Skip to content

Introduce AI agent auto-capture via shell integration #3

@melonattacker

Description

@melonattacker

Motivation

Users currently need to explicitly wrap AI agent commands:

logira run -- claude
logira run -- codex

This is easy to forget.

For logira's primary use case — local auditing of AI coding agents — users should be able to keep using familiar commands such as claude and codex while logira captures runs automatically.

Goal

Introduce a lightweight auto-capture mechanism for common AI coding agent commands.

For v0.3.0, prefer shell integration or shim-based wrapping over fully transparent daemon-based process discovery.

Proposed UX

logira init shell --targets claude,codex

After installation, the user can run:

claude
codex

and logira will internally run:

logira run --tool claude -- claude ...
logira run --tool codex -- codex ...

Possible command design

Option A:

logira init shell --targets claude,codex
logira shell status
logira shell uninstall

Option B:

logira watch install --target claude
logira watch install --target codex
logira watch status
logira watch uninstall

The exact command naming can be decided during implementation.

Scope

The first version should support:

  • bash
  • zsh
  • wrapping configured target commands
  • preserving original args
  • preserving original exit code
  • avoiding recursive wrapping
  • idempotent install
  • easy uninstall
  • status/debug output

Non-goals

This issue should not implement fully transparent process monitoring.

The following should be deferred:

  • discovering already-running processes
  • monitoring arbitrary binaries without shell integration
  • system-wide process watch
  • team/fleet monitoring
  • cloud upload
  • hosted UI

A future experimental feature can explore:

logira watch --target claude --experimental

Acceptance criteria

  • user can configure auto-capture for claude
  • user can configure auto-capture for codex
  • running the target command creates a normal logira run
  • original command args are preserved
  • original command exit code is preserved
  • repeated installation is idempotent
  • uninstall removes the integration cleanly
  • status command shows whether integration is active
  • documentation explains how to install, verify, and uninstall

Notes

This issue intentionally narrows the original watch-mode idea into a practical v0.3.0 MVP.

The larger transparent watch mode remains a future direction.

Metadata

Metadata

Assignees

No one assigned

    Labels

    area:shellShell integration, shims, and command wrappingpriority:mustRequired for the target milestonetype:featureNew functionality or user-visible behavior change

    Projects

    No projects

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions