5 (default) |
+| **Codex** | `codex exec -m <model>` | stdin pipe: `cat prompt.md \| codex exec -` | `-o /path/output.md` | GPT-5.5 + GPT-5.4 (both roster entries, `-m` per entry) |
 | **Cursor Agent** | `agent -p` | File reference (stdin NOT supported) | stdout redirect: `> output.md` | Composer-2.5 (default) |
-| **Claude Code** | Agent tool (internal) | Direct prompt string | Write to workspace file | Opus 4.6 (session) |
-| **Claude CLI (4.7)** | `claude -p --model claude-opus-4-7 --bare` | stdin pipe: `cat prompt.md \| claude -p --model claude-opus-4-7 --bare` | stdout redirect: `> output.md` | Opus 4.7 |
+| **Claude Code** | Agent tool (internal) | Direct prompt string | Write to workspace file | Fable 5 (session) |
+| **Claude CLI (4.6)** | `claude -p --model claude-opus-4-6 --bare` | stdin pipe: `cat prompt.md \| claude -p --model claude-opus-4-6 --bare` | stdout redirect: `> output.md` | Opus 4.6 |
+| **Claude CLI (4.8)** | `claude -p --model claude-opus-4-8 --bare` | stdin pipe: `cat prompt.md \| claude -p --model claude-opus-4-8 --bare` | stdout redirect: `> output.md` | Opus 4.8 |
 
 ### Thinking Effort Configuration (validated 2026-04-20)
 
@@ -421,23 +422,30 @@ Based on cross-evaluation experiment (7 models × 4 tasks + Nomic, 518 CLI calls
 
 | Role | Model | Effort Flag | Rationale |
 |------|-------|-------------|-----------|
-| **Primary (orchestrator)** | Opus 4.6 | `--effort medium` | Sufficient for integration, dialogue, judgment |
-| **Reviewer: Agent Team** | Opus 4.6 | `--effort medium` | Evaluator quality adequate at medium |
-| **Reviewer: Claude CLI** | Opus 4.7 | `--effort low` | Evaluator quality is effort-independent (low≈high: 8.35 vs 8.16) |
+| **Primary (orchestrator)** | Fable 5 (session default) | (default) | Sufficient for integration, dialogue, judgment |
+| **Reviewer: Agent Team** | = orchestrator (Fable 5) | (default) | Personas inherit orchestrator model |
+| **Reviewer: Claude CLI** | Opus 4.6 / Opus 4.8 | (default; config `effort: medium`) | Evaluator quality is effort-independent (low≈high: 8.35 vs 8.16) — per 2026-04-29 policy reviewers stay at default |
 | **Coding sub-agent** | Opus 4.7 | `--effort medium` | Cost-effective default; use `high` for complex tasks |
 | **Design sub-agent** | Opus 4.7 | `--effort medium` | Cost-effective default; use `high` for complex tasks |
 | **Codex** | GPT-5.5 (default) | (no flag) | Fixed effort |
 | **Cursor Agent** | Composer-2.5 | (no flag) | Fixed effort |
 
+Note (2026-06-10): the effort experiment data is from the Opus 4.6/4.7
+generation. Fable 5 and Opus 4.8 effort sensitivity is not yet calibrated;
+defaults apply until re-measured.
+
 Key findings:
 - **Opus 4.6** high effort improves Evaluator/Strategy (+0.43/+0.200 Nomic), not Response
 - **Opus 4.7** high effort improves Response/Thinking (+0.81 code, +0.53 philosophy), not Evaluator
 - **Opus 4.7 low > Opus 4.6 high** in combined score — model generation > effort setting
 
-**Effort escalation**: For particularly complex tasks (Tier 3+ architecture, security-critical
-code, multi-component refactoring), the LLM accessing this skill SHOULD escalate to `--effort high`
-at its own judgment. No human approval is needed for effort escalation — it is a cost/quality
-tradeoff that the executing LLM is best positioned to evaluate in context.
+**Effort escalation** (coding/design sub-agents and the post-aggregation revision
+phase only — NOT reviewers, who stay at default per the 2026-04-29 policy): For
+particularly complex tasks (Tier 3+ architecture, security-critical code,
+multi-component refactoring), the LLM accessing this skill SHOULD escalate to
+`--effort high` at its own judgment. No human approval is needed for effort
+escalation — it is a cost/quality tradeoff that the executing LLM is best
+positioned to evaluate in context.
 
 ### Model Detection
 
@@ -454,15 +462,17 @@ agent --list-models 2>&1 | grep "(current\|default)"
 **Rule**: When invoking `multi_llm_review` (or running this workflow manually), the
 orchestrating LLM MUST pass its own model identifier as `orchestrator_model`.
 
-**Rationale**: The reviewer roster typically contains both Opus 4.6 and Opus 4.7
-entries. To avoid the orchestrator reviewing its own output (no independent signal),
-the dispatcher excludes any roster entry whose `model` matches `orchestrator_model`.
-This keeps the same SkillSet useful regardless of which Opus version the user has
-toggled to via `/model` — review composition adapts automatically.
+**Rationale**: The reviewer roster contains multiple Claude entries (Fable 5
+team slot, Opus 4.6 CLI, Opus 4.8 CLI). To avoid the orchestrator reviewing its
+own output (no independent signal), the dispatcher excludes or delegates the
+roster entry whose `model` matches `orchestrator_model` (per
+`orchestrator_strategy`). This keeps the same SkillSet useful
+regardless of which Claude model the user has toggled to via `/model` — review
+composition adapts automatically.
 
 **Why "argument-passing" not "file-introspection"**:
 - The orchestrator's model identity lives in *its own context* (system prompt
-  declares e.g. "You are powered by Opus 4.7"). No external file or env var is
+  declares e.g. "You are powered by Fable 5"). No external file or env var is
   authoritative — `/model` switches change context immediately.
 - MCP protocol does not transmit caller-model info; only the orchestrator can
   truthfully report its own identity. This is genuine self-reference: the system
@@ -473,7 +483,8 @@ toggled to via `/model` — review composition adapts automatically.
 
 **How orchestrator obtains its model ID**:
 - Claude Code sessions: read the system prompt line "You are powered by the
-  model named ... The exact model ID is `claude-opus-X-Y`". Use the exact ID.
+  model named ... The exact model ID is ...". Use the exact ID as stated,
+  whatever its form (e.g. `claude-fable-5`, `claude-opus-4-8`).
 - Other hosts: use whatever introspection the host provides; if none, pass
   `null` and accept that no exclusion happens.
 
@@ -482,27 +493,33 @@ toggled to via `/model` — review composition adapts automatically.
 multi_llm_review(
   artifact_path: "log/design.md",
   review_type: "design",
-  orchestrator_model: "claude-opus-4-7"   # MUST be set by caller
+  orchestrator_model: "claude-fable-5"    # MUST be set by caller
 )
 ```
 
 **Dispatcher behavior** (config: `exclude_orchestrator_model: true`, default `true`):
 - If `orchestrator_model` matches a roster entry's `model`, that entry is skipped.
 - `min_quorum` and `convergence_rule` apply to the remaining reviewers.
-- 4-reviewer roster → 3 reviewer; recommended `convergence_rule: "2/3 APPROVE"`
-  when one Opus is excluded.
+- 6-reviewer roster → 5 reviewers; `convergence_rule_after_exclusion: "3/5 APPROVE"`
+  (from config) replaces the full-roster rule. This reduced count applies to the
+  "exclude" strategy only. The "subprocess" strategy keeps the full roster (the
+  matching entry runs as a fresh CLI process instead of being skipped). Under the
+  default "delegate" strategy, the matching entry is dropped at dispatch but
+  re-added at collect as the persona-team entry, so the voter count returns to 6
+  and the full-roster rule (4/6 APPROVE) applies — verified live 2026-06-10.
 - If `orchestrator_model` is `null` or unmatched, full roster runs (back-compat).
 
 **Manual-mode equivalent**: When orchestrating by hand, do not assign yourself
-as a reviewer. Pick the *other* Opus version for the Claude CLI subprocess
-reviewer (4.6 if you are 4.7, and vice versa).
+as a subprocess reviewer. Run the Claude CLI subprocess reviewers (Opus 4.6 and
+Opus 4.8); if your own model matches one of them, skip that entry and use the
+after-exclusion convergence rule.
 
-### Orchestrator Delegation Protocol (Two-Phase, opt-in)
+### Orchestrator Delegation Protocol (Two-Phase, default)
 
 The `delegate` strategy lets the orchestrator perform persona-based "Agent Team"
 review in its own context — preserving inherited project context that a fresh
-`claude -p` subprocess loses. Subprocess reviewers (codex, cursor, opposite-Opus)
-remain single-LLM.
+`claude -p` subprocess loses. Subprocess reviewers (codex, cursor, Claude CLI
+Opus 4.6/4.8) remain single-LLM.
 
 **Why**: The orchestrator already holds the artifact in context with full project
 awareness. Re-shipping it to a sandboxed subprocess discards that context. Same-
@@ -537,8 +554,10 @@ cross-model subprocess reviewers give epistemic diversity. The two are complemen
   required fields. Fix and retry collect with the same token.
 - All-subprocess-failed at Call 1: returns error immediately; no token issued.
 
-**Default**: `orchestrator_strategy` defaults to `"exclude"` (back-compat). Use
-`"delegate"` explicitly until validated by use.
+**Default**: `orchestrator_strategy` defaults to `"delegate"` (config key
+`default_orchestrator_strategy`). `"exclude"` remains available as the legacy
+strategy. (Historical note: delegate was opt-in until validated by use; it has
+been the config default since v3.x.)
 
 #### Async/Parallel Collect Timing — Iron Rule
 
@@ -603,8 +622,8 @@ readable until GC. Read them directly and synthesize manually, then re-run
 - **Cursor Agent trust**: `--trust` required for headless/non-interactive mode
 - **Codex workspace**: `-C /path/to/workspace` to set working directory
 - **Claude Agent paths**: Write within workspace (e.g., `log/`), not `/tmp`
-- **Claude CLI (Opus 4.7)**: `claude -p --model claude-opus-4-7 --bare` runs as external process. Uses stdin pipe (like Codex). `--bare` required for review tasks (skips hooks, CLAUDE.md, avoids bias from project instructions). Without `--bare`, CLAUDE.md's three-layer response structure may distort review output
-- **Claude CLI parallelism**: Agent tool (internal, Opus 4.6) + Bash `claude -p` (external, Opus 4.7) run truly in parallel as separate processes
+- **Claude CLI (Opus 4.6 / 4.8)**: `claude -p --model claude-opus-4-6 --bare` (likewise `claude-opus-4-8`) runs as external process. Uses stdin pipe (like Codex). `--bare` required for review tasks (skips hooks, CLAUDE.md, avoids bias from project instructions). Without `--bare`, CLAUDE.md's three-layer response structure may distort review output
+- **Claude CLI parallelism**: Agent tool (internal, orchestrator model = Fable 5) + Bash `claude -p` (external, Opus 4.6 / 4.8) run truly in parallel as separate processes
 - **Claude CLI file access**: `claude -p` with `--bare` has no MCP tools or file access. Ensure review prompt includes all artifact content inline (rule #6). Use `--add-dir` + `--allowedTools "Read,Glob,Grep"` if file access is needed (but note: this loads CLAUDE.md unless `--bare` is also used)
 
 ## Prompt Generation Rules
@@ -709,13 +728,15 @@ Step 1: Generate review prompt
 Step 2: Detect environment and models
   - Run: which codex && which agent && which claude
   - Detect default models
-  - Report: "Auto mode: Codex (gpt-5.5), Agent (composer-2.5), Claude (opus-4.6), Claude CLI (opus-4.7)"
+  - Report: "Auto mode: Codex (gpt-5.5, gpt-5.4), Agent (composer-2.5), Claude Team (claude-fable-5), Claude CLI (opus-4.6, opus-4.8)"
 
-Step 3: Execute N reviews in parallel (default 4 reviewers)
-  - Bash(background): cat prompt.md | codex exec -C workspace -o log/review_codex.md -
+Step 3: Execute N reviews in parallel (default 6 reviewers)
+  - Bash(background): cat prompt.md | codex exec -m gpt-5.5 -C workspace -o log/review_codex_gpt5.5.md -
+  - Bash(background): cat prompt.md | codex exec -m gpt-5.4 -C workspace -o log/review_codex_gpt5.4.md -
   - Bash(background): agent -p --trust "Read prompt and review..." > log/review_cursor.md
-  - Agent(background): Claude Team (Opus 4.6) → write to log/review_claude_opus4.6.md
-  - Bash(background): cat prompt.md | claude -p --model claude-opus-4-7 --bare > log/review_claude_opus4.7.md 2>log/review_claude_opus4.7.stderr.log
+  - Agent(background): Claude Team (orchestrator model, Fable 5) → write to log/review_claude_team_fable5.md
+  - Bash(background): cat prompt.md | claude -p --model claude-opus-4-6 --bare > log/review_claude_opus4.6.md 2>log/review_claude_opus4.6.stderr.log
+  - Bash(background): cat prompt.md | claude -p --model claude-opus-4-8 --bare > log/review_claude_opus4.8.md 2>log/review_claude_opus4.8.stderr.log
 
 Step 4: Collect and validate
   - Wait for all to complete (background task notifications)
@@ -752,9 +773,11 @@ log/{artifact}_review{N}_{llm_id}_{date}.md       # Individual reviews
 log/{artifact}_review{N}_consensus_{date}.md       # Consensus analysis
 ```
 
-LLM identifiers: `claude_opus4.6`, `claude_team_opus4.6`,
-`claude_cli_opus4.7`, `codex_gpt5.5`, `codex_gpt5.4`, `cursor_composer2`, `cursor_gpt5.4`,
+LLM identifiers: `claude_team_fable5`, `claude_cli_opus4.6`, `claude_cli_opus4.8`,
+`codex_gpt5.5`, `codex_gpt5.4`, `cursor_composer2.5`, `cursor_gpt5.4`,
 `cursor_premium`
+(legacy, pre-2026-06-10: `claude_opus4.6`, `claude_team_opus4.6`, `claude_team_opus4.7`,
+`claude_cli_opus4.7`, `cursor_composer2`)
 
 ## Internal Agent Team Review
 
@@ -774,7 +797,7 @@ Compression ratio: parallel agent raw → Assembly ≈ 2:1
 - Don't advance to Phase N+1 before Phase N review converges
 - Don't re-review from scratch — each round checks only the delta
 - Don't use only internal agent team — different providers catch different bugs
-- Don't dismiss 1/4 (or 1/N) findings without evaluating substance
+- Don't dismiss 1/N findings without evaluating substance
 - Don't use Persona Assembly in every intermediate round (save for final gate)
 
 ---
@@ -793,6 +816,8 @@ Compression ratio: parallel agent raw → Assembly ≈ 2:1
 - Codex convergence: REJECT → REJECT → REJECT → APPROVE (4 rounds)
 - Self-referential review: v3.0 of this skill reviewed by its own process → v3.1
 - Self-referential review: v3.2 (4-reviewer update, 2026-04-19) reviewed with new 4-reviewer default (Opus 4.6 + 4.7 + Codex + Composer-2). 4/4 APPROVE WITH CHANGES R1. Findings integrated → v3.3
+- Roster update (v3.5, 2026-06-10): Fable 5 replaces Opus 4.7 as orchestrator/team slot; Opus 4.8 added as second subprocess CLI reviewer alongside Opus 4.6. 4.6 retained for its documented complementary bias (ambiguity-preserving, self-reference-friendly); 4.7 retired as its register is covered by 4.8 and Fable 5. 4.8/Fable 5 bias profiles uncalibrated — record (a)/(b)/(c) breakdowns per round until profiles accumulate in `multi_llm_reviewer_evaluation`
+- Self-referential review of v3.5 (2 rounds, 2026-06-10/11, first run of the 6-reviewer roster): R1 REVISE (1 APPROVE / 4 REJECT — stale pre-v3.5 passages) → fixes → R2 3/6 APPROVE (4.6, 4.8, codex 5.4) with Cursor contributing a code-grounded correction (subprocess strategy keeps the full roster). 4.6/4.8 verdicts split along the predicted lenient/strict axis in R1 and converged to APPROVE in R2
 
 **Key insight**: Design reviews and implementation reviews find
-**categorically different bugs**. Both phases are necessary.
+**categorically different bugs**. Both phases are necessary.

data/templates/skillsets/multi_llm_review/config/multi_llm_review.yml

@@ -4,21 +4,24 @@
 # to avoid duplication.
 
 # Convergence rules
-# Roster has 5 reviewers (claude_team_opus4.7, claude_cli_opus4.6,
-# codex_gpt5.4, codex_gpt5.5, cursor_composer2.5). Rules are ratio-based
-# (parser interprets "N/M" as N/M fraction applied to successful count),
-# so the literal numerator/denominator is informational; what matters is
-# the ratio.
-convergence_rule: "3/5 APPROVE"  # 60% of successful reviewers must APPROVE
+# Roster has 6 reviewers (claude_team_fable5, claude_cli_opus4.6,
+# claude_cli_opus4.8, codex_gpt5.4, codex_gpt5.5, cursor_composer2.5).
+# Rules are ratio-based (parser interprets "N/M" as N/M fraction applied
+# to successful count), so the literal numerator/denominator is
+# informational; what matters is the ratio.
+convergence_rule: "4/6 APPROVE"  # ceil(6 * 0.6) = 4 of the 6-reviewer roster
 min_quorum: 2                    # minimum successful reviews for any verdict
 
 # Self-referential orchestrator exclusion.
-# When the caller passes orchestrator_model (e.g. "claude-opus-4-7"), any
+# When the caller passes orchestrator_model (e.g. "claude-fable-5"), any
 # roster entry with that exact model is dropped before dispatch — the
 # orchestrator should not review its own output.
-# When at least one entry is excluded, convergence_rule_after_exclusion
-# replaces convergence_rule for that dispatch (4 reviewers → 3 reviewers
-# makes "3/4 APPROVE" require unanimity, which is too strict).
+# Under the "exclude" strategy, when at least one entry is excluded,
+# convergence_rule_after_exclusion replaces convergence_rule for that
+# dispatch (the literal rule for the full roster would otherwise be too
+# strict for the reduced count). The "subprocess" strategy keeps the full
+# roster; the default "delegate" strategy re-adds the slot at collect, so
+# the full-roster rule governs there.
 exclude_orchestrator_model: true
 
 # Default orchestrator_strategy when the caller does not specify one.
@@ -29,9 +32,15 @@ exclude_orchestrator_model: true
 # "exclude": legacy behavior — drop the matching reviewer entirely.
 # "subprocess": spawn fresh claude -p for the matching reviewer.
 default_orchestrator_strategy: "delegate"
-# After excluding 1 orchestrator from 5 → 4 reviewers. Maintain the same
-# 60% ratio (ceil(4 * 0.6) = 3 → 3 of 4 must APPROVE).
-convergence_rule_after_exclusion: "3/4 APPROVE"
+# After excluding 1 orchestrator from 6 → 5 reviewers. Same ceil(N * 0.6)
+# majority basis (ceil(5 * 0.6) = 3 → 3 of 5 must APPROVE). Note the two
+# rules are not the same literal ratio (4/6 ≈ 0.67 vs 3/5 = 0.60); since the
+# parser applies the ratio to the successful count, the full-roster rule is
+# slightly stricter when some reviewers fail. Accepted as-is.
+# Applies to the "exclude" strategy only — "subprocess" keeps the full
+# roster, and "delegate" re-adds the orchestrator slot at collect, so the
+# full-roster rule (4/6) governs both (verified live 2026-06-10).
+convergence_rule_after_exclusion: "3/5 APPROVE"
 
 # Two-phase delegation (orchestrator_strategy: "delegate").
 # Phase 1 dispatches subprocess reviewers synchronously, persists their
@@ -73,7 +82,9 @@ delegation:
     wait_still_pending_streak_limit: 3        # consecutive still_pending returns before crashed/wait_exhausted
 
 # Dispatch settings
-timeout_seconds: 300              # global deadline for all reviewers
+timeout_seconds: 600              # global deadline for all reviewers
+                                  # (raised 300 -> 600 on 2026-06-10: live 6-roster
+                                  # run took 381s wall-clock with max_concurrent: 2)
 max_concurrent: 2                 # semaphore limit (2 for laptop, 4 for CI)
 
 # Reviewer roster
@@ -84,16 +95,31 @@ max_concurrent: 2                 # semaphore limit (2 for laptop, 4 for CI)
 # tool receives a `complexity` argument (or auto-detects it), these defaults
 # are overridden per-dispatch by the effort_map below.
 reviewers:
+  # Orchestrator slot. With orchestrator_strategy "delegate" (the default),
+  # this entry is replaced by the orchestrator's own persona Agent Team
+  # review when orchestrator_model matches. Updated 2026-06-10: Fable 5
+  # replaces Opus 4.7 as the session default model.
   - provider: claude_code
-    model: claude-opus-4-7
+    model: claude-fable-5
     effort: medium
-    role_label: claude_team_opus4.7
+    role_label: claude_team_fable5
 
   - provider: claude_code
     model: claude-opus-4-6
     effort: medium
     role_label: claude_cli_opus4.6
 
+  # Opus 4.8 added 2026-06-10, replacing Opus 4.7 in the roster. Rationale:
+  # 4.7's register (strictness / systematizing) is covered by its direct
+  # successor 4.8 and by Fable 5; 4.6 stays for its documented complementary
+  # bias (ambiguity-preserving, self-reference-friendly). 4.8's reviewer
+  # bias profile is not yet calibrated — record (a)/(b)/(c) breakdowns per
+  # round (see multi_llm_reviewer_evaluation) until a profile accumulates.
+  - provider: claude_code
+    model: claude-opus-4-8
+    effort: medium
+    role_label: claude_cli_opus4.8
+
   - provider: codex
     model: gpt-5.4
     effort: medium

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: kairos-chain
 version: !ruby/object:Gem::Version
-  version: 3.30.0
+  version: 3.31.0
 platform: ruby
 authors:
 - Masaomi Hatakeyama
 autorequire:
 bindir: bin
 cert_chain: []
-date: 2026-06-08 00:00:00.000000000 Z
+date: 2026-06-10 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: minitest
@@ -128,6 +128,7 @@ files:
 - lib/kairos_mcp/daemon/wal.rb
 - lib/kairos_mcp/daemon/wal_phase_recorder.rb
 - lib/kairos_mcp/daemon/wal_recovery.rb
+- lib/kairos_mcp/drift_detection/correspondence_checker.rb
 - lib/kairos_mcp/dsl_ast/ast_engine.rb
 - lib/kairos_mcp/dsl_ast/decompiler.rb
 - lib/kairos_mcp/dsl_ast/drift_detector.rb