dev-trajectory 0.1.1

Local, single-user analysis of Claude Code session histories — surfaces engineering trajectory over time per domain, with evidence on hover.

dev-trajectory

A local, single-user CLI that ingests your Claude Code session histories, scores each session along five engineering capability dimensions using an LLM judge, and renders your trajectory over time as a per-domain or per-language summary with evidence quotes.

Status: alpha. Phases 0–2 of the project plan are complete — discover → parse → sample → score → aggregate, plus a local browser dashboard. Phase 4 (MCP server) and Phase 5 (Claude skill) are not built yet.

What it does

Claude Code writes JSONL transcripts of every session you run. They sit in ~/.claude/projects/ and are read by no one. Across hundreds of sessions, patterns aggregate: how you frame ambiguous tasks, whether you catch the model's mistakes, how you redirect when an approach is going wrong. dev-trajectory summarizes each session, scores it, and aggregates weekly so you can see where you're operating like a senior engineer and where you're still learning.

The local dashboard (devtraj serve): per-session scatter with per-domain best-fit lines on the left, evidence quotes from the clicked session on the right.

Privacy

• Local-only storage. Sessions and scores live in a SQLite file under your user data directory. Nothing is uploaded.
• Your own API key. Scoring uses your Anthropic API key. No proxying.
• Zero telemetry in this build. A consent flag and toggle command exist (devtraj telemetry on|off|status) for a planned future opt-in aggregate study, but the uploader is hard-disabled — no endpoint is wired up.
• No automatic redaction. Evidence quotes are stored verbatim because they're what makes the dashboard credible. Since data never leaves the machine, this is a deliberate trade-off rather than a leak risk. See docs/PROJECT.md §10 for the full posture.

What this can't tell you

These shape the architecture, not just the README — read them before drawing strong conclusions from the output:

• You only use AI for things you don't already know. This tool measures your behavior on those topics, which are usually your blind spots. A staff backend engineer learning Rust will look junior in their Rust sessions. The score reflects growth areas, not absolute capability.
• LLM judges drift. Single-shot scoring on subjective work has real variance. Median-of-3 voting, forced verbatim evidence quotes, and a fixed anchored rubric narrow the noise but don't eliminate it. Treat individual session scores as soft; trust trends over many weeks.
• A flat or declining line is real information you didn't ask for. During a hard quarter the chart will reflect that. The dashboard slices per-domain so a slump in one area doesn't drown out growth elsewhere — but the signal is honest, not encouraging.

See docs/PROJECT.md §2 for the longer treatment.

Install

pipx install dev-trajectory
export ANTHROPIC_API_KEY=sk-ant-...

Requires Python 3.11+ and an Anthropic API key. To persist the key instead of exporting it each session, run devtraj init. Plain pip install dev-trajectory inside a venv works too. From source: see CONTRIBUTING.md.

Quickstart

After install + API key setup:

devtraj analyze                     # summarize + score new sessions
devtraj report                      # per-week trajectory by domain (text)
devtraj report --slice-by language  # alternate slice
devtraj explain 2026-04-13          # evidence quotes for a given week
devtraj serve                       # interactive dashboard at http://127.0.0.1:8765/

By default analyze samples up to 3 sessions per ISO week to control cost. Sampling is deterministic — re-runs pick the same sessions every time, so re-running analyze is safely idempotent. Set sessions_per_week = 0 in config (or pass --sessions-per-week 0) to score every session.

Sessions are processed in parallel (default 10 at a time). Inside each session the median-of-3 scoring runs 3 calls in parallel, so total in-flight LLM calls is roughly max_concurrent_sessions * 4. Override with --max-concurrent-sessions N, or set max_concurrent_sessions in config.

CLI reference

Command	What it does
devtraj init	Detect Claude Code session paths, write a default config, verify the rubric loads, and surface the telemetry toggle.
devtraj analyze	Walk new sessions, summarize each (Haiku 4.5), score 3× with median voting (Sonnet 4.6), persist to SQLite. Skips sessions already scored at the current rubric version.
devtraj analyze --rebuild	Force re-summary and re-score for every sampled session.
devtraj analyze --sessions-per-week N	Override the per-week sampling cap for this run (0 disables).
devtraj analyze --max-concurrent-sessions N	Override how many sessions are processed in parallel (default 10).
devtraj report [--slice-by domain|language|all]	Print a weekly median table per slice.
devtraj report --min-confidence 0.7	Drop scores below the given confidence floor.
devtraj explain YYYY-MM-DD	Show evidence quotes for the ISO week starting on that date.
devtraj rescore	Re-score sessions whose stored rubric version is older than the current one.
devtraj serve [--port N] [--no-open]	Launch the local dashboard. Two chart modes: weekly-median lines per domain, and a per-session scatter with per-domain best-fit. Click any point or bucket to load evidence quotes in the side panel; expand a session card to see the full user-turn text. Default 127.0.0.1:8765.
devtraj telemetry status	Show the local consent flag (uploader is disabled regardless).
devtraj paths	Print the config and database paths.
devtraj rubric [--prompt]	Show the active rubric (version, source, bands, dimensions). With --prompt, dump the full system prompt sent to the judge.

How scoring works

Five dimensions, each scored 1 (junior) to 5 (principal):

• problem_decomposition — how well you break ambiguous tasks into structured sub-problems.
• systems_thinking — whether you reason about effects beyond the immediate change.
• debugging_methodology — hypothesis quality, use of evidence, willingness to bisect.
• technical_communication — precision of language; whether you give the model what it needs to answer well.
• model_redirection — whether you catch and correct the model when it's wrong. Weighted most heavily because it's the dimension least confounded by topic familiarity.

Each session also gets a seniority_band (junior / mid / senior / staff / principal — the primary surfaced signal), an inferred_ic_level (IC2–IC6), a domain_category, a language, and 3–5 verbatim evidence quotes. See docs/PROJECT.md §4.4 for the full schema.

The rubric anchors are paraphrased from CircleCI's published engineering competency matrix. The exact rendered system prompt lives in src/dev_trajectory/rubric/v1.yaml — edit it and bump the version to iterate.

Configuration

Config lives at ~/.config/dev-trajectory/config.toml (Linux/macOS) or the platform equivalent. Run devtraj paths to see the exact location.

api_key = "sk-ant-..."          # optional; ANTHROPIC_API_KEY env var takes precedence
rubric_version = "v1"
min_confidence = 0.5            # scores below this are hidden from the report by default
tool_output_token_limit = 2000  # tool outputs longer than this are stripped from the user-turn extraction
sessions_per_week = 3           # 0 = no cap (score every discovered session)
max_concurrent_sessions = 10    # how many sessions to process in parallel
telemetry_enabled = false
custom_session_paths = []       # extra directories to scan for *.jsonl beyond ~/.claude/projects/

Cost

A year of heavy Claude Code use is roughly 500–2000 sessions. With prompt caching on the rubric system prompt, median-of-3 scoring, and the default 3/week sampling cap, total spend on a fresh run is single-digit dollars. Subsequent runs are effectively free since they only score new weeks.

Roadmap & contributing

See docs/ROADMAP.md for what's planned and roughly when. Suggestions to reprioritize, add, or drop items are welcome — open an issue or a PR against the roadmap itself.

For everything else (bug reports, rubric tweaks, new tool support, dashboard polish), see CONTRIBUTING.md.

License

MIT. See LICENSE.