Local observable AI work timeline

This document describes the source candidate following RC6, not a replacement for the immutable published RC6/RC6.1 artifacts. The timeline uses the existing Linux user service journal. Windows journal viewing remains unsupported.

commandcore-agent activity
commandcore-agent activity --json
commandcore-agent activity --last 100
commandcore-agent activity --errors
commandcore-agent activity --service commandcore-agent-canary.service

The default follows live events as UTC timestamped REQUEST, RESULT and CONNECTION blocks with pretty JSON. --json produces one redacted JSON object per line. --last reads recent journal entries and exits. Unstructured historical debug messages are ignored. Requests and results share both the protocol request ID and execution ID; concurrent work can interleave without losing correlation.

Requests show exact sanitized commands, effective cwd, file/source/destination paths, offsets, lengths, timeout, script/process arguments, and source/client information when the server supplies it. Results show operation status, process ID, exit code, duration, returned/written bytes, change status, and bounded text previews. Shell streams are collected across chunk boundaries before redaction, then shown in the final RESULT. For long managed jobs, process.output requests show the output returned by that poll. Connection and retry events are separate from execution results; interrupted requests receive a closing cancellation event.

An operator can reconstruct this sequence from the terminal:

1. Write config.py (path and byte count).
2. Read config.py (path, bytes and value = 1 preview).
3. Write check_config.py (path and byte count).
4. Run Python check_config.py in the workspace; observe configuration/test output.
5. Patch config.py (one patch, changed result).
6. Read the updated value = 2.
7. Repeat the Python check and observe exit code zero and tests passing.

Redaction boundary

Python and Rust share activity_policy.json. Agents sanitize objects before writing structured events to the journal; viewers apply the same policy again before rendering. Server audit storage retains its separate metadata-only policy. The wire requests, wire results and execution permissions are unchanged.

• Only reviewed contextual fields are retained; unknown payload fields are omitted.
• Known device credentials/private keys, secret environment values and supplied environment values are masked wherever they recur in text.
• Credential assignments, secret flags, auth/cookie headers, URL credentials, private-key blocks and long encoded values are redacted. Ordinary commands, including python -u, and ordinary paths remain readable.
• fs.write emits byte counts, not content. Patches emit counts, not replacement text. Keyboard typing, clipboard and transfer payloads are metadata-only.
• fs.read suppresses encoded/binary previews and previews of recognized sensitive paths such as .env, SSH keys, credential files and Agent state.
• Text fields retain at most 2,048 characters with a truncation marker; output counts describe original streamed bytes. Stream collection holds at most 16,384 characters per stream. Total structured events are bounded to 16 KiB; oversized context fields are replaced with a bounded-field marker.
• ANSI/control characters are removed and string contents are JSON escaped.

Redaction cannot infer that an arbitrary short, unlabelled string in an otherwise ordinary text file is a secret. Sensitive data must be recognizable through its field, path, credential format or registered value. Local journal access and retention remain controlled by the operating system; text previews can contain ordinary private project information.

Disposable acceptance

tests/test_local_activity.py exercises the actual Python request handler with disposable file reads/patches and subprocess execution. Rust activity tests execute the same file/edit/command pipeline through the actual executor. Redaction tests cover both render formats, split output, byte counts, sensitive payloads, long paths and useful commands. scripts/local_activity_fixture.py supplies a bounded local workload suitable for an isolated user service; it does not enroll a device or connect to a production server.