trajrl

The official Python CLI for TrajectoryRL — an open skill factory that uses Bittensor's distributed compute and incentive layer to produce state-of-the-art agent skills.

trajrl watches live SN11 state: winner, challenger, queue, validators, miners, packs, eval logs, and deep validator analysis.

CLI output is Rich tables in a TTY and JSON when piped.

For generic Bittensor on-chain queries (metagraph, hyperparams) use btcli — the official Bittensor CLI.

Install

pip install trajrl

Live SN11 state (v6 winner-challenger)

trajrl challenge                    # in-flight epoch — challenger pack + per-validator scores so far
trajrl winner                       # current seated winner + last 5 change events
trajrl winner --history 20          # show more history
trajrl queue                        # pending eval queue
trajrl queue --eligible-only        # filter to submissions eligible right now

Validators / miners / packs

trajrl validators                   # roster table
trajrl validators --detail          # adds stake / weightTargets / benchVersion
trajrl miner --uid 63               # miner detail by UID
trajrl miner HOTKEY
trajrl pack HOTKEY PACK_HASH        # specific pack + eval results
trajrl submissions                  # recent submissions across the network
trajrl submissions --failed

Subnet analysis

trajrl analyze                              # last 24h, all reports
trajrl analyze --epochs 50                  # explicit epoch window
trajrl analyze --last 6                     # last 6 hours
trajrl analyze --scenario cancel-async-tasks   # focus per-scenario report
trajrl analyze --no-compare                 # skip the validator-sync table
trajrl analyze --deep                       # drill into eval logs for top packs

analyze produces, in one report:

1. Throughput — epochs in window, decisions submitted, decisions/hour, rejection rate
2. Competition Health — distinct challenger packs/miners, outcomes (held/replaced), replace rate, current winner tenure, mean inter-replacement gap
3. Score Distribution — N qualified, mean, p50/p75/p90/p99, histogram of consensus_score
4. Per-Scenario — pass rate, mean score, eval count, top-scoring pack per scenario (filterable with --scenario)
5. Top 10 Challenger Packs — best consensus_score in window, with outcome
6. Rejection Breakdown — bucketed reasons (integrity:copy, eval_error, schema_validation, etc.) with sample details
7. Miner Pool — distinct miners, distinct packs, top-10 most-active
8. Validator Sync — per-validator mean Δ vs peer mean; outlier flag with adaptive threshold (1.5× network median)
9. Recent Winner Changes — last 10 replacement events
10. With --deep: drill into eval logs for the top challenger packs in the window

Network state & feeds

trajrl stats                         # network totals (reports, cost USD, LLM calls, tokens)
trajrl epoch 12345                   # full epoch detail + per-validator submissions
trajrl leaderboard                   # v6 leaderboard — winner + ranked miners
trajrl events --limit 20             # recent eval events (validator → miner reports)
trajrl events --since 2026-05-21T00:00:00Z
trajrl winners                       # winner takeover history + cumulative α emission
trajrl prices                        # TAO + SN11 alpha price strip + 30d stake APY
trajrl nodes                         # node heartbeat summary
trajrl nodes --type validator

Validator drill-downs

trajrl scores HOTKEY                 # one validator's recent per-miner reports (last 24h)
trajrl scores --uid 7
trajrl progress                      # per-validator progress through current cycle
trajrl activity HOTKEY --limit 50    # per-epoch participation history

Bench / scenario quality

trajrl bench                         # per-scenario quality leaderboard (latest spec)
trajrl bench --spec 7
trajrl bench-miner HOTKEY            # one miner's per-scenario best scores
trajrl bench-miner HOTKEY --spec 7

Eval logs (debug + audit)

trajrl logs --validator HOTKEY --limit 20
trajrl logs --eval-id 20260329_1430_w42 --show
trajrl logs --eval-id 20260329_1430_w42 --dump-to ./debug/

A miner eval archive contains:

SKILL.md                                 # miner's product
JUDGE.md                                 # scoring rubric used
metadata.json                            # final_score, delta, episode qualities
world.json                               # scenario context + salt
episodes/episode_N/
  testee_transcript.txt                  # agent's session log
  judge_transcript.txt                   # judge agent's grading log
  evaluation.json                        # per-criterion scores + summary
  episode.json                           # fixtures + instruction

Global options

Every command accepts:

JSON output

Piped output is JSON for every command — handy with jq:

trajrl winner | jq '.current.winner.uid'
trajrl queue --eligible-only | jq '.queue | length'
trajrl validators | jq '.validators[] | {uid, name, version, weightTargets}'

Migration from v1.x

v1.x shipped three binaries: trajrl, trajectoryrl-inspector, bittensor-subnet-inspector. v2.0 collapses everything into one trajrl binary.

New v2.0 commands: trajrl challenge, trajrl winner, trajrl queue (v6 dual-seat winner-challenger).

v2.1 (this release): removed trajrl chain group — it was redundant with btcli and pulled in the heavy bittensor SDK as a hard dependency. v2.1 has no chain-query commands; use btcli subnet metagraph / btcli subnet hyperparameters instead. Result: much lighter install (no substrate-interface, scalecodec, websockets, etc.).

Links

• Subnet repo: https://github.com/trajectoryRL/trajectoryRL — incentive mechanism, evaluation framework
• Bench: https://github.com/trajectoryRL/trajrl-bench — eval sandbox
• Website: https://trajrl.com — leaderboard, live subnet data, skill hub
• Public API: PUBLIC_API.md — read-only, no auth, base URL https://trajrl.com