@smartmemory/compose 0.2.52-beta → 0.2.54-beta

package/.claude/skills/compose/references/hermes-tools.md

@@ -1,80 +0,0 @@
-# Compose on Hermes Agent — Tool Mapping
-
-SKILL.md is written against **Claude Code** tool names. This file maps each
-primitive the compose lifecycle uses to its **Nous Hermes Agent** equivalent, so
-the lifecycle can run under Hermes. Scope: the *interactive* `/compose` lifecycle
-(the headless `compose` CLI is already harness-agnostic — it runs from any shell
-and delegates agent execution to Stratum).
-
-Hermes shares the three primitives the lifecycle depends on — **skills**, an
-**MCP client**, and **parallel subagent delegation** — so this is a name/packaging
-adaptation, not an architecture port. The one genuine gap is plan mode (Phase 6);
-see below.
-
-## Action mapping
-
-| Action compose requests (Claude Code) | Hermes equivalent |
-|---|---|
-| Invoke a skill (`/compose`, `superpowers:*`) — **Skill** tool | Skills load natively from `~/.hermes/skills/`. Install with `hermes skills install <git-url\|file\|registry-id>` (SKILL.md frontmatter convention). |
-| Dispatch a subagent (`compose-explorer`, `compose-architect`, `general-purpose`) — **Agent/Task** tool | `delegate_task` (runtime tool) — spawns an isolated child agent with its own conversation, terminal, and toolset; only the final summary returns. |
-| Dispatch 2-3 subagents **in parallel** (Phase 1 explorers, Phase 3 architects) | `delegate_task` with a **`tasks` array** — runs in parallel via a thread pool, `max_concurrent_children` (default **3**, matches compose's fan-out). Orchestrator→leaf nesting needs `delegation.max_spawn_depth: 2`. |
-| **Plan mode** (Phase 6) — `EnterPlanMode` / `ExitPlanMode` | **No direct equivalent.** Degrade to a flat gate: write `plan.md`, present it in chat, human approves before Phase 7. Compose already supports this fallback ("If Stratum unavailable, continue with flat prompt chain"). |
-| Task tracking ("create a todo", "mark complete") — **TodoWrite** | `hermes kanban` (statuses todo/triage/scheduled/blocked/running/done; `hermes kanban create "<title>"`). Optional/advisory — skip if not wired. |
-| Read a file — **Read** | Native file read, `@file` context injection, or shell (`cat`/`head`). |
-| Create / edit a file — **Write / Edit** | Native file tools, or a filesystem MCP server (`hermes mcp add fs --command ...`). |
-| Run a shell command — **Bash** | Hermes shell / per-subagent terminal session. |
-| Search contents / find files — **Grep / Glob** | Shell (`rg`, `grep`, `find`). |
-| compose tracker (`mcp__compose__*`) | `hermes mcp add compose --command "<compose-mcp-launch>"`, then call natively. |
-| Stratum (`mcp__stratum__stratum_*`, incl. `stratum_agent_run type=codex`) | `hermes mcp add stratum --command "<stratum-mcp-launch>"`, then call natively. **Codex review works unchanged** — it's an MCP call, harness-agnostic. |
-| Bind flow → vision item — `POST /api/stratum/bind` (HTTP) | Shell `curl` (harness-agnostic). |
-| Project instructions — **CLAUDE.md** | **AGENTS.md** (Hermes auto-injects `AGENTS.md`/`SOUL.md`/`USER.md`/`MEMORY.md` unless `--ignore-rules`). |
-
-## Setup (one-time, per workspace)
-
-```bash
-# 1. Install the compose skill into Hermes
-hermes skills install <compose-skill-git-url>        # → ~/.hermes/skills/compose/
-
-# 2. Register the MCP servers the lifecycle reads/writes through
-hermes mcp add compose  --command "<compose-mcp-launch-command>"
-hermes mcp add stratum  --command "<stratum-mcp-launch-command>"
-hermes mcp test compose && hermes mcp test stratum   # verify discovery
-
-# 3. Mirror compose's project guidance for auto-injection
-#    (compose's CLAUDE.md content → AGENTS.md at the project root)
-```
-
-## What does NOT port 1:1 (the real work)
-
-1. **`compose-explorer` / `compose-architect` are Claude agent files** (installed to
-   `~/.claude/agents/` by `compose setup` → `syncSkills()`). Hermes has no agent-file
-   format; re-express each as a **`delegate_task` prompt template** carrying the same
-   mandate (explorer = read-only codebase mapping; architect = competing-mandate
-   proposals). The prompts in SKILL.md Phases 1/3/4 are the source of truth — port
-   those strings, not the `.md` agent files.
-
-2. **`compose setup` targets Claude paths** (`~/.claude/skills/`, `~/.claude/agents/`,
-   `.compose-skills.json`). A Hermes path would install to `~/.hermes/skills/` via
-   `hermes skills install`. This is a `bin/compose.js syncSkills()` branch, not a
-   lifecycle change.
-
-3. **Plan mode (Phase 6)** has no Hermes analog — use the flat-gate degrade above.
-   This is the only lifecycle *behavior* that changes shape; everything else is a
-   rename.
-
-4. **Tier→model mapping is Claude-centric** (`server/model-tiers.js` maps
-   `critical/standard/fast` → Opus/Sonnet/Haiku). Under Hermes-as-driver this only
-   matters if you also make Hermes the *executing agent* — which is a **Stratum**
-   change (add a `hermes` provider to Stratum's agent runner), out of scope for this
-   harness-driver adaptation. See SKILL.md "Stratum Integration".
-
-## Quick decision guide
-
-- **Just want to launch builds from Hermes?** Don't port anything — run the headless
-  `compose build`/`compose fix` CLI from the Hermes shell. It executes via Claude/Codex
-  under Stratum regardless of the launching harness.
-- **Want the interactive lifecycle inside Hermes?** Use this mapping: install the skill,
-  register the two MCP servers, port the two agent prompts to `delegate_task`, accept the
-  flat-gate Phase-6 degrade.
-- **Want Hermes to be the building agent?** That's the Stratum provider-addition path,
-  not this file.