Skip to main content
Version: main

Status Dashboard

aide status shows a comprehensive view of AIDE's internal state.

Usage

aide status # Full dashboard
aide status --json # Machine-readable JSON

Output

The dashboard includes:

  • Version and project info -- aide version, project root, current mode
  • Server -- whether MCP/gRPC server is running, uptime
  • File watcher -- watched paths, directory count, debounce delay, pending files
  • Code index -- files, symbols, and references indexed
  • Findings analysers -- per-analyser status, last run time, finding counts by severity
  • MCP tools -- all tools with per-tool execution counts
  • Stores -- paths and sizes of all data files
  • Environment -- active AIDE_* environment variables

Status Line (Claude Code)

AIDE can display session info in Claude Code's status line, showing mode, duration, task counts, and token usage.

Add to .claude/settings.json:

{
"statusLine": {
"type": "command",
"command": "bun ~/.claude/bin/aide-hud.ts"
}
}

Example output:

[aide(0.0.40)] mode:idle | 12m | tasks:done(6) wip(0) todo(0) | 5h:115K ~<1m

Shows: version, current mode, session duration, task status, 5-hour token usage, and estimated time to rate limit.

Web dashboard: token evidence

Run aide dashboard (or aide dash) and open Tokens for a project. Session and date filters apply to the report.

Overview shows aide's contribution alongside recorded model usage:

  • Output reductions compare measured text before and after processing. Proposed rewrites remain separate because the host may ignore them.
  • Potential retrieval reductions compare recorded retrieval text with verified full-file references, counting each file version once per context window. Added text remains visible, and incomplete windows do not receive a savings estimate.
  • Aide operations and prepared context show work performed and context overhead.
  • Model counters include input, output, cached and uncached input, cache writes and reasoning where the host supplies them. Missing values remain unknown.

Token estimates are labelled. Cache and reasoning counters can be subsets of input and output; text comparisons can also overlap. The report does not add these values into a combined savings total or claim a billing reduction. Details holds individual operations and context-window evidence; Accounting explains measurement boundaries and coverage.

Recent events are filtered by session and date in the daemon before the result limit is applied. A failed load shows unknown event coverage with a retry action; it is not displayed as an empty history. Upgrade and restart both the daemon and dashboard together so date-filtered queries use compatible versions.

Context and subagents

Context windows are scoped by host, session and actor. Explicit subagent-start hooks in Claude Code and Codex initialize the child's own window. Replayed starts preserve existing state, including a pending compaction or newer reset. This requires a CLI and daemon that support atomic state initialization; older or unavailable components leave the child's context unknown.

Codex installations also need refreshed hook registration. In a development checkout, rerun ./aide-dev-toggle.sh dev; for an installed release, rerun aide-plugin install --platform codex. Restart the host and daemon after the upgrade. A rebuild alone does not add missing lifecycle hooks. Existing child actors without a recorded start remain unknown; test with a newly started actor.

Compaction hooks use the reported actor identity. OpenCode uses each session's identity, including child sessions, through creation and compaction. When a host does not supply the necessary lifecycle or identity evidence, the report retains that uncertainty. A provider cache expiry alone does not reset context evidence.

Historical observations are not reassigned after an upgrade. An unknown context can prevent comparisons in otherwise matching windows: event receipt time alone cannot prove which window produced a delayed result. These coverage limits do not establish that aide saved nothing; they mean a reduction is not established by the available records.