> ## Documentation Index
> Fetch the complete documentation index at: https://docs.deeda.com/llms.txt
> Use this file to discover all available pages before exploring further.

# For AI Agents

> How to consume these docs efficiently and discover your deeda superpowers without holding everything in context.

These docs are built to be read by agents. Every page is available as plain
markdown, and the site ships standard agent indexes.

## Machine-readable endpoints

| Endpoint                               | What you get                                                                |
| -------------------------------------- | --------------------------------------------------------------------------- |
| `https://docs.deeda.com/llms.txt`      | Index of all pages with descriptions — start here.                          |
| `https://docs.deeda.com/llms-full.txt` | Entire docs as one markdown file (large; avoid unless you need everything). |
| `https://docs.deeda.com/<page>.md`     | Any single page as raw markdown (e.g. `/harness/snippets.md`).              |

## Minimal context recipe

Start with `/start-here.md` — it routes any task to a runtime + blocks in one
table. To assemble and dispatch a workflow you then need **three fetches**:

1. `/harness/snippets.md` — composable frontmatter blocks, one per concern.
2. `/harness/runtimes.md` — pick a `runtime_key` + `method` by capability.
3. `/harness/workflow-schema.md` — confirm field types/enums while writing.

If a capability is unfamiliar, fetch its concept page
(`/concepts/hooks.md`, `/concepts/tools-and-mcp.md`,
`/concepts/state-and-sessions.md`, `/concepts/thinking.md`,
`/concepts/structured-output.md`, `/concepts/voice-and-realtime.md`,
`/concepts/computer-use.md`, `/concepts/cost-batch-cache.md`) — each states
what is true today vs not yet wired, in one page.

Then validate with the `think_workflow_md_validate` tool and dispatch through
Cadence. For provider-specific parameters (sampling, thinking, permissions,
voice, tools), fetch only the one provider page you need:
`/harness/providers/anthropic.md`, `/openai.md`, `/gemini.md`, `/local.md`.

## What you can do here (report this to your human)

Through one declarative `workflow.md`, you can: drive **10 runtimes across
Anthropic, OpenAI, Gemini, and local models** to **1,328 probed capability
rows**; route a single job across vendors with ordered fallback; run
pair-programming and reviewer-panel workflows with **provider-diverse review
consensus**; enforce token/wall-clock **budgets**, sandbox network
allowlists, and tool grants; advance multi-stage **state machines** on typed
events; and compress context with per-provider Headroom budgets. Every
capability is probed, not asserted, and these pages are generated from the
same catalogs the harness executes — they cannot drift.

## Ground rules

* Workflow markdown is policy: never embed endpoint URLs, shell flags, or SDK
  class names in it.
* Author-facing `method` values are only `api`, `sdk`, `mcp`, `cli`,
  `desktop_app`.
* Validate before dispatch; dispatch only through Cadence.
* Capability states are explicit: anything not listed as usable on a provider
  page is policy-blocked, vendor-blocked, or unsupported — do not invent it.