agent-trace-capture 0.0.0-stage → 0.4.0
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/COVERAGE.md +38 -0
- package/README.md +182 -2
- package/THIRD_PARTY_NOTICES.txt +1041 -0
- package/bin/cli.cjs +18 -0
- package/binaries/darwin-arm64/agent-trace-capture +0 -0
- package/binaries/darwin-x64/agent-trace-capture +0 -0
- package/binaries/linux-arm64/agent-trace-capture +0 -0
- package/binaries/linux-x64/agent-trace-capture +0 -0
- package/binaries/win32-arm64/agent-trace-capture.exe +0 -0
- package/binaries/win32-x64/agent-trace-capture.exe +0 -0
- package/package.json +12 -4
package/COVERAGE.md
ADDED
|
@@ -0,0 +1,38 @@
|
|
|
1
|
+
# Coding tool coverage
|
|
2
|
+
|
|
3
|
+
The npm package and standalone apps use the same native Go collector and detect existing histories for the 11 providers below. The original Python collector remains available for private deployments. Missing history is reported as zero candidate files. Capture retains raw histories; it does not convert every tool into a common event schema.
|
|
4
|
+
|
|
5
|
+
| CLI provider | Product | Capture | Validation |
|
|
6
|
+
| --- | --- | --- | --- |
|
|
7
|
+
| `claude` | Claude Code | Project JSONL and subagent JSONL | Synthetic archive/hash tests; local discovery |
|
|
8
|
+
| `codex` | Codex | Active/archived JSONL and JSONL.ZST | Synthetic archive/hash tests; local discovery |
|
|
9
|
+
| `cursor` | Cursor IDE | Agent transcripts; filtered SQLite chat rows | SQLite/WAL fixtures; local discovery |
|
|
10
|
+
| `opencode` | OpenCode | Allowlisted SQLite session/message/part tables; older JSON stores | Real OpenCode 1.19.5 free-model run in Docker; deployed Vercel/R2 flow; SQLite/WAL fixtures |
|
|
11
|
+
| `kimi` | Kimi Code CLI | Context and wire JSONL, including subagents | Primary-source layout check; synthetic files, exclusions, ZIP tests |
|
|
12
|
+
| `gemini` | Gemini CLI | Session JSON and JSONL under project chat directories | Primary-source layout check; synthetic old/new layouts, exclusions, ZIP tests |
|
|
13
|
+
| `copilot` | GitHub Copilot CLI | Session `events.jsonl` | Official storage docs; synthetic files, exclusions, ZIP tests |
|
|
14
|
+
| `cline` | Cline | API conversation, UI events, context history, task metadata | Primary-source filenames; synthetic editor/standalone layouts and exclusions |
|
|
15
|
+
| `roo` | Roo Code | Same allowlisted task histories in its extension storage | Primary-source filenames; synthetic editor/profile layouts and exclusions |
|
|
16
|
+
| `continue` | Continue | Saved session JSON files | Primary-source paths; synthetic files, exclusions, ZIP tests |
|
|
17
|
+
| `aider` | Aider | Project chat-history Markdown | Official option defaults; synthetic project selection/boundary and ZIP tests |
|
|
18
|
+
|
|
19
|
+
All adapters skip symlinks and preserve originals. macOS/Windows/Linux editor root calculations have fixture checks. This is not a claim of live testing each product/version on each OS. Aider captures selected projects only; alternate homes/profiles require flags. Histories can still contain secrets pasted into prompts or tool outputs, so excluding separate credential files does not make traces redacted.
|
|
20
|
+
|
|
21
|
+
Amp and Windsurf Cascade remain outside automatic capture. Amp documents cloud-backed thread export; this client does not access its account or export API. No verified local Cascade decoder is included. Collecting Cline/Roo extensions inside the Windsurf editor is distinct from capturing Cascade itself.
|
|
22
|
+
|
|
23
|
+
Sources checked during implementation:
|
|
24
|
+
|
|
25
|
+
- [Kimi data locations](https://github.com/MoonshotAI/kimi-cli/blob/main/docs/en/configuration/data-locations.md)
|
|
26
|
+
- [Gemini chat recording](https://github.com/google-gemini/gemini-cli/blob/main/packages/core/src/services/chatRecordingService.ts) and [storage paths](https://github.com/google-gemini/gemini-cli/blob/main/packages/core/src/config/storage.ts)
|
|
27
|
+
- [Copilot session data](https://docs.github.com/en/enterprise-cloud%40latest/copilot/concepts/agents/copilot-cli/chronicle)
|
|
28
|
+
- [Cline storage filenames](https://github.com/cline/cline/blob/main/apps/vscode/src/core/storage/disk.ts)
|
|
29
|
+
- [Roo storage filenames](https://github.com/RooCodeInc/Roo-Code/blob/main/src/shared/globalFileNames.ts)
|
|
30
|
+
- [Continue session paths](https://github.com/continuedev/continue/blob/main/core/util/paths.ts)
|
|
31
|
+
- [Aider history defaults](https://aider.chat/docs/config/aider_conf.html)
|
|
32
|
+
- [Amp thread export](https://ampcode.com/docs/threads)
|
|
33
|
+
|
|
34
|
+
## Distribution validation
|
|
35
|
+
|
|
36
|
+
The native collector has separate fixture tests for allowlisted file adapters, SQLite/WAL reads and credential-table exclusions, archive hashes, partial capture, output boundaries, installation IDs, and upload HTTP/checksum behavior. The npm launcher is tested without Python or shell access in PATH and has no install hooks or runtime npm dependencies. Release builds include fallback trusted certificate roots for minimal systems with no OS certificate bundle; normal OS certificate validation is preserved.
|
|
37
|
+
|
|
38
|
+
Executables are cross-built for Windows, macOS, and Linux on x64/ARM64. macOS ARM64 collection and Linux ARM64 collection/upload are runtime-tested; Linux x64 version execution is tested through Docker emulation. Windows and Intel macOS builds have not been run on native devices and are not signed/notarized with a trusted publisher identity.
|
package/README.md
CHANGED
|
@@ -1,3 +1,183 @@
|
|
|
1
|
-
#
|
|
1
|
+
# Agent Trace Capture
|
|
2
2
|
|
|
3
|
-
|
|
3
|
+
Capture histories from 11 coding tools with one command:
|
|
4
|
+
|
|
5
|
+
```sh
|
|
6
|
+
npx agent-trace-capture
|
|
7
|
+
```
|
|
8
|
+
|
|
9
|
+
Or install once and run whenever needed:
|
|
10
|
+
|
|
11
|
+
```sh
|
|
12
|
+
npm install -g agent-trace-capture
|
|
13
|
+
agent-trace-capture
|
|
14
|
+
```
|
|
15
|
+
|
|
16
|
+
Node.js 18+ is required for npm/npx. The package bundles native executables for Windows, macOS, and Linux on x64 and ARM64; it has no runtime npm dependencies or install hooks. It does not need Python and does not collect anything during installation. The command discovers existing histories, creates a ZIP on your device, then uploads it to `https://agent-trace-capture.vercel.app` using a short-lived upload URL. No bucket credentials, enrollment token, or user configuration are needed for the public service.
|
|
17
|
+
|
|
18
|
+
Without Node.js, download the standalone Windows executable, macOS app, or Linux executable from [the download page](https://agent-trace-capture.vercel.app/downloads). Windows/macOS builds are not signed with a trusted publisher certificate or notarized, so the OS may display a security prompt. Native Windows execution still requires testing on a Windows device; cross-compilation is not equivalent to runtime testing.
|
|
19
|
+
|
|
20
|
+
Histories are **not redacted** and can contain private prompts, source code, personal data, and secrets pasted into conversations. Uploads occur when you run the collector. Preview with `agent-trace-capture --dry-run` or save without uploading with `agent-trace-capture --local-only`. Each upload stays in a separate installation-ID folder. Aider histories are per project: run from a project or pass `--project-dir`.
|
|
21
|
+
|
|
22
|
+
See [COVERAGE.md](COVERAGE.md) for exact formats and validation limits. The Python client remains available for private/custom server deployments; it is not required by the npm or standalone releases.
|
|
23
|
+
|
|
24
|
+
## Vercel hosting
|
|
25
|
+
|
|
26
|
+
This repository includes a Flask entrypoint (`app.py`), Python version pin, and `vercel.json`. Configure the server environment variables below in your Vercel project's Production environment and deploy this folder using the Flask framework preset. `AGENT_TRACE_PUBLIC_URL` must be the stable production HTTPS URL. The application authenticates legacy installer downloads and private upload-ticket requests with its enrollment token. Set `AGENT_TRACE_PUBLIC_UPLOADS=1` to enable the separate public, upload-only endpoint used by npm and standalone clients; this permits anyone to request a bounded upload URL, so apply rate limits/quotas at your deployment edge as appropriate. It does not permit reading or listing captured histories. `AGENT_TRACE_RELEASE_MANIFEST` maps public release ZIP names to their R2 object keys and SHA-256 hashes. Only those allowlisted release objects can be downloaded without authentication. Keep preview deployments protected; participants need to reach the production API without signing into Vercel.
|
|
27
|
+
|
|
28
|
+
Vercel only handles small JSON requests and the configured client download; trace ZIPs go directly from participant devices to R2/S3. No Vercel disk storage, long-running process, or large HTTP body is needed. Project dependencies include Flask and boto3 so Vercel's normal Python dependency installation bundles them. The legacy Python download remains a standard-library-only zipapp; npm and standalone downloads use the Go collector.
|
|
29
|
+
|
|
30
|
+
## Server setup
|
|
31
|
+
|
|
32
|
+
Copy `.env.example` to `.env` and configure:
|
|
33
|
+
|
|
34
|
+
| Environment variable | Purpose |
|
|
35
|
+
| --- | --- |
|
|
36
|
+
| `AGENT_TRACE_PUBLIC_URL` | Public HTTPS address of this server, e.g. `https://traces.example.com` |
|
|
37
|
+
| `AGENT_TRACE_BUCKET` | Destination S3/R2 bucket |
|
|
38
|
+
| `AWS_ACCESS_KEY_ID` | Server's storage access key; AWS roles/profiles also work |
|
|
39
|
+
| `AWS_SECRET_ACCESS_KEY` | Server's storage secret |
|
|
40
|
+
| `AWS_SESSION_TOKEN` | Optional, for temporary AWS credentials |
|
|
41
|
+
| `AGENT_TRACE_ENDPOINT_URL` | R2/S3-compatible endpoint; omit for AWS S3 |
|
|
42
|
+
| `AGENT_TRACE_REGION` | AWS region; use `auto` for R2 |
|
|
43
|
+
| `AGENT_TRACE_UPLOAD_TOKEN` | Random enrollment token, at least 32 printable ASCII characters |
|
|
44
|
+
| `AGENT_TRACE_PREFIX` | Bucket prefix; defaults to `agent-traces` |
|
|
45
|
+
| `AGENT_TRACE_MAX_BYTES` | Maximum archive size; defaults to 1 GiB; single PUT supports up to 5 GiB |
|
|
46
|
+
| `AGENT_TRACE_URL_TTL` | Upload URL lifetime; defaults to 900 seconds, allowed range 60–3600 |
|
|
47
|
+
|
|
48
|
+
Generate an enrollment token with `python3 -c 'import secrets; print(secrets.token_urlsafe(32))'`. This token allows issuing upload URLs and downloading the configured client. It does not allow reading/listing bucket objects and is separate from bucket credentials.
|
|
49
|
+
|
|
50
|
+
For R2, set:
|
|
51
|
+
|
|
52
|
+
```text
|
|
53
|
+
AGENT_TRACE_ENDPOINT_URL=https://ACCOUNT_ID.r2.cloudflarestorage.com
|
|
54
|
+
AGENT_TRACE_REGION=auto
|
|
55
|
+
```
|
|
56
|
+
|
|
57
|
+
Use the S3 API endpoint, including a jurisdiction-specific endpoint if applicable, rather than an `r2.dev` URL. Create the bucket ahead of time. Give the server storage credentials permission to write to the configured bucket/prefix. Configuration is loaded at process startup; restart the server after changing it. `.env` files are ignored by Git and excluded from Docker's build context.
|
|
58
|
+
|
|
59
|
+
Run with Docker:
|
|
60
|
+
|
|
61
|
+
```sh
|
|
62
|
+
docker build -t agent-trace-server .
|
|
63
|
+
docker run --rm --env-file .env -p 127.0.0.1:8000:8000 agent-trace-server
|
|
64
|
+
```
|
|
65
|
+
|
|
66
|
+
Put an HTTPS reverse proxy/load balancer in front of port 8000. Configure upload-ticket request rate limits there; the server limits individual object sizes but does not implement per-participant quotas. The image runs Gunicorn as a non-root user. It does not need a persistent volume or database. `GET /healthz` is a process health check, not a test of bucket permissions.
|
|
67
|
+
|
|
68
|
+
Or install/run directly on a Linux server:
|
|
69
|
+
|
|
70
|
+
```sh
|
|
71
|
+
python3 -m venv .venv
|
|
72
|
+
. .venv/bin/activate
|
|
73
|
+
python3 -m pip install '.[server]'
|
|
74
|
+
# Export the variables from your server's environment manager before starting.
|
|
75
|
+
gunicorn --bind 127.0.0.1:8000 --workers 2 'agent_trace_server:create_app()'
|
|
76
|
+
```
|
|
77
|
+
|
|
78
|
+
The application does not automatically load `.env`; Docker's `--env-file` or your process manager supplies the environment. For local development, `AGENT_TRACE_PUBLIC_URL=http://127.0.0.1:8000` is permitted. HTTP URLs are otherwise rejected by the client.
|
|
79
|
+
|
|
80
|
+
## Legacy Python client installation (private servers)
|
|
81
|
+
|
|
82
|
+
After the server is running, macOS/Linux participants install using the enrollment token provided by the administrator:
|
|
83
|
+
|
|
84
|
+
```sh
|
|
85
|
+
curl -fsS -H 'Authorization: Bearer YOUR_ENROLLMENT_TOKEN' \
|
|
86
|
+
https://traces.example.com/install.sh -o install-agent-trace-capture.sh
|
|
87
|
+
sh install-agent-trace-capture.sh
|
|
88
|
+
```
|
|
89
|
+
|
|
90
|
+
The installer puts a standalone Python zipapp at `~/.local/bin/agent-trace-capture`. Add `~/.local/bin` to PATH if it is not already there. The zipapp contains the server URL and an upload-only enrollment token, so subsequent runs need no flags, environment variables, Python packages, or bucket credentials. Administrators can also distribute the downloaded executable directly through their normal software installation process.
|
|
91
|
+
|
|
92
|
+
The configured client and installer contain the enrollment token: distribute them only to your participants. It is a shared upload capability, not participant authentication. Token rotation requires distributing a new client. This version does not include accounts, per-device enrollment/revocation, or a public anonymous registration flow.
|
|
93
|
+
|
|
94
|
+
Windows participants can download the same `/client.pyz` artifact and use a small command wrapper. PowerShell, with Python 3.10+ installed:
|
|
95
|
+
|
|
96
|
+
```powershell
|
|
97
|
+
$headers = @{ Authorization = 'Bearer YOUR_ENROLLMENT_TOKEN' }
|
|
98
|
+
$destination = "$env:LOCALAPPDATA\AgentTraceCapture"
|
|
99
|
+
New-Item -ItemType Directory -Force $destination | Out-Null
|
|
100
|
+
Invoke-WebRequest https://traces.example.com/client.pyz -Headers $headers -OutFile "$destination\client.pyz"
|
|
101
|
+
'@py -3 "%~dp0client.pyz" %*' | Set-Content "$destination\agent-trace-capture.cmd" -Encoding ascii
|
|
102
|
+
$env:Path += ";$destination"
|
|
103
|
+
```
|
|
104
|
+
|
|
105
|
+
Add that directory to the user's persistent PATH to keep the command available in new terminals. The same run command then works on Windows:
|
|
106
|
+
|
|
107
|
+
```sh
|
|
108
|
+
agent-trace-capture
|
|
109
|
+
```
|
|
110
|
+
|
|
111
|
+
The raw package is not published on PyPI. An unconfigured source checkout needs `--server-url`/`--upload-token` or equivalent `AGENT_TRACE_SERVER_URL`/`AGENT_TRACE_UPLOAD_TOKEN` environment variables; the downloaded client already includes those defaults.
|
|
112
|
+
|
|
113
|
+
## Participant identity and uploads
|
|
114
|
+
|
|
115
|
+
The first upload generates a random UUID saved in `~/.agent-trace-capture/participant-id`. Repeat uploads use that ID:
|
|
116
|
+
|
|
117
|
+
```text
|
|
118
|
+
s3://BUCKET/agent-traces/<participant-id>/YYYY/MM/DD/<upload-uuid>.zip
|
|
119
|
+
```
|
|
120
|
+
|
|
121
|
+
An ID identifies an installation. To group one person across devices, assign the same `--participant-id study_001` or `AGENT_TRACE_PARTICIPANT_ID` on each device. The UUID is not linked to a name unless you maintain a separate mapping. Private IP addresses are not collected. Shared/reused/changing addresses make them unsuitable person identifiers even when hashed. IDs are client-supplied labels, not authenticated ownership boundaries.
|
|
122
|
+
|
|
123
|
+
The server generates the object key; clients cannot choose arbitrary bucket keys. Upload URLs permit one key for a limited time and bind the declared size, content type, MD5, and participant/SHA-256 metadata using S3 Signature V4. URLs can be reused until expiry for that same content/key. Object storage verifies Content-MD5; the SHA-256 value is client-supplied metadata. The server never receives or extracts the archive and has no upload-completion database. The CLI reports success only after object storage returns success. Presigned URLs must remain private.
|
|
124
|
+
|
|
125
|
+
Captures remain on the participant's disk after upload or failure. Partial captures are not uploaded. This version has no multipart upload or automatic retry/resume; rerunning the command captures again with a new object key.
|
|
126
|
+
|
|
127
|
+
## Capture options
|
|
128
|
+
|
|
129
|
+
```sh
|
|
130
|
+
agent-trace-capture --dry-run
|
|
131
|
+
agent-trace-capture --local-only --output ./capture.zip
|
|
132
|
+
agent-trace-capture --providers claude codex
|
|
133
|
+
agent-trace-capture --codex-home /custom/codex --claude-home /custom/claude
|
|
134
|
+
agent-trace-capture --cursor-home /custom/cursor --cursor-user-dir /custom/Cursor/User
|
|
135
|
+
```
|
|
136
|
+
|
|
137
|
+
`--dry-run` lists candidate files without reading transcript contents, saving an identity, or making network calls. `--local-only` creates a ZIP without uploading. Default output: `agent-traces-<UTC timestamp>.zip` in the current folder. Existing outputs are never overwritten. `--identity-file` relocates the saved participant ID.
|
|
138
|
+
|
|
139
|
+
| Tool | Captured locations |
|
|
140
|
+
| --- | --- |
|
|
141
|
+
| Claude Code | `${CLAUDE_CONFIG_DIR:-~/.claude}/projects/**/*.jsonl`, including subagents |
|
|
142
|
+
| Codex | `${CODEX_HOME:-~/.codex}/sessions/` and `archived_sessions/`; `.jsonl` and `.jsonl.zst` |
|
|
143
|
+
| Cursor IDE | `~/.cursor/projects/*/agent-transcripts/`, recursively; `.jsonl`, `.txt`, `.md` |
|
|
144
|
+
| Cursor chat databases | `User/globalStorage/**/state.vscdb` and `User/workspaceStorage/**/state.vscdb` |
|
|
145
|
+
| OpenCode | `${XDG_DATA_HOME:-~/.local/share}/opencode/opencode.db`; legacy `storage/session`, `message`, `part`, and `session_diff` JSON files |
|
|
146
|
+
| Kimi Code | `~/.kimi/sessions/` and `imported_sessions/`: `context.jsonl`, `wire.jsonl`, including subagents |
|
|
147
|
+
| Gemini CLI | `~/.gemini/tmp/**/chats/**/session-*.json` and `.jsonl` |
|
|
148
|
+
| Copilot CLI | `~/.copilot/session-state/**/events.jsonl` |
|
|
149
|
+
| Continue | `~/.continue/sessions/*.json`; honors `CONTINUE_GLOBAL_DIR` |
|
|
150
|
+
| Cline | Editor `User/globalStorage/saoudrizwan.claude-dev/tasks/` and `~/.cline/data/tasks/` |
|
|
151
|
+
| Roo Code | Editor `User/globalStorage/rooveterinaryinc.roo-cline/tasks/` |
|
|
152
|
+
| Aider | `.aider.chat.history.md` in the current directory/Git root; extra projects via `--project-dir` |
|
|
153
|
+
|
|
154
|
+
Cursor's User folder defaults to `~/Library/Application Support/Cursor/User` on macOS, `%APPDATA%\Cursor\User` on Windows (fallback `%USERPROFILE%\AppData\Roaming\Cursor\User`), and `${XDG_CONFIG_HOME:-~/.config}/Cursor/User` on Linux. Claude/Codex use the home directory on all three platforms. WSL and remote hosts have separate stores: run the client there to capture them. Custom profiles and portable installs require explicit overrides.
|
|
155
|
+
|
|
156
|
+
Cursor databases are read in a read-only SQLite transaction including committed WAL data. Recognized chat keys from `ItemTable` and `cursorDiskKV` are exported as JSONL rather than copying entire settings/auth databases. This internal format may change; a zero-row export can indicate an unsupported schema. Cursor CLI `~/.cursor/chats`, cloud-only conversations, attachments, and non-transcript sidecars are not included. This is a trace export, not a restorable application backup.
|
|
157
|
+
|
|
158
|
+
OpenCode databases are also read in a read-only transaction and exported as JSONL records containing `table` and `row`. Only allowlisted project/session/message/part/context/todo tables are read; `credential`, `account`, authentication files, and unrelated tables are excluded. The adapter was checked against OpenCode 1.19.5. A database without the expected `session` table produces an explicit partial-capture error. Use `--opencode-home` for a custom data directory, or `--opencode-db`/`OPENCODE_DB` for a custom database (relative paths resolve under the data directory). Future storage schemas require a verified adapter update.
|
|
159
|
+
|
|
160
|
+
The file adapters preserve existing transcript bytes rather than parsing every tool's schema. Kimi, Gemini, Copilot, Continue, Cline, Roo, and Aider adapters are covered by synthetic fixture tests and primary-source layout checks; they have not each been exercised with live authenticated accounts. The OpenCode adapter was exercised with a real free-model session inside Docker and a verified R2 upload.
|
|
161
|
+
|
|
162
|
+
Cline/Roo discovery includes Code, Code Insiders, Cursor, VSCodium, and Windsurf editor User folders, named editor profiles, and standard VS Code remote server folders. This does not capture Windsurf's own Cascade conversations. Only the known task conversation/event/metadata filenames are copied; editor settings and global credential databases are excluded. Add custom editor User folders with repeated `--editor-user-dir` flags. Root overrides are available as `--kimi-home`, `--gemini-home`, `--copilot-home`, `--continue-home`, and `--cline-home`. For Aider, run from a project or pass repeated `--project-dir` flags; the collector does not recursively search the entire disk.
|
|
163
|
+
|
|
164
|
+
Amp and Windsurf Cascade are not advertised as automatically captured. Amp's documented export uses its cloud-backed thread interface, and no verified local Cascade transcript adapter is included. Adding their names without a usable collector would give a misleading capture result.
|
|
165
|
+
|
|
166
|
+
Raw traces are not redacted and may contain private code, prompts, personal data, or secrets pasted into conversations. The ZIP's manifest also includes original paths, which may contain a username. Each run prints that it is capturing unredacted traces; the configured client uploads them automatically. Symlink files/directories are skipped. POSIX ZIP permissions are owner-only; Windows inherits filesystem permissions.
|
|
167
|
+
|
|
168
|
+
`manifest.json` records paths, provider labels, sizes, SHA-256 hashes, modification times, database row counts, and errors. Close agents for the most stable capture; files can grow during capture, and detected changes are marked. Multiple databases are captured independently. Exit codes: `0` success (including an empty dry run), `1` no traces, `2` partial capture/configuration/capture/upload failure.
|
|
169
|
+
|
|
170
|
+
## Development
|
|
171
|
+
|
|
172
|
+
The participant-facing native collector is in `native/`. Run its tests in the Go build container with `GOTOOLCHAIN=auto go test ./...`. `scripts/build-release.py` builds six targets and generates complete linked-dependency notices; `npm test` checks the launcher, and `npm pack --dry-run` verifies the publish allowlist. Standalone ZIPs and their hashes are written under `releases/`. Publishing requires an npm account with 2FA.
|
|
173
|
+
|
|
174
|
+
|
|
175
|
+
Install `.[server]` in a development environment, then run:
|
|
176
|
+
|
|
177
|
+
```sh
|
|
178
|
+
python3 -m unittest discover -s tests -v
|
|
179
|
+
```
|
|
180
|
+
|
|
181
|
+
Tests use synthetic traces and SQLite fixtures, real boto3 signing with fake credentials, and a local HTTP flow running the downloaded zipapp with no arguments and no installed dependencies. They do not upload personal traces. Native Windows execution and real S3/R2 storage need integration testing with those environments.
|
|
182
|
+
|
|
183
|
+
References: [R2 presigned uploads](https://developers.cloudflare.com/r2/api/s3/presigned-urls/), [boto3 credentials](https://docs.aws.amazon.com/boto3/latest/guide/credentials.html), [Flask production deployment](https://flask.palletsprojects.com/en/stable/deploying/), [Claude storage configuration](https://code.claude.com/docs/en/settings), [Codex rollout storage](https://github.com/openai/codex/blob/main/codex-rs/tui/src/session_archive_commands_tests.rs), [Cursor transcript storage](https://forum.cursor.com/t/allow-storing-conversation-history-inside-the-project-directory/160515).
|