agentdust 1.5.1 → 1.6.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +16 -4
- package/bin/agentdust +0 -0
- package/package.json +1 -1
package/README.md
CHANGED
|
@@ -4,7 +4,7 @@
|
|
|
4
4
|
|
|
5
5
|
AgentDust is a macOS tool that finds processes left running after Claude Code and Codex sessions. Manual cleanup requires your typed approval. Version 1.1.0 adds opt-in automatic cleanup for proven leftovers in directories you enable. It also reports Claude Code disk usage without deleting files. Codex tracking uses native hooks and the shell session marker. Automatic cleanup waits for every recorded Codex host process to exit, including a shared desktop server. Cursor support is deferred.
|
|
6
6
|
|
|
7
|
-
**AgentDust 1.
|
|
7
|
+
**AgentDust 1.6.0 supports Claude Code and Codex process tracking on Apple silicon, with local cleanup outcome metrics. Automatic cleanup results include their record time. Codex desktop cleanup waits for its shared host to exit. Captured fixtures cover the listed concurrent session, subagent and lifecycle scenarios on Claude Code 2.1.292, including an abrupt CLI exit. Agent Teams remain outside the support claim. The installed 1.0.0 Homebrew binary passed controlled human decline and typed approval checks on the maintainer's Mac ([release evidence](docs/v1/readiness.md)).** See [install and connect](#install-and-connect). To try the non-destructive approval flow, run the [development probe](#running-the-development-probe).
|
|
8
8
|
|
|
9
9
|
Agents start dev servers, MCP servers and helpers. When a session ends or crashes, some of them keep running under `launchd`, holding memory, ports and sometimes CPU. The upstream reports are open: [anthropics/claude-code#1935](https://github.com/anthropics/claude-code/issues/1935) and [openai/codex#21008](https://github.com/openai/codex/issues/21008).
|
|
10
10
|
|
|
@@ -12,13 +12,14 @@ Security and privacy: [SECURITY.md](SECURITY.md), the [threat model](docs/threat
|
|
|
12
12
|
|
|
13
13
|
## Status
|
|
14
14
|
|
|
15
|
-
AgentDust 1.
|
|
15
|
+
AgentDust 1.6.0 provides process analysis, opt-in automatic cleanup, approved manual cleanup and local cleanup outcome metrics for Claude Code and Codex, plus a read-only Claude Code disk report. Automatic cleanup results include the time they were recorded. Codex support retains the exact host-exit requirement; it does not stop helpers immediately when a desktop chat ends. Cursor remains deferred. The milestones are in [ROADMAP.md](ROADMAP.md).
|
|
16
16
|
|
|
17
17
|
Releases follow [Semantic Versioning](docs/release.md#semantic-versioning): compatible fixes increment the patch version, compatible features increment the minor version, and incompatible public API changes increment the major version.
|
|
18
18
|
|
|
19
19
|
| Part | State |
|
|
20
20
|
| --- | --- |
|
|
21
21
|
| TCP port diagnosis and recovery | `agentdust port PORT [--resolve] [--json]` and `agentdust_port`. Recovery uses the existing automatic policy ([scope and limits](docs/port-conflicts.md)) |
|
|
22
|
+
| Local cleanup outcome metrics | `agentdust metrics [--json]` and `agentdust_metrics`; private local history with size based retention ([limits](docs/automatic-cleanup.md#cleanup-metrics)) |
|
|
22
23
|
| Opt-in automatic cleanup | Available since `1.1.0`. Directory policy, owner-exit worker and read-only agent results with per-result record times ([commands and limits](docs/automatic-cleanup.md)) |
|
|
23
24
|
| Process identity, named session markers read from another process, `KERN_PROCARGS2` parser (fuzzed) | Built |
|
|
24
25
|
| Claude Code and Codex hooks that record session and shell events in a local journal | Built |
|
|
@@ -70,6 +71,17 @@ In Claude Code, ask “Use AgentDust to show where Claude Code uses disk space
|
|
|
70
71
|
2. A plan keeps only processes owned by an ended session and suspects, at most 10 per call. You approve by typing a 4-character code that AgentDust generates: one code for the whole batch of owned processes, and one code per suspect with its evidence. The form is in your agent, or in the terminal with `agentdust apply`.
|
|
71
72
|
3. Before signalling each process, AgentDust classifies it again and compares boot session, PID, start time, user and executable path. It sends `SIGTERM` to that one PID and nothing else.
|
|
72
73
|
|
|
74
|
+
## Cleanup metrics
|
|
75
|
+
|
|
76
|
+
```bash
|
|
77
|
+
agentdust metrics
|
|
78
|
+
agentdust metrics --json
|
|
79
|
+
```
|
|
80
|
+
|
|
81
|
+
AgentDust 1.6.0 records manual and automatic executor outcomes, plus automatic policy skips, in a private local JSONL file. The JSON summary includes totals by result and cleanup mode, classifier counts, and daily buckets for charts. The read-only `agentdust_metrics` MCP tool returns the same summary to agents. Calls rejected before an item reaches the executor are not included. Counts are outcomes, not unique processes; repeated attempts may add more rows. Records contain only the time, mode, class and result. They do not contain PIDs, process names, paths, commands, session IDs or project names. AgentDust does not upload metrics. An MCP host may include the returned summary in requests to its model service, and timestamps and counts can reveal activity patterns. The log keeps its current 5 MiB file and one rotated file. Retention is based on file size, with no age-based expiration or guaranteed history window. Metrics history starts with the first cleanup result recorded by a build that supports this feature. If a metrics write fails, AgentDust prints a sanitized warning to stderr and continues the cleanup result.
|
|
82
|
+
|
|
83
|
+
`terminated` means the selected process disappeared during AgentDust's exit check. It does not estimate time, CPU or memory saved. See [metrics data and limits](docs/automatic-cleanup.md#cleanup-metrics).
|
|
84
|
+
|
|
73
85
|
The design is in [docs/superpowers/specs](docs/superpowers/specs).
|
|
74
86
|
|
|
75
87
|
## Why I built AgentDust
|
|
@@ -100,9 +112,9 @@ Cleanup controls:
|
|
|
100
112
|
|
|
101
113
|
Report a vulnerability privately as described in [SECURITY.md](SECURITY.md).
|
|
102
114
|
|
|
103
|
-
## Using AgentDust 1.
|
|
115
|
+
## Using AgentDust 1.6.0
|
|
104
116
|
|
|
105
|
-
Release 1.
|
|
117
|
+
Release 1.6.0 supports Apple silicon with Claude Code. Homebrew and npm distribute the native binary. Automatic cleanup status includes per-result record times, and cleanup metrics summarize recorded manual and automatic outcomes. Codex process tracking and automatic cleanup wait for every recorded host to exit. Codex desktop cleanup waits while a shared server remains alive. Automatic cleanup passed isolated normal-exit, abrupt-exit and worker-restart checks. The 1.0.0 Homebrew installation, setup, live session discovery and controlled human cleanup acceptance passed on the maintainer's Mac. The listed M6 lifecycle and attribution scenarios are captured and replayed for Claude Code 2.1.292. Agent Teams remain outside the support claim. See the [M6 fixture matrix](docs/m6/fixtures.md) for the tested scenarios and their limits. Use `agentdust metrics` for totals or `agentdust metrics --json` for daily buckets; agents can read the same summary with `agentdust_metrics` ([scope and limits](#cleanup-metrics)).
|
|
106
118
|
|
|
107
119
|
### Install and connect
|
|
108
120
|
|
package/bin/agentdust
CHANGED
|
Binary file
|