Skip to content

The agent app asks before it acts

From v0.123.2, an agent task on Codex, Claude Code or Gemini CLI is held to the workflow's permits as far as the seat allows, and pauses beyond them.

A Nika workflow says what it may touch in its permits: block: files, hosts, programs and tools. nika check judges the file against that block before anything runs, and the run judges each effect again.

That model has a weak spot when a task is handed to an agent application. An agent: task can run on Codex, Claude Code or Gemini CLI, an application you are already signed into. The engine's guide calls this a harness seat. The application then runs its own tools: its file edits, its shell, its fetches. Nika does not execute them, so it cannot judge each one the way it judges an exec: task or a built-in nika:* tool.

The commit that fixed it says that before the change, a Codex seat "was never held to permits:." The 0.123.2 release notes describe the result: "On codex-acp the session kept the adapter's default agent mode, whose own reviewer approves most actions, so a Codex task wrote ./out and read files outside the project without a single permit_checked frame." A permit_checked frame is the line Nika writes in the run's journal for each permission decision. None was written: the adapter's own reviewer did the approving.

Hold the session, judge every ask #

v0.123.2 changes how Nika opens such a session. The harness guide describes two levers.

The first acts before the prompt. The session works in the project the run was admitted for, never in whatever directory the Nika process started from. Nika selects the application's measured asking mode, the mode in which it asks before acting, and reads that mode back from the application's own answer, or refuses to start. In the notes' words, it "never picks a mode that approves on its own". For Codex that mode is read-only. For Claude Code it is default, instead of a user's auto. For Gemini CLI it is default. If the application reports a mode change during the turn, the session ends. Nika also offers the application no file or terminal methods: a call to one gets an error, and the refusal is journaled.

The second lever acts on every permission the application asks for. Inside your permits: grants, the ask is allowed once, never "allow always". Outside them, the run pauses with the application's question and waits for a person. Three kinds of ask pause whatever the grants say, even under permits.exec: true:

  • network access, including Codex's SOCKS approvals, which the notes say "carry no URL and were answered allow_once as plain commands" before the fix;
  • a request to widen the application's own sandbox (Codex additionalPermissions, Claude Code dangerouslyDisableSandbox);
  • an approval for one of the application's own MCP tools, because permits.tools names your workflow's MCP tools, not the application's.

Paths are judged where the operating system will land. Each path is walked through every symlink and every ... The notes give the case that used to pass: "out/link/../x with out/link pointing outside the project was judged as out/x and allowed". A path that cannot be resolved now pauses.

Running on a seat #

A seat is declared in the file with run: { access: { via: codex, protocol: acp } } (the declaration told in The file says how it reaches its model), or chosen at launch. The harness guide of the documentation site (guides/harness-access.mdx in nika-docs) shows the launch form:

terminal
nika check flow.nika --access codex   # installed? signed in? fit for each task?
nika run flow.nika --access codex

On every agent: task a seat will run, nika check prints a [harness] hint that says what permits: does not cover on that seat. nika explain harness prints the same lesson.

When an ask falls outside your grants, the run pauses with the question in view. The guide gives the line that answers it, with false to refuse:

terminal
nika run --resume <trace> --answer <task>=true

The application is shut down at the pause. The resumed run starts the task again in a new session, and your answer decides its first ask outside the grants.

Every decision lands in the trace. Each harness task records one permit_checked frame for the session mode, with the decision confine when the measured mode was selected, then one frame per ask: allow, deny (your answer) or escalate (paused for you).

v0.124.0 closed the gap 0.123.2 named #

The 0.123.2 guide said that in read-only, Codex runs ordinary commands inside its own sandbox without asking, so a command it does ask about usually wants to run outside that sandbox, and that permits.exec: true admitted such a command "like any command". Its advice was to leave permits.exec unset.

v0.124.0, published the next day, removed the need for that advice. Its notes say Nika used to read such an ask "as execute · <prose>: under permits: { exec: true } it was allowed, and Codex fetched a site outside its sandbox." Such an ask now always pauses, with the command and Codex's reason in the question.

The same release tightened three more things:

  • The pause question is one line the application cannot steer: control characters become spaces and invisible ones (direction overrides, zero-width characters) are removed, so a title cannot erase itself on your terminal.
  • One --answer decides one ask for the task. An attempt that retry: replays never reuses it.
  • A turn Codex could not complete now fails its task, instead of succeeding with the error as its output.

Why it matters #

Before 0.123.2, a task handed to a Codex seat inherited whatever the adapter's default mode allowed, and the boundary written in the workflow did not reach it.

Now the workflow file stays the reference for what the application asks. Nika picks the mode that asks, judges each ask against the file, writes each decision down and stops for a person on any ask the file does not grant. Where it cannot see, nika check says so on every such task.

What it does not do #

The guarantee holds "as far as the seat allows", in the notes' own words. What the selected mode lets the application do without asking never reaches Nika, so permits: does not bound it. The guide lists it per seat:

  • Codex reads any file; writes the project, /tmp, $TMPDIR and its own ~/.codex; runs any program inside its own sandbox, and its tools through a JavaScript "code mode"; calls its MCP tools marked readOnlyHint: true; opens pages with its hosted web tool, "even under permits: {}"; starts every configured MCP server; applies your global ~/.codex/AGENTS.md and your Codex plugins. In the guide's words, "even with no fs.write grant, the project stays writable." The 0.123.2 guide already listed the reads, the writes to the project, /tmp and $TMPDIR, and the sandboxed programs; the rest was added with v0.124.0.
  • Claude Code reads files in the project, runs commands it classifies as read-only and fetches documentation hosts it pre-approves.
  • Gemini CLI reads files in the project, runs shell commands it judges safe reads and runs Google web searches. It only acknowledges the mode switch, and it reports a mode change only as message text, which Nika does not read.
  • On qwen-code, kimi-code, opencode, copilot and grok-build, no mode is measured, so none is selected, and Nika judges only what they ask.

On every seat, an action Nika allows runs under the application's own sandbox, not Nika's exec sandbox, so permits.fs and permits.net do not confine it. The application's own permission rules, hooks and MCP servers still act unless they ask. A failed turn is only detected when the application reports it; one without that report still returns its error as text.

The seat's program itself is a separate question. The 0.118.0 changelog section, a tag with no public release, said that "nothing yet hashes, pins or sandboxes the seat's binary, so an executable that impersonates a seat on PATH runs with the user's rights." The 0.124.2 changelog later says that, for seats whose identity is their initialize answer, an adapter that "answers as another program or sits outside its version pin" is unavailable. That judges what the adapter says about itself.

The guide's advice follows from all this: "If a task must not write anything, or must not read outside the project, do not rely on permits: alone on a harness seat: run it on an API or local model (the native agent loop judges every tool call), or in a container or a throwaway checkout."

Read the v0.123.2 release and the v0.124.0 release, or continue with The run that waits for you.