MiniScira

Instructions, tools, skills, and the researcher subagent. Everything the model can see or do.

Everything under agent/ is compiled by eve, not by Next.js. The directory layout is the configuration: eve discovers tools, skills, instructions, and subagents from the filesystem.

The definition

agent/agent.ts is small on purpose. It sets the model, the compaction policy, and the spend limits. Nothing else.

agent/agent.ts

export default defineAgent({
  model: defineDynamic({
    fallback: DEFAULT_CHAT_MODEL,
    events: {
      "step.started": async (_event, ctx) => {
        const modelId = chatModelFromMessages(ctx.messages) ?? DEFAULT_CHAT_MODEL
        const credential = await turnCredential(ctx)
        // No credential means nobody to bill: fail here rather than fall
        // through to `fallback` on the deployment's own key.
        if (!credential) throw new NoGatewayCredentialError()
        return {
          model: createGateway({ apiKey: credential.apiKey })(modelId),
          modelContextWindowTokens: await contextWindowFor(
            modelId,
            credential.apiKey
          ),
        }
      },
    },
  }),
  // …compaction, limits
})

Two non-obvious details

The resolver runs on step.started, not turn.started. At turn.started ctx.messages is still empty, so there is no client context to read and every turn falls through to the compiled model.

fallback is a required field and is the compiled build-time model. It is what subagents and headless lookout runs use, since neither carries a picker selection. It is not a fallback in the behavioural sense.

The composer sends clientContext.chatModel with each message and the resolver reads the last marker in the transcript. The featured list and the compiled default live in lib/models.ts; see Models.

Cost and lifetime guards

Limit	Value	Why
`maxOutputTokensPerSession`	200K	A runaway session pauses with eve's Continue/Stop prompt.
`maxInputTokensPerSession`	10M	A deep-research turn re-reads a growing transcript; input dominates the bill. eve's default of 40M is ~160 such turns.
`sessionTimeoutMs`	180 days	eve 0.28 started completing sessions after 30 days. A research chat is exactly the thing someone reopens months later.
`compaction.model`	`google/gemini-3.5-flash`	Summarize on something cheap with a big window, not at flagship rates.

Instructions

agent/instructions/ is concatenated in filename order. Files can be static Markdown or a .ts module that returns a string, which is how per-user context gets injected:

File	Contributes
`00-core.md`	The standing system prompt: role, research method, citation rules.
`05-now.ts`	The current date and time.
`10-memories.ts`	The user's saved memories.
`20-personalization.ts`	Tone and personalization settings.

See Editing instructions.

Tools

Each file in agent/tools/ is one tool, named after the file.

Group	Tools
Web research	`exa_search`, `firecrawl_search`, `firecrawl_scrape`, `firecrawl_map`, `x_search`, `reddit_search`
Documents	`search_documents`
Memory	`remember`, `list_memories`, `forget`
MCP	`mcp_servers`, `mcp_list_tools`, `mcp_call`
Output	`artifact`, `generate_image`
Compute	`run_code`, `sleep`

Tools whose provider key is unset degrade rather than crash. The agent simply has fewer options that turn.

The disabled defaults

agent/tools/web_search.ts and agent/tools/web_fetch.ts both contain a single line:

agent/tools/web_search.ts

export default disableTool()

eve ships built-in web_search and web_fetch tools. They were turned off deliberately: the agent otherwise reached for the generic pair instead of the provider tools that produce the structured results the timeline and citation pills render from.

Skills

agent/skills/*.md are on-demand playbooks. The agent sees only their names and descriptions until it calls load_skill, at which point the full body enters context. That keeps a long playbook out of every prompt.

Skill	For
`deep_research.md`	Multi-source reports with a plan and a synthesis pass.
`document_research.md`	Questions answered from uploaded documents.
`fact_check.md`	Verifying a specific claim.
`compare_options.md`	Structured comparisons.
`news_brief.md`	Recent-events summaries.
`social_pulse.md`	X and Reddit sentiment.
`build_ui.md`	Interactive artifacts.

The researcher subagent

agent/subagents/researcher/ is a full agent in its own right, with its own instructions and its own copy of the search tools. The main agent delegates a scoped research question to it and gets back a summary rather than the raw transcript. That is the whole point: the subagent burns its own context window on twenty search results, and only the conclusion comes back.

In the UI, a delegated run renders as a nested timeline branching off the main rail.

Subagents carry no client context, so they always run on DEFAULT_CHAT_MODEL, never the model showing in the composer.

The sandbox

agent/sandbox.ts backs run_code with a Vercel Sandbox. bootstrap installs pandas/numpy/matplotlib once with egress open and reuses the snapshot; onSession then sets a deny-all network policy so user code cannot call out.

The template prewarm is eager

Per-session sandboxes open lazily, but eve start initializes the template before it begins listening and refuses to serve if that fails. On Vercel this is transparent. Anywhere else the agent will not start without working sandbox credentials, whether or not run_code is ever called. See Self-hosting.

The agent

On this page