nyldn/claude-octopus v8.12.0

on GitHub

Added

New /octo:deck Slide Deck Skill (Issue #29):

• skill-deck.md — 4-step pipeline: gather brief, optional research, outline approval gate, PPTX render
• deck.md command — /octo:deck entry point with usage examples
• AskUserQuestion-based outline wireframe gate — user approves slide structure before rendering
• Delegates to document-skills:pptx for PowerPoint generation (no new dependencies)
• Audience-specific slide templates: executives, engineers, investors, general
• Integration with /octo:discover for research-backed presentations

Changed

• Skill count: 43 → 44
• Command count: 38 → 39

[8.11.0] - 2026-02-13

Added

New OpenRouter-Backed Agent Types (P0 - Issue #24):

• openrouter-glm5 agent type → z-ai/glm-5 (77.8% SWE-bench, $0.80/$2.56 per MTok, 203K ctx)
• openrouter-kimi agent type → moonshotai/kimi-k2.5 (cheapest, $0.45/$2.25 per MTok, 262K ctx)
• openrouter-deepseek agent type → deepseek/deepseek-r1 (visible reasoning traces, $0.70/$2.50 per MTok, 164K ctx)
• openrouter_execute_model() function for fixed-model OpenRouter API calls
• All three types integrated into get_agent_command(), get_agent_command_array(), AVAILABLE_AGENTS, get_agent_model(), get_model_pricing(), is_agent_available_v2(), get_fallback_agent(), get_tiered_agent_v2()

Intelligent Task-Based Routing (P1):

• get_tiered_agent_v2() now routes OpenRouter provider to model-specific agents based on task type
• review tasks → openrouter-glm5 (lowest hallucination rate, best SWE-bench)
• research/design tasks → openrouter-kimi (262K context window, cheapest per token)
• security/reasoning tasks → openrouter-deepseek (visible <think> reasoning chains)
• Model-specific context windows: GLM-5 203K, Kimi 262K, DeepSeek R1 164K

[8.10.0] - 2026-02-13

Fixed

Gemini CLI Headless Mode (P0 - Issue #25):

• Fixed Gemini CLI launching in interactive REPL mode instead of headless
• Added -p "" flag + stdin-based prompt delivery for reliable headless execution
• Added -o text for clean output, --approval-mode yolo replaces deprecated -y
• Removed invalid --pipe flag; fixed enforce_context_budget() ordering

[8.9.0] - 2026-02-13

Added

Contextual Codex Model Routing (P0):

• select_codex_model_for_context() function - automatically selects the best Codex model based on workflow phase, task type, and user config
• get_codex_agent_for_phase() helper - maps phases to appropriate codex-* agent types
• Per-phase model routing in providers.json phase_routing section
• 5-tier model precedence: env var > task hints > phase routing > config defaults > hard-coded

New Agent Types (P0):

• codex-spark agent type - routes to GPT-5.3-Codex-Spark (1000+ tok/s, 15x faster, 128K context)
• codex-reasoning agent type - routes to o3 (deep reasoning, 200K context)
• codex-large-context agent type - routes to gpt-4.1 (1M context window)
• All new types integrated into get_agent_command(), get_agent_command_array(), AVAILABLE_AGENTS, is_agent_available_v2(), get_fallback_agent()

New Agent Personas (P1):

• codebase-analyst - large-context agent using gpt-4.1 for analyzing entire codebases
• reasoning-analyst - deep reasoning agent using o3 for complex trade-off analysis

Enhanced Model Support (P0):

• GPT-5.3-Codex-Spark pricing and model entry
• o3 and o4-mini reasoning model entries
• gpt-4.1 and gpt-4.1-mini large-context model entries (1M token window)
• gpt-5.1 and gpt-5-codex legacy model entries

Changed

Updated API Pricing (P0):

• Corrected gpt-5.3-codex to $1.75/$14.00 per MTok (was $4.00/$16.00)
• Corrected gpt-5.2-codex to $1.75/$14.00 per MTok (was $2.00/$10.00)
• Corrected gpt-5.1-codex-mini to $0.30/$1.25 per MTok (was $0.50/$2.00)
• Added 6 new model price entries (spark, o3, o4-mini, gpt-4.1, gpt-4.1-mini, gpt-5.1)

Agent Config v2.0 (P1):

• agents/config.yaml upgraded to version 2.0
• Added phase_model_routing section with per-phase Codex model defaults
• Added fallback_cli field for graceful agent degradation
• Code reviewer switched to codex-spark for fast PR feedback (15x faster)
• Performance engineer switched to codex-spark for rapid analysis
• Security auditor stays on full gpt-5.3-codex for thorough analysis

Model Config Command v2.0 (P1):

• model-config.md rewritten for v2.0 with comprehensive model catalog
• Added phase <phase> <model> subcommand for per-phase routing
• Added reset phases subcommand
• Full Spark vs Codex comparison table
• Pricing table for all 12+ supported models

Config Schema v2.0 (P1):

• providers.json schema upgraded to v2.0
• Added phase_routing section (9 phase-to-model mappings)
• Added spark_model, mini_model, reasoning_model, large_context_model fields to codex provider
• Backward compatible with v1.0 configs

Fallback Chains (P1):

• Extended get_fallback_agent() with codex-spark → codex → gemini chain
• Extended with codex-reasoning → codex → gemini chain
• Extended with codex-large-context → codex → gemini chain
• Spark falls back gracefully to standard Codex when unavailable

Cost Awareness (P2):

• Updated CLAUDE.md cost estimates to reflect Feb 2026 API pricing
• Cost range updated from ~$0.02-0.10 to ~$0.01-0.15 per query

Tier Model Selection (P1):

• get_tier_model() extended with codex-spark (always spark), codex-reasoning (o4-mini/o3), codex-large-context (gpt-4.1-mini/gpt-4.1) tiers