ELAI agents

Canonical list of the 15 ELAI sub-agents. Each agent runs in its own context window that is garbage-collected on return. The main LLM dispatches via the Agent tool with subagent_type=general-purpose and a prompt that asks the agent to read its own .md file under .claude/agents/. See .claude/rules/orchestrate-protocol.md for the dispatch contract.

This file is the workflow-root agent catalog. It is a hand-written digest of the full agent definitions in .claude/agents/*.md (the source of truth). It lives at the workflow root so any agent harness (Claude Code, OpenAI Codex, Cursor, Devin) loads it as always-on context. It follows the cross-vendor AGENTS.md convention; agents are alphabetized for stable diffs.

NOTE: this is distinct from the runtime export target. The Rust PlaybookCurator::export_verified (I-12 / ARM-4) writes verified-rule blocks to <project_root>/AGENTS.md inside .elai_cc; that slot is for generated output, not this hand-written catalog.

Source-grounded by:

• code.claude.com/docs/en/skills (Anthropic SKILL.md v1 spec)
• developers.openai.com/codex/skills (Codex skills format)

auditor

Sole test runner. Reads contracts, runs cargo check / cargo test / pnpm svelte-check, classifies failures via .claude/tools/failure_taxonomy.py. Returns PASS / WARN / BLOCK with a failure category. Main LLM does NOT run tests directly.

context-compiler

Pre-execution context packer. Reads plan + spec references, summarises into a context-package-{phase}.md file the executor reads. Owns memory-bank reads in HEAVY flows; the main LLM never reads memory-bank directly in those flows.

dashboard-smoker

Post-edit dashboard verification. Triggers on Write/Edit of files under .elai_cc/crates/dashboard-v2/. Runs pnpm svelte-check + cargo check on dashboard-v2, optionally launches tauri:dev for manual smoke.

fixer

Targeted defect-repair agent. Spawned after auditor returns BLOCK with a specific failure category. Reads the failure, the failing source, the plan, and authors a focused fix. Returns the patch + test outcome.

impact-mapper

Pre-edit blast-radius analysis. Reads .claude/tools/elai.map to find every callsite of the symbols a planned change will touch. Returns the set of files that need review.

integration-explorer-elai

Cross-system wiring agent. Walks orchestrator / dashboard-v2 / sidecar / CLI seams to find existing extension points before the executor lands new wiring. Output is an exploration-plan.md the main LLM consumes indirectly via auditor sign-off.

memory-writer

Sole writer to memory-bank/. Updates activeContext.md, changelog.md, progress.md, projectIndex.md after every audit pass. Optionally appends to gotchas.md, systemPatterns.md, operatingManual.md when the audit surfaces a reusable pattern. Calls mcp_memorize / mcp_log_change for RAG indexing.

paper-ingestor

Source-grounding agent. Modes:

• mode=paper — given an arXiv ID, downloads + summarises the paper into workflow/papers/<paper-slug>-interface.md.
• mode=source — given a file or SDK name, ingests the source and writes an interface map. Decompiled output is marked LOW reliability.

prompt-engineer

Reformats user prompts before implementation. Picks the GRC mcp_think budget for the task (simple=3/600, moderate=5/1000, complex=8/1500).

scope-planner-elai

Multi-area partition planner. When a plan touches more than one subsystem (orchestrator + dashboard-v2 + sidecar), this agent computes the per-subsystem partition + dispatch order so the executor doesnt re-explore work the explorer already covered.

sidecar-health

Remote sidecar boot + health check. Runs the canonical 3-step health check from CLAUDE.md before any live inference call: ssh process check, local Test-NetConnection, curl /v1/messages probe.

spec-grounder

Verifies "Spec references:" + "Reads (mandatory)" blocks in a plan resolve to real spec line ranges. Used by orchestrate runs to catch spec_drift before executor lands code.

synthesizer-elai

Cross-explorer merge agent. When multiple integration-explorer-elai agents return parallel exploration plans, the synthesizer merges into a single exploration-plan.md the executor consumes.

test-attributer

After a test passes, links the passing test back to the source claim it verifies. Builds the contract-test coverage map used by the auditor.

wiring-contract-validator

Verifies a plans WIRING_CONTRACT block (produces / consumes / reachable / stub_ban / observable_effect) against the actual code after executor lands the change. Used inside orchestrate sub-phase audits.

Dispatch invocation template

Per .claude/rules/orchestrate-protocol.md H-8: these agents are NOT in the Agent tools subagent_type enum. Dispatch via:

subagent_type = general-purpose
prompt        = "Read F:\\.ELAI_workflow\\.claude\\agents\\<name>.md
                 and follow it strictly. <task-specific args here>."

The main LLM does NOT read agent .md files itself and does NOT ad-lib agent instructions. See .claude/rules/main-llm-role.md.