RFD D15: Structured Logging Infrastructure

• Status: Draft
• Category: Design
• Authors: Jean Mertz git@jeanmertz.com
• Date: 2026-04-05
• Extends: RFD 048

Summary

This RFD replaces JP's temporary-file tracing with persistent, structured log files at ~/.local/share/jp/logs/, adds an in-memory buffering layer for deferred file path resolution, and introduces a post-run log report. It redefines the semantics of -v, -q, JP_DEBUG, and --log-file to give users clear control over what gets logged and where.

Motivation

RFD 048 separates JP's output into four channels: stdout (assistant content), stderr (chrome), /dev/tty (prompts), and a log file (tracing). Phases 1 and 2 are implemented. Phase 3 — moving tracing to a log file — was deferred because the original design (a simple flag to redirect tracing) lost important nuance around -v behavior, log discoverability, and post-mortem debugging.

The current tracing setup writes to a temporary file that is silently discarded on success. Users who run jp -v query see tracing on stderr mixed with chrome. Users who don't use -v have no persistent log history. When something goes wrong, the only recourse is re-running with -v and hoping to reproduce the issue.

This RFD addresses these problems:

1. No persistent logs by default. Tracing data is lost unless the user explicitly enables it or the process fails.
2. -v mixes tracing with chrome on stderr. This makes 2> chrome.log capture both chrome and tracing noise, defeating the channel separation from Phase 1.
3. No post-mortem summary. When warnings or errors occur during a run, the user has no summary — they must scroll through interleaved output or grep a log file.
4. Log file naming is opaque. The current temp file has no connection to the session or conversation it belongs to.

Design

Log file location and naming

Every JP invocation writes a log file to:

~/.local/share/jp/logs/<session-id>-<conversation-id>-<timestamp>.log

• session-id is the resolved session identity (from session::resolve()). Falls back to unknown if no session is available.
• conversation-id is the primary conversation ID for the command. Falls back to none for commands that don't use a conversation (e.g., jp init).
• timestamp is the UTC start time in YYYYMMDD-HHMMSS format.

Example: ~/.local/share/jp/logs/12345-a1b2c3-20260405-143022.log

Deferred file path resolution

The log file path depends on values (session ID, conversation ID) that are not available at process start. Logging must begin immediately to capture early startup events. The solution is an in-memory buffer layer:

1. At startup, configure_logging() installs a custom tracing layer that buffers events in memory.
2. After session and conversation resolution, run_inner() calls guard.activate(session_id, conversation_id) which:
   • Builds the log file path
   • Creates the ~/.local/share/jp/logs/ directory if needed
   • Opens the file
   • Flushes all buffered events to it
   • Switches to tee mode (memory + file) for subsequent events
3. If activation never happens (e.g., jp init or early error), the buffer is written to a fallback file named unknown-none-<timestamp>.log.

In-memory event layer

A custom tracing layer that serves three purposes:

1. Buffering: Holds events in a Vec until the file path is resolved.
2. Counting: Tracks warn_count and error_count for the log report.
3. Recent errors: Keeps the last 5 WARN/ERROR messages in a ring buffer for the post-run summary.

After activation, the layer tees events to both the file and the in-memory counters. The buffer is cleared after flushing to the file.

Verbosity semantics

Flag	File log level	Stderr tracing	Log report
(none)	TRACE	off	on error only
-v	WARN	WARN	yes
-vv	INFO	INFO	yes
-vvv	DEBUG	DEBUG	yes
-vvvv	TRACE	TRACE (jp crates)	yes
-vvvvv	TRACE	TRACE (+ third-party)	yes
JP_DEBUG=1	TRACE	TRACE (jp crates)	yes
-q	TRACE	off	off

Key changes from current behavior:

• Default (no -v): The file captures TRACE. Stderr has no tracing layer. The log report is printed only on error.
• -v once or more: The file captures at the specified level (not full TRACE). Stderr shows tracing at the same level. The log report is always printed.
• JP_DEBUG=1: Equivalent to -vvvv. Resolved before configure_logging so it affects both file and stderr layers from the start.
• -q: Suppresses tracing output and the log report entirely. The file still captures TRACE for post-mortem use.

--log-file flag

--log-file <PATH>

Writes tracing to an additional destination:

• --log-file=- writes to stderr (useful for piping to external tools).
• --log-file=/path/to/file writes to the specified file.
• --log-file=/dev/null suppresses the additional output (the default log file is still written).

This is additive — the log at ~/.local/share/jp/logs/ is always written regardless of --log-file.

Log report

When enabled (any -v, JP_DEBUG, or non-zero exit code), the report is printed to stderr after the run completes:

⚠ 3 warnings, 1 error during run (2.4s)

  WARN  Rate limited, retrying (1/3)
  WARN  Tool fs_create_file skipped by user
  ERROR Stream error: connection reset
  ... (2 more warnings, see log file)

Log: ~/.local/share/jp/logs/12345-a1b2c3-20260405-143022.log

The report includes:

• Total warning and error counts, plus run duration
• Up to 5 most recent WARN/ERROR messages (one line each)
• Truncation notice if more than 5
• Log file path

The report is suppressed by -q.

--log-format

Controls the format of stderr tracing output (when enabled via -v or --log-file=-). Unchanged from current behavior:

• auto: text for terminals, JSON for pipes
• text: compact human-readable format
• json: NDJSON format

The persistent log file always uses JSON format for machine parseability.

Drawbacks

Default TRACE logging produces large files. Without log rotation, the ~/.local/share/jp/logs/ directory will grow unbounded. This is mitigated by a planned log rotation task (not in this RFD's scope) and by the fact that individual log files for typical queries are small (< 1 MB).

The in-memory buffer layer adds complexity. A custom tracing layer with buffering, flushing, counting, and tee semantics is approximately 150-250 lines of non-trivial code. It must be thread-safe and handle the transition from buffer-only to tee mode atomically.

-v changes file verbosity. A user who runs with -v and later needs the full trace for debugging won't have it. This is an intentional trade-off — users who explicitly set a verbosity level are communicating what level of detail they want. The default (no -v) still captures full TRACE.

Alternatives

Keep the temp file approach

Continue writing to NamedUtf8TempFile and persisting on error. This loses log history and makes post-mortem debugging harder for intermittent issues.

Always write full TRACE regardless of -v

Simpler but creates large files on every run and removes the user's ability to control file verbosity. Rejected in favor of the "default = TRACE, -v = user's choice" model.

Use the tracing-appender crate

Provides rolling file appenders out of the box. However, it doesn't support the deferred path resolution or in-memory buffering needed here. The custom layer is simpler than wrapping tracing-appender with the required lifecycle.

Non-Goals

• Log rotation and cleanup. Planned as a follow-up task. The directory will accumulate files until that work is done.
• Graduated -q levels (-qq = no chrome, -qqq = no reasoning, etc.). This requires changes across the Printer, ChatResponseRenderer, and interactivity systems. Tracked separately.
• Structured log querying. The JSON log files can be parsed with jq, but this RFD does not add a jp log subcommand or similar.

Risks and Open Questions

Buffer memory usage

The in-memory buffer holds all events from process start until the file path is resolved. For typical startup sequences, this is a few hundred events (< 100 KB). For pathological cases (e.g., workspace with thousands of conversations triggering warnings during index load), the buffer could grow larger. A cap (e.g., 10,000 events, dropping oldest) would bound memory usage at the cost of losing early events.

File I/O on the hot path

Opening the log file and flushing the buffer happens once during startup, not on the streaming hot path. Subsequent writes go through the tracing subscriber's normal I/O path. No performance concern for typical usage.

Log file format stability

The JSON log format is an internal implementation detail, not a public API. However, users may build tooling around it (e.g., jq scripts). A format version field in each log entry would allow future changes without silent breakage.

Implementation Plan

Phase 1: Persistent log directory and file layer

Replace the NamedUtf8TempFile with a file at ~/.local/share/jp/logs/. Use a timestamp-only filename initially. Wire JP_DEBUG as a -vvvv alias. Update the TracingGuard to always persist (no conditional).

Can be merged independently.

Phase 2: In-memory buffer layer

Implement the custom tracing layer with buffering, counting, and tee semantics. Add guard.activate(session_id, conversation_id) to run_inner(). Rename the log file to include session and conversation IDs.

Depends on Phase 1.

Phase 3: Log report and -v semantics

Add the post-run log report. Change -v to control file log level and enable the stderr layer. Add --log-file flag.

Depends on Phase 2.

References

• RFD 048: Four-Channel Output Model — parent RFD; this extends Phase 3 (tracing to log file).
• crates/jp_cli/src/lib.rs — current configure_logging and TracingGuard.
• crates/jp_cli/src/session.rs — session identity resolution.