Kratos v2.43.0 — Anthropic Best Practices Alignment

What Changed

Full diagnostic and rewrite of 12 agent prompts to align with Anthropic's Claude 4 Best Practices.

Key Improvements

Reasoning over commands — Replaced heavy-handed CRITICAL, MUST, NEVER, ALWAYS directives with explanations of why the behavior matters. Claude 4.6 responds better to reasoning than commands.

Deferred Arena reads — Ares, Artemis, and Hades no longer front-load Arena file reads. In pipeline mode, upstream agent summaries in status.json provide sufficient context. Arena is consulted only when a specific question arises that summaries don't answer.

Pipeline-aware boundaries — Every agent's boundary section now explains why staying in lane matters (avoids duplicate work, prevents circular dependencies, respects pipeline gates).

Agents Modified (12/17)

Agent	Change
Ananke	CRITICAL: Execution Rules → Two-Path Storage System with fallback reasoning
Prometheus	CRITICAL output blocks → explains Kratos parses programmatically
Hermes	MANDATORY checklist → leads with SubagentStop hook gate reasoning
Ares	Exploration ban reasoning; Arena reads deferred in pipeline mode
Hades	Logs-first debugging protocol; Arena deferred to Phase 3
Athena	Boundary reasoning: tech decisions here contradict spec phase
Hephaestus	Boundary reasoning: requirements bypass PRD review gate
Apollo	Boundary reasoning: fixes skip implementation pipeline
Artemis	Arena reads deferred; test code boundary explains duplication
Daedalus	Boundary reasoning: mixing concerns creates circular dependencies
Metis	READ-ONLY reasoning: Arena neutrality trust model
Cassandra	Boundary reasoning: fixing skips pipeline; delta focus

Unchanged Agents (5/17)

Nemesis, Hera, Themis, Clio, Mimir — already aligned or low severity.

v2.42.3

Summary

• restore the 533b39d43ae simplification style across Kratos pipeline docs and agent prompts
• normalize Kratos to a contiguous 1-10 pipeline with Athena removed from spec review
• align hooks, status/schema docs, README examples, and Go check/hook tests with the updated stage flow

Kratos v2.42.1

Version 2.42.1 release

v2.41.0 — Token Optimizations

What's Changed

Token Cost Reductions

• Model pinning: All agents now use explicit model IDs (claude-sonnet-4-6, claude-opus-4-6) instead of shorthands that were silently resolving to older model versions
• Hermes + Apollo → sonnet: Code review and SA spec review downgraded from opus to sonnet — saves 2 opus calls per pipeline run
• Stage 6 removed: Athena PM spec review eliminated; Apollo's SA review provides sufficient coverage — saves 1 opus call per pipeline run
• Gap analysis collapse: Athena now writes the PRD directly in Phase 1 when requirements are already clear (ambiguity ≤ 0.20), skipping Phase 2 — collapses 2 opus calls to 1 on well-specified features

Output Length Reductions

• Conditional tech-spec template: Sections 3–6 (Data Model, API Design, Security, Performance) now have skip guards — agents omit inapplicable sections instead of writing N/A boilerplate
• Artemis scoped down: Test plan agent now produces structured test cases (name/scenario/input/expected) instead of full runnable code — reduces output tokens significantly

Net Impact (measured, full pipeline run)

• Pipeline goes from 11 stages to 10 stages
• Happy-path opus calls reduced from 6 to 3 (Athena, Hephaestus only on normal features)
• Baseline cost: ~$6.86/run on Max 5x (~7.6% of 5-hour budget)

Go CLI

• Version bumped to v2.41.0-go

v2.40.0 — CLI-only status.json enforcement

What's Changed

Breaking: All agents now use CLI exclusively for status.json

All pipeline agents (Ares, Artemis, Cassandra, Daedalus, Hephaestus, Hera, Nemesis, Themis) have had direct Write/Edit fallbacks removed. Status updates must go through the kratos binary — no exceptions.

New CLI flags

• pipeline update --summary <text> — attach a stage summary in one call; downstream agents read this before deciding to open the full deliverable doc
• pipeline set-tasks <json> — register the Ares task list in status.json
• pipeline task-done <task-id> — mark a single Ares task complete

Removed: feature README.md

pipeline/start.md no longer creates .claude/feature/<name>/README.md. It duplicated prd.md and status.json. PRD.md is the feature overview; status.json is the authoritative pipeline state.

Test harness additions

• 9 new cli-compliance-* tasks (one per pipeline agent) in test-harness/src/tasks.mjs
• New test-harness/src/cli-compliance-validator.mjs — scans messages.jsonl for violations (direct Write/Edit on status.json) and compliance signals (Bash calls with kratos pipeline)
• npm run validate:cli-compliance — run the validator against the most recent cli-compliance test run

Bug fixed

The plugin cache binary at ~/.claude/plugins/cache/lizard-plugins/kratos/ was stale (missing --summary), causing Artemis to fall back to direct Edit. The binary is now correctly rebuilt and must be refreshed in the cache after each binary update.

Kratos v2.39.0 — Output Token Reduction (-52%)

What's Changed

Token Optimization — Pipeline Deliverable Verbosity Reduction

Measured via test harness (full 10-agent pipeline, same feature type):

Metric	v2.38.0	v2.39.0	Δ
Output tokens	179,518	86,852	-52%
Total tokens	15,851,865	10,881,852	-31%
Duration	71 min	35 min	-51%

Agent Changes

• tech-spec-template: Removed Section 8 (Testing Strategy) and Section 9 (Rollout Plan) — Artemis owns testing; rollout is rarely applicable
• prd-template: Sections 8 (External APIs) and 9 (Research Summary) are now conditional — only rendered when Mimir was called or external APIs are involved; Section 10 appendix removed
• apollo: Replaced "maximum scrutiny regardless of priority" with proportional review rule — depth scales to feature surface area
• hermes: Review output now omits clean tiers; only tiers with findings are listed
• nemesis: Merged dual-part (Devil's Advocate / User Advocate) BLOCKING/MAJOR/MINOR sections into a single ## Findings list with flag type tags; score table collapsed to one line
• athena: decisions.md uses flat bullet list instead of markdown table (same information, ~40% fewer lines)
• hera: Steps 2 and 3 collapsed into a single mapping+verification pass; prd-alignment.md outputs a count summary instead of re-enumerating all passing criteria

Test Harness Fixes (v2.38.x → v2.39.0)

• Fixed KRATOS_PLUGIN_PATH (was pointing to lizard-market/ instead of plugins/kratos/)
• Task prompts now use explicit /kratos:main slash commands
• Project reset before each run (clears .claude/feature, .claude/.Arena, impl dirs)
• Token aggregation now sums all assistant message usage including subagents (raw + cacheCreate + cacheRead)
• False error fix: process exit-code-1 after a success result no longer marks task as failed
• Removed invalid optimization experiments (main-optimized, context-optimizer, test-optimized skill)