Work with Coding Agents

Tascarrel supports three coding-agent harnesses:

Tasci is bundled with Tascarrel and connects to configured OpenAI-compatible Chat Completions endpoints.
Codex and Claude Code are installed and authenticated for each workspace.

The harness and its tools run inside the task pod. The model endpoint may run on your host, elsewhere on your network, or at a hosted provider.

Understand the Permission Model

Tascarrel uses isolation, not per-action approval, as its agent safety model. Codex and Claude Code start without command or file approval prompts, and Tasci’s built-in tools likewise run without per-action confirmation. The workspace VM and its declared access policies limit what an agent can reach on the host and network.

Tasci resolves relative file-tool paths from /workspace. Its read, edit, and write tools also accept absolute paths and paths containing parent components, so they can access any UTF-8 file permitted to the Tasci process inside the task pod. The pod and workspace VM remain the isolation boundary.

That safety boundary makes autonomy practical. Agents keep working without permission prompts while the UI tracks their activity and surfaces decisions, failures, and completed turns. You give up some moment-to-moment control, but you can leave agents working and return when attention is needed to review changes before publishing.

Configure a Harness

Choose one of the supported harnesses:

For Tasci, open Workspace → Settings → Tasci, add an API endpoint and model, and select a default model. Tasci is already installed.
For Codex, install the pinned harness and choose Sign in with ChatGPT. Complete the displayed device-code flow under Settings → Harnesses.
For Claude Code, install the pinned harness, run claude setup-token separately, and enter the resulting token under Settings → Harnesses.

Use different workspaces when projects should not share agent credentials.

Tasci endpoint authorization contains only a placeholder-bearing header template. Configure a matching host-side HTTP secret-injection rule so the credential is inserted after the request leaves the workspace. The endpoint must also be reachable under the workspace’s network policy.

Add MCP Tools to Coding Harnesses

The Model Context Protocol (MCP) lets an agent call tools provided by an external service. Add Streamable HTTP servers under Workspace → Settings → MCP. The catalog is shared by default, or a server can select only specific harnesses:

{
  "chat": {
    "mcpServers": {
      "exa": {
        "displayName": "Exa",
        "endpoint": "https://mcp.exa.ai/mcp",
        "harnesses": ["Tasci", "Codex", "ClaudeCode"]
      }
    }
  }
}

Omit harnesses to select all three harnesses. Tascarrel supplies the selected servers whenever it starts or attaches a new harness session. Tasci receives the servers in its native harness configuration, Codex receives mcp_servers.<name> process overrides, and Claude Code receives an additive --mcp-config without strict mode. Native MCP sources that Codex and Claude Code already load are not disabled.

The harness discovers every tool advertised by a configured server. Tasci prefixes model-visible names with the server’s settings name, so Exa’s web_search_exa tool becomes mcp__exa__web_search_exa. Codex and Claude Code retain their native tool naming. Enabling a server therefore trusts all tools and tool descriptions that it advertises. Tool arguments, including search queries and URLs, leave the workspace and are sent to that server. Tasci accepts only text results and limits how much result text enters the model context.

Each server may define arbitrary HTTP header templates:

{
  "headers": {
    "Authorization": "Bearer tascarrel-secret:mcp-token",
    "X-Workspace": "development"
  }
}

Configure matching host-side secret-injection rules for placeholder-bearing values. Set each rule’s paths to the endpoint path, such as ["/mcp"], so the credential cannot be injected into another endpoint on the same host. The workspace network policy must also permit access to the endpoint. If one server is unavailable, startup and warning behavior follows the selected harness’s native MCP client.

Start and Resume Chats

Open a pod’s Agent view, choose a harness and model, and send a prompt. Chats retain messages, attachments, status, and harness metadata even when the harness process stops. Navigating away does not terminate an active turn.

Each harness can resume its model conversation after a harness or workspace restart. Codex and Claude Code use their own session stores. Tasci keeps a versioned native session journal under ~/.tasci/sessions in the associated pod. The shared chat database stores only an opaque Tasci resume cursor, not Tasci’s messages or compaction records.

The Tasci journal belongs to the pod that owns the chat. Removing the pod or its journal prevents that chat from resuming its native model context.

A pod can have multiple chats. Interrupt asks an active harness to stop cooperatively. Archiving removes a completed chat from the active list; the web UI cannot currently restore it.

Deliver a Prompt at the Right Time

When a turn is active, the composer offers:

Immediate steers the active turn when supported. Tasci does not currently support steering.
When Idle queues a separate follow-up until the turn finishes.
Interrupt and Send cancels the active turn before sending the new prompt when the harness supports that operation. Tasci does not yet support it.

Prompt queues are runtime state and do not survive an agent-engine restart. Interruption is cooperative, so child processes may not stop immediately.

Choose Models and Read Usage

Codex and Claude Code supply their available models and options. Set their workspace defaults and ordering under Settings → Harnesses.

Tasci models are aliases that map to provider-native model identifiers and endpoints under Settings → Tasci. One endpoint can serve several models, and the model can be changed between turns. Tascarrel stores both kinds of settings in settings.json.

Tasci preserves model-step order in the chat timeline. Text emitted before a tool call, the tool activity, and the response after the tool appear as separate consecutive items.

Reasoning-capable OpenAI-compatible endpoints may stream reasoning_content. Tasci presents that content as a reasoning item and retains it in assistant history across tool steps. The llama-server default reasoning mode is auto, allowing the model chat template to enable reasoning.

Token usage and cost estimates appear only when the harness reports enough data. Tasci can calculate costs from optional per-model pricing in settings.json. Provider billing remains authoritative.

The chat status bar distinguishes cumulative Tokens from the model’s current Context. Context appears as the occupied token count and, when known, the effective capacity, such as 42k / 200k. A leading ~ marks a Tasci estimate. N/A means that the harness has not supplied a current observation, including after compaction and before the next model response. Codex reports context usage directly, Claude Code reports it with primary assistant usage, and Tasci estimates any context that its provider usage does not cover.

To track work for different clients or internal projects, add cost centers under Workspace → Settings → Usage and optionally select a default for new chats. The chat composer shows the assignment for a new chat, and an existing chat’s assignment can be changed beside its status. Reassigning a chat moves its whole recorded history to the new cost center.

The monthly report on the Usage settings page includes active and archived chats across all pods in the workspace. It keeps unassigned usage visible and shows when cost, provisional usage, or delegated-agent coverage is incomplete. Calculated costs use the pricing snapshot stored with each turn and are estimates rather than a replacement for provider billing.

Compact Long Tasci Chats

Tasci compacts model context when a configured contextWindow approaches its limit. It asks the selected model for a structured checkpoint, retains a verbatim suffix at a complete user or assistant boundary, and uses the checkpoint plus that suffix for later requests. A provider context-overflow response triggers the same process followed by one retry.

Compaction does not delete Tasci’s native conversation data. User, assistant, tool-result, and compaction records remain in the durable append-only session journal. Tasci commits each successful turn or compaction as one recoverable transaction before reporting completion. Only the context projected into subsequent model requests changes. The timeline reports successful and failed compaction attempts.

After compaction, the status bar replaces the previous context count with Tasci’s estimate for the compacted context. Other harnesses may briefly show N/A until they report their next context observation.

The chat compaction action can also request this process while Tasci is idle. There must be enough older context to summarize; a short chat has no useful compaction boundary.

Provide Workspace Guidance

Put Codex guidance that applies to every task in agents/AGENTS.md. Place compatible reusable skills below agents/skills/; Tascarrel mounts workspace skills read-only and pins them to the pod’s input generation. Tasci currently reads /workspace/AGENTS.md when starting a new model conversation. New workspaces also contain agents/CLAUDE.md as a relative symlink to AGENTS.md, so Claude Code and Codex receive the same workspace-level instructions.

Define prompt shortcuts in config.toml:

[chat.commands.prepare-review]
text = """
Inspect the current changes, run the smallest relevant checks, and summarize
anything that should block publication.
"""

Commands insert their text into the composer, where it can be adapted before sending.

Edit this page

PreviousBuild Pod Images NextUse Development Tools