agentpprof 0.2.32

pprof-style semantic profiler for local AI coding-agent sessions
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145

# agentpprof

`agentpprof` turns local AI coding-agent sessions into pprof-compatible semantic
profiles. It reads Codex and Claude Code JSONL history through AgentSight's
`agent-session` crate, assigns one-word tags to sessions, prompts, and LLM
calls, and writes one explicit output file.

The profiles are not CPU profiles. They are projections over agent activity:
tool events, file effects, network effects, or token usage.

## Install

```bash
cargo install agentpprof
```

From this repository:

```bash
cargo run --manifest-path agentpprof/Cargo.toml -- -o agent.pb.gz
```

## pprof Output

Generate a semantic task profile for the current repository:

```bash
agentpprof --project-root . -o agent.pb.gz
```

Open it with standard Go pprof:

```bash
go tool pprof -top agent.pb.gz
go tool pprof -http=:0 agent.pb.gz
```

The default `tasks` view makes prompt tags the pprof leaf frame, so `pprof -top`
shows where the agent spent most of its session activity semantically.
Folded, SVG, and JSON outputs keep the full context-first task stack for
drilldown.

## Views

Use `--view` to choose the projection:

```bash
agentpprof -o tasks.pb.gz --view tasks
agentpprof -o system.pb.gz --view system
agentpprof -o tools.pb.gz --view tools
agentpprof -o tokens.pb.gz --view tokens
agentpprof -o files.pb.gz --view files
agentpprof -o network.pb.gz --view network
```

Widths mean different things by view:

- `tasks`: event count across tool and LLM-call activity.
- `system`: system-effect count, including tool category, process chain,
  effect, path/domain, and status frames.
- `tools`: compatibility alias for the system-effect projection.
- `tokens`: token count when reported by the agent log; otherwise bounded text
  estimates. Very large unsafe estimates are recorded as `unknown=1` so one
  replayed transcript cannot dominate the profile with bogus token width.
- `files`: file/path effect count.
- `network`: network/domain effect count.

## Other Formats

The default format is pprof protobuf, gzipped when the output path ends in
`.gz`. The output extension also selects common formats:

```bash
agentpprof -o tasks.folded --view tasks
agentpprof -o tokens.svg --view tokens
agentpprof -o files.json --view files
```

Folded stacks are compatible with common flamegraph tooling. SVG output is a
single prefix-merged flamegraph built from the folded stacks. JSON output
includes redacted session summaries and the stack table. Passing
`--include-previews` writes prompt, command, and LLM-output previews into JSON;
avoid it for public artifacts unless the source sessions are already sanitized.
Path frames outside the selected project root are grouped into stable
`external/*` buckets so home-directory names are not emitted in public
profiles.
See `../docs/flamegraph-example/` for a flamegraph gallery and view-by-view usage
examples.

## Tags

The default tagger is deterministic:

```bash
agentpprof -o agent.pb.gz --tagger regex
```

Add project-specific deterministic rules with repeated `--tag-rule`
arguments. Rules use `KIND:TAG=REGEX`, are tried in command-line order before
the built-in rules, and support `session`, `prompt`, `llm`, or `all` as
`KIND`:

```bash
agentpprof -o tasks.svg \
  --tagger regex \
  --tag-rule prompt:review='(?i)review|diff|regression' \
  --tag-rule prompt:test='(?i)cargo test|pytest|unit test'
```

For model-produced one-word tags, run a llama.cpp-compatible server and use:

```bash
llama-server -m /path/to/model.gguf --port 8080
agentpprof -o agent.pb.gz --tagger llm --llama-url http://127.0.0.1:8080
```

LLM tags are cached under the user cache directory by default, for example
`$XDG_CACHE_HOME/agentpprof/tags.json`. Override with `--cache`, or pass
`--no-cache` to avoid saving new entries.

## Selecting Sessions

By default, `agentpprof` scans recent local Codex and Claude Code sessions that
match `--project-root`.
Those logs can contain prompts, paths, model outputs, and tool results. For
repeatable private investigations, use explicit `--session-file` inputs.

Useful selectors:

```bash
agentpprof -o agent.pb.gz --session-file ~/.codex/sessions/.../session.jsonl
agentpprof -o agent.pb.gz --agent codex
agentpprof -o agent.pb.gz --session-id 019ec5
agentpprof -o agent.pb.gz --session-tag profile
agentpprof -o agent.pb.gz --prompt-tag review
```

No output directory is created unless the explicit `-o/--output` path contains
one.

## Development

```bash
cargo test --manifest-path agentpprof/Cargo.toml
```