agent-exploration

Files (7)

• SKILL.md
• AGENT.md · assets/
• analysis-template.md · assets/
• checklist.md · references/
• dispatch-rules.md · references/
• dispatch-slices.sh · scripts/
• install-explorer.sh · scripts/

SKILL.md (first lines)

# Agent Exploration

Generic parallel-research workflow. Use when a question requires deep reads across multiple distinct areas and the operator needs written artifacts (not chat output). The skill dispatches `explorer` agents in parallel through `compozy exec`; each invocation writes one analysis file. The parent then synthesizes a final summary.

The skill is self-contained and harness-agnostic. The explorer is a Compozy agent — a single definition (`assets/AGENT.md`) installed once at `~/.compozy/agents/explorer/AGENT.md`, discoverable by `compozy exec --agent explorer` from any harness (Claude Code, Codex CLI, Cursor, Droid, OpenCode, Pi, Gemini, Copilot, …). The parent runtime is selected per invocation via `--ide`, `--model`, and `--reasoning`. Parallel dispatch uses whatever async/background tool calls your harness exposes — the skill does not prescribe a specific tool.

## Required Reading Router

Match your step to the row. Read the listed files **in full before** producing output. They are not appendices — they are load-bearing. Inline content in this SKILL.md is a pointer, not a substitute.

| Step                                                    | MUST read                                                              |
| ------------------------------------------------------- | ---------------------------------------------------------------------- |
| Step 4 — composing every slice prompt                   | `references/dispatch-rules.md` + `assets/analysis-template.md`         |
| Step 5 — verifying outputs                              | `references/checklist.md` + `assets/analysis-template.md`              |
| Step 6 — synthesizing `summary.md`                      | every `<path>/analysis/NN_analysis_<slug>.md` from this round          |
| Any contract violation, fabricated evidence, or retry   | `references/dispatch-rules.md` (re-read; do not paraphrase from memory)|

## Reference Index

- `references/dispatch-rules.md` — the scoped-write contract: what the dispatched agent may write, may read, may run; tool allow/forbid lists; parent responsibilities; parallelism cap; failure handling. **Must be embedded verbatim in every slice prompt.**
- `references/checklist.md` — seven-section output validation checklist (installation, inputs, scout, dispatch, files, schema, summary). Run before authoring `summary.md`.
- `assets/analysis-template.md` — the canonical seven-section schema every dispatched agent fills (Overview, Mechanisms/Patterns, Relevant Sources, Transferable Patterns, Risks/Mismatches, Open Questions, Evidence) plus a Scope header.
- `assets/AGENT.md` — the Compozy explorer agent definition (frontmatter: `title`, `description`, `ide`, `model`, `reasoning_effort`, `access_mode`; body: the scoped-write contract and workflow). Installed by `scripts/install-explorer.sh` to `~/.compozy/agents/explorer/AGENT.md`.
- `scripts/install-explorer.sh` — bootstrap helper. Writes the bundled `assets/AGENT.md` to the Compozy global registry. Refuses to overwrite.
- `scripts/dispatch-slices.sh` — parallel dispatch runner. Takes `--ide`/`--model`/`--reasoning` plus 1-8 prompt files, backgrounds one `compozy exec` per file, waits via `wait $pid`, captures per-slice stdout/stderr/exit, and reports a summary. Zero external dependencies (native bash + the `compozy` binary).

## Bundled Path Rule

Resolve every bundled helper relative to the directory that holds this `SKILL.md`. When a command appears below as `scripts/<name>`, treat the actual invocation as `<agent-exploration-dir>/scripts/<name>` — expand `<agent-exploration-dir>` to the absolute skill directory before running.

## Required Inputs

- `--path <dir>` (required): Output directory. Analysis files are written under `<path>/analysis/`. Any project-relative or absolute directory works (for example `docs/research/<topic>/`, `tasks/<slug>/`, or a path outside the repo). The skill is not tied to any specific project layout.
- `--agents <num>` (optional, default 3, hard cap 8): Number of explorer invocations to dispatch in parallel. Caps prevent runaway dispatch when the prompt is vague.
- `--prompt <text>` (required): The research question. Quoted multi-line strings are supported. If omitted, the parent asks the operator before continuing.
- `--ide <ide>` (optional, default `claude`): Compozy runtime for each dispatched invocation. Forwarded to `compozy exec --ide`. Accepted values mirror `compozy exec`: `codex`, `claude`, `cursor-agent`, `droid`, `opencode`, `pi`, `gemini`, `copilot`.
- `--model <name>` (optional, default `opus`): Forwarded to `compozy exec --model`. See **Model Options** for the two supported profiles (Opus default; GLM 5.2 via `--ide pi --model openrouter/z-ai/glm-5.2`). When the chosen IDE does not support the requested model, `compozy` surfaces the error and the slice fails — re-dispatch with a compatible model.
- `--reasoning <effort>` (optional, default `xhigh`): Forwarded to `compozy exec --reasoning-effort`. Accepted values: `low`, `medium`, `high`, `xhigh`.

If `--path` or `--prompt` is missing, the parent asks the operator a single clarification before continuing. Never invent defaults for either. Apply the documented defaults for `--ide`, `--model`, `--reasoning` silently when omitted; reject an invalid `--ide` rather than falling back.

## Model Options

Two dispatch profiles are supported. Opus is the default; GLM 5.2 is a cheaper/faster alternative that, on deep multi-file reasoning, matched Opus on the core findings when its citations were verified against source. Pick per run via `--ide`/`--model`.

| Profile                         | Flags                                            | When                                                              |
| ------------------------------- | ------------------------------------------------ | ----------------------------------------------------------------- |
| Opus (default, high-fidelity)   | `--ide claude --model opus --reasoning xhigh`    | Hardest slices; when citation precision must be maximal.          |
| GLM 5.2 (fast/cheap, via pi)    | `--ide pi --model openrouter/z-ai/glm-5.2`       | Breadth research, cost-sensitive runs, high slice counts.         |

GLM-via-pi caveats (non-negotiable — a GLM run that ignores these fails silently):

- **Use the `pi` harness, never `opencode`, for GLM.** `opencode` routes GLM through an internal explorer subagent that emits little ACP output, starving Compozy's activity watchdog — the run stalls and is killed with zero output (reproduced twice). `pi` streams tool-calls continuously and completes.
- **`--reasoning` is NOT forwarded to `pi`** (its Compozy runtime spec drops reasoning args). Raise GLM's thinking either on the model — `--model openrouter/z-ai/glm-5.2:xhigh` — or once via `"defaultThinkingLevel": "xhigh"` in `~/.pi/agent/settings.json`. pi's built-in default is `medium`.
- **GLM auth** lives in `~/.pi/agent/auth.json` (`{"openrouter":{"type":"api_key","key":"sk-or-..."}}`), independent of the daemon env. Requires `pi` + `pi-acp` on PATH (`npm i -g @earendil-works/pi-coding-agent pi-acp`).
- GLM output can carry cosmetic identifier typos (dropped letters, e.g. `atempt`); Step 5's citation spot-check is the gate that catches them — never skip it on a GLM run.

## Output Layout

```
<path>/analysis/
├── 01_analysis_<slug-a>.md
├── 02_analysis_<slug-b>.md
├── 03_analysis_<slug-c>.md
└── summary.md
```

- File numbering is zero-padded to two digits (`01`, `02`, …, `08`).
- Each slug is a short kebab-case identifier the parent assigns during the scout (Step 3), reflecting that slice's focus.
- `summary.md` is parent-authored synthesis, not a dispatched output.

## Procedures

**Step 1: Verify the explorer agent is installed**

1. Confirm the `compozy` binary is on `PATH` (e.g. `command -v compozy`). If missing, abort with a one-line message instructing the operator to install Compozy from `/Users/pedronauck/dev/compozy/looper`. Do not fall back to harness-native subagents.
2. Check for the explorer agent in the Compozy registry:
   - Global (preferred): `~/.compozy/agents/explorer/AGENT.md`.
   - Workspace override (optional, takes precedence): `<repo>/.compozy/agents/explorer/AGENT.md`.
   At least one must be present.
3. If both are absent, ask the operator a single question: "The `explorer` Compozy agent is not installed at `~/.compozy/agents/explorer/AGENT.md`. Install it now? [yes/no]". Do not proceed silently.
4. On "yes", run the bundled bootstrap helper: `<agent-exploration-dir>/scripts/install-explorer.sh`. The helper installs to `~/.compozy/agents/explorer/AGENT.md` and refuses to overwrite an existing file.
5. On "no", abort the dispatch with a one-line message.
6. After installation, re-check that the file exists before continuing.
7. (Optional sanity check) Run `compozy agents inspect explorer` and confirm the entry is discoverable. Any error here blocks dispatch.

**Step 2: Resolve inputs**

1. Parse `--path`, `--agents`, `--prompt`, `--ide`, `--model`, `--reasoning` from the invocation. If `--path` or `--prompt` is missing, ask the operator and stop.
2. Default `--agents` to `3` when omitted. Reject values below 1 or above 8 — ask the operator to choose a value in range.
3. Default `--ide` to `claude`, `--model` to `opus`, `--reasoning` to `xhigh` when omitted. Validate `--ide` against the accepted list (`codex`, `claude`, `cursor-agent`, `droid`, `opencode`, `pi`, `gemini`, `copilot`); reject invalid values with a clear message instead of silent fallback. Do not validate `--model` ahead of time — let `compozy` surface incompatibilities.
4. Resolve `--path` to an absolute path. If the directory does not exist, ask the operator whether to create it before continuing.
5. Create `<path>/analysis/` if absent. The dispatched agents refuse to write into a missing directory.

**Step 3: Parent-led initial scout (MANDATORY — do not skip)**

The scout is the load-bearing step that prevents wasted parallel dispatch. The parent must do this work itself before any slice is launched.

1. Perform a brief read-only exploration of the problem space using `Glob`, `Grep`, and targeted `Read` calls. The scout's job is to learn enough about the territory to divide it well — not to produce analysis content. Cap the scout at 8–15 tool calls; deep reading belongs to the dispatched agents.
2. From the scout, identify exactly `--agents` distinct slices that are:
   - **Non-overlapping** — two slices should not require reading the same primary files for the same purpose.
   - **Independently answerable** — a slice's analysis must not depend on another slice's output.
   - **Aligned with the operator's `--prompt`** — every slice serves the original research question.
3. For each slice, assign:
   - A two-digit ordinal (`01`..`08`).
   - A short kebab-case slug (≤ 4 words) reflecting that slice's focus (e.g. `state-machine`, `event-bus`, `auth-boundaries`).
   - A focused per-slice prompt that names the slice question, the primary source paths/URLs to read, and any cross-references the dispatched agent should use.
4. Briefly tell the operator the slice list (one line per slice: `NN – slug – focus`) before dispatching. Do not ask for approval unless the slices look thin or overlap; just announce and proceed.

If the scout reveals that fewer than `--agents` non-overlapping slices exist, reduce the dispatch count and tell the operator. Do not pad slices to hit the requested count.

**Step 4: Dispatch explorer agents in parallel**

Gist tripwires — the contract items the parent must enforce in every dispatched prompt:

- The prompt names three things: slice scope, slug+ordinal, exact target file path. If any is missing, the dispatched agent must refuse and ask back.
- The dispatched agent gets exactly one file-write — at the named target path — and nothing else. No edits, no `git`/`make`/package managers, no writes outside `<path>/analysis/`.
- All slices dispatch in parallel via `compozy exec`, with `--ide`/`--model`/`--reasoning-effort` forwarded from the operator's inputs. Wait for every process to exit (code 0) before verification.

**STOP. Read `references/dispatch-rules.md` in full before composing any slice prompt.** That file contains the complete scoped-write contract, tool allow/forbid lists, parent responsibilities, and failure handling. The bullets above are tripwires, not the contract — the contract must be embedded verbatim in every slice prompt.

From haryelramalho/skills, MIT.

More from haryelramalho/skills