/orchestrating-agents

orchestrating-agents - Changelog

All notable changes to the orchestrating-agents skill are documented in this file. The format is based on Keep a Changelog.

[0.7.0] - 2026-09-09

Other

• prompt-audit: dated prompting patterns across the skill catalogue (#791)

[0.6.0] - 2026-08-12

Fixed

• Surface routing no longer claims the native runtime lacks inter-agent messaging. The "reach back into this skill for what the runtime lacks" line listed AgentPool messaging alongside stall detection and ConversationThread. Claude Code and Cowork ship SendMessage and ListAgents; AgentPool reimplements them worse. The skill previously mentioned neither tool anywhere, so a reader following it would hand-roll a fan-out for messaging the runtime already provides.

Added

• Native inter-agent messaging subsection under the native-subagents branch, covering four behaviors measured 2026-08-12 that the official docs do not state: the incoming envelope's from attribute is the agent type and fails as a reply address; subagents have no ListAgents at all, making the topology a star through the main conversation rather than a mesh; delivery queues and never interrupts a running tool; and a send resumes a completed agent with full context that the agent cannot detect, at transcript-replay cost.
• Note that anthropics/claude-code#48160 and ruvnet/ruflo#2028 report subagents cannot originate SendMessage, contradicted by a CCotw measurement on 2026-08-12 — flagged as environment-dependent and to be verified locally rather than designed around.

[0.5.0] - 2026-07-30

Other

• orchestrating-agents v0.5.0 — three-surface routing, Gemini via CF, credential-handling fix (#749)
• Deprecate mapping-codebases; adopt ruff 0.16.0 baseline (#747)

[0.5.0] - 2026-07-30

Changed

• Surface routing rewritten as a three-surface matrix (claude.ai / Cowork / Claude Code) against three engines (native subagents / Gemini via CF AI Gateway / this skill's httpx fan-out). Previously the block knew only two surfaces and had no Gemini row. Primary discriminator is now whether an Agent/Task/Workflow tool is callable, rather than filesystem inspection.
• Cowork documented as its own surface, including plugin-declared subagents (agents/*.md, frontmatter fields, plugin:agent naming) and the security restriction that hooks, mcpServers, and permissionMode are refused in plugin-shipped agents — a declared agent inherits the session's MCP connections and cannot bring its own.
• Setup no longer instructs storing ANTHROPIC_API_KEY.txt in project knowledge. Credentials must arrive by a path the shell reads directly.

Added

• Gemini-via-Cloudflare as an explicit option on all surfaces, not just claude.ai — for mechanical-but-large work and for model-family diversity in judge panels. Mechanics stay in invoking-gemini; the three failure modes that bite are named here (stale flash alias, thinking_level='minimal' for mechanical work, BYOK through the gateway).
• "Review is not delegable" stated as a cross-surface rule, with the note that cross-model review tools keep their own model config.
• See Also split into routing companions (agent-routing, invoking-gemini, subagent-delegation-protocol) versus this skill's own internals.

Fixed

• project_read warned against for credentials on every surface: small docs are returned inline, and the documented "large text is written to a local file" branch does not fire even at 64 KB (measured 2026-07-30). Writing is safe — project_write with local_path keeps contents out of context — reading is not.
• Token Efficiency claim corrected from "~800 tokens" to ~2k.

[0.4.0] - 2026-04-08

Added

• add AgentPool, EXECUTE_MODE, inter-agent messaging (v0.4.0) (#537)
• add mapping-features skill for behavioral web app documentation (#432)

Fixed

• remove shim and local _parse_json workarounds from tiling-tree (#314)
• streaming passes system=None to API when no system prompt given

Other

• Regenerate _MAP.md files after @lat: backlink insertion (#504)
• Lattice v2: bidirectional source-anchored knowledge graph (#503)

[0.3.0] - 2026-03-05

Added

• implement Symphony orchestration patterns (#349)
• Add orchestrating-skills skill (#319)

[0.3.0] - 2026-03-05

Added

• Continuation Turn Protocol (Task 1): ConversationThread now supports send_continuation(), turn_count property, max_turns limit, and configurable continuation_prompt
• Stall Detection (Task 2): New StallDetector class with activity timestamps, configurable timeout, heartbeat tracking, and background monitoring thread
• Task Lifecycle State Machine (Task 3): New task_state.py module with TaskTracker, TaskState enum, formal state transitions (Unclaimed → Claimed → Running → Completed/Failed/Cancelled), retry queuing, and category-based filtering
• Exponential Backoff (Task 4): New invoke_with_retry() in orchestration.py with configurable backoff (1s fixed for continuations, exponential for failures capped at max_ms)
• Reconciliation Hook (Task 5): New invoke_parallel_with_reconciliation() accepts optional reconcile callback to validate/prune tasks before dispatch
• Concurrency Control (Task 6): New ConcurrencyLimiter class with global and per-category semaphore-based limits
• Managed Parallel (Task 6): New invoke_parallel_managed() combining all Symphony patterns: retry, reconciliation, concurrency control, stall detection, and task tracking

Changed

• All new parameters are optional with backward-compatible defaults — existing interfaces unchanged

[0.2.0] - 2026-02-28

Added

• add line numbers, markdown ToC, and other files listing
• add code maps and CLAUDE.md integration guidance
• Delete VERSION files, complete migration to frontmatter
• Migrate all 27 skills from VERSION files to frontmatter

Changed

• migrate API credential management to project knowledge files

Fixed

• resolve issues #311 and #312 in claude_client.py
• limit markdown ToC to h1/h2 headings only

Other

• Update subagent models: default to Sonnet 4.6, add Haiku 4.5 support