Known Agent Issues

This page tracks current workflow issues by code agent. Issues are grouped by agent below.

Claude Code

Medium-effort models can miss workflow steps

In strict multi-step aitask-* workflows, medium-effort models may skip required checkpoints or finalization steps.

Use stronger reasoning/model settings when you need reliable workflow compliance. If the agent stops mid-workflow before a final step (for example, the satisfaction-rating prompt), nudge it to “continue the workflow” or “finish the workflow” to complete the remaining steps.

Codex CLI

Interactive checkpoints

ait setup enables the default_mode_request_user_input feature in the generated Codex config (.codex/config.toml), so request_user_input is available in Codex’s default mode. This makes the interactive workflow checkpoints — task confirmation, plan approval, and commit review — available during the aitask-* workflow. Availability is necessary but not sufficient for reliable compliance: see Reasoning effort and workflow compliance below.

ait codeagent invoke launches every Codex skill in Codex’s default mode.

Composer animation and idle detection

With some models (currently the Astra family, such as gpt-6-astra), Codex animates a field of twinkling dots around the idle input box. The dots are visible characters, so ait monitor and ait minimonitor would never see such a pane go idle, and a shadow agent would never read as ready for an automatic recheck. The framework therefore turns off that one effect, and only that one:

• ait codeagent invoke and ait skillrun pass -c tui.whimsy=false -c tui.effects.starfield=false to every Codex launch, including shadow agents and restored sessions. Both keys are passed because older Codex versions recognize only tui.whimsy and newer ones use tui.effects.starfield. A command-line override takes precedence over every config file, so framework-launched Codex panes never draw the dots.
• The same launches also pass -c tui.animations=true. tui.animations is Codex’s switch for all of its TUI animations: with it off, the Working (Ns • esc to interrupt) timer stops updating while the agent works, so the monitor would read a working Codex pane as idle.
• ait setup adds whimsy = false under [tui] and starfield = false under [tui.effects] to .codex/config.toml, each only when the project config does not already set it, so an explicit project value is kept. The two keys mean the same thing, so if your project sets one of them (for example whimsy = true to keep the dots), setup gives the other one the same value. Codex reads a project’s .codex/config.toml only when the project is trusted.
• If .codex/config.toml sets animations = false under [tui], has no whimsy key, and has no [tui.effects] table (not even one for an unrelated effect such as welcome), ait setup removes that line and prints a notice, but only in a run where it also adds the two keys above. Because of that, this happens at most once: an animations = false you add back afterwards is kept.

If you start Codex by hand outside a trusted project, pass -c tui.whimsy=false -c tui.effects.starfield=false, or set the same keys in ~/.codex/config.toml. Avoid animations = false unless you accept that the working timer stops updating. A desktop “reduce motion” accessibility setting has the same effect on Codex, so the monitor can misread working Codex panes on a host that has it on.

Scrollback in Codex panes

Recent Codex versions draw their interface on the terminal’s alternate screen, where tmux keeps no scrollback. A pane in that mode loses its earlier output for tmux copy mode and for shadow-agent and monitor captures. ait codeagent invoke and ait skillrun therefore pass -c tui.alternate_screen=never to every Codex launch, including shadow agents and restored sessions, so framework-launched Codex panes run inline and keep their scrollback. If you start Codex by hand and want the same behavior, pass the same option.

Reasoning effort and workflow compliance

Set Codex’s reasoning effort to at least high when running the aitask-* workflow. At lower effort, Codex may silently skip required non-skippable workflow steps and treat the archive as the end of the workflow even when the interactive prompts are available. Raising the effort to high resolves most of these compliance problems.

Some skipped prompts are not compliance failures: execution profiles such as fast deliberately pre-answer prompts like task confirmation, email, and worktree creation. Those skips are expected — the caveat here is only about steps no profile pre-answers.

Workaround: even at high effort, Codex may occasionally stop mid-workflow before a final step (for example, the satisfaction-rating prompt). If it does, prompt it to “continue the workflow” or “finish the workflow” so it completes the remaining steps. This is not unique to Codex (see Medium-effort models can miss workflow steps).

Model self-identification is unreliable

Codex CLI models cannot reliably self-report their model ID when prompted. The framework falls back to reading the configured model from ~/.codex/config.toml, but this may not reflect the actual model if it was overridden at invocation time via CLI flags.

Workaround: Launch Codex CLI via ait codeagent invoke instead of calling codex directly. The wrapper sets the AITASK_AGENT_STRING environment variable with the correct agent string, ensuring accurate implemented_with metadata.

OpenCode

Plan mode may skip task locking

When OpenCode runs in plan mode, interactive skills (aitask-pick, aitask-explore, aitask-review, aitask-fold) may skip the task locking step because plan mode restricts the agent to read-only tools.

Recommendation: Use OpenCode in regular mode (not plan mode) for interactive skills that acquire task locks. These skills have their own internal planning phases.

Shallow implementation plans

OpenCode may produce high-level overviews instead of detailed step-by-step implementation plans during the task-workflow planning phase. The opencode_planmode_prereqs.md file contains explicit instructions to mitigate this, but results may vary by model.

Workaround: If the agent produces a shallow plan, prompt it directly: “Please make a detailed plan of which files will be edited with which changes.” This usually triggers the agent to expand the plan with specific file paths, exact modifications, and code snippets.

References

• Codex CLI docs: Codex CLI overview and approval modes
• aitasks Codex mapping note: .agents/skills/codex_tool_mapping.md
• OpenCode plan mode prereqs: .opencode/skills/opencode_planmode_prereqs.md
• ait codeagent — unified agent wrapper with AITASK_AGENT_STRING support

Next: Git Remotes