xrun v0.7.1

Added

• xrun install skill --codex|--claude installs repository-local agent
  instructions for Codex and Claude.
• Installers now install the Python TUI by default, check for Python 3.11+
  and pip, and can try ensurepip via --install-pip / -InstallPip.
• xrun update checks GitHub Releases, prompts before installing, and runs
  the official installer. Interactive xrun startup also prompts when a
  newer release exists; set XRUN_NO_UPDATE_CHECK=1 to skip that check.

Changed

• Installer and README links now target the master branch.
• The xrun agent skill text is harness-neutral instead of Claude-only.

---

xrun v0.7.0

Pluggable metric sinks. WandB joins MLflow as a first-class fan-out
target; the poller now mirrors one xrun run to N tracking servers in
parallel, with per-sink failure isolation. Comet ML and TensorBoard
plug into the same trait in v0.8 with no further refactor.

Skipped 0.6.x — the field-feedback ramp was already absorbed into 0.5.4
and the next milestone deserves a clear minor bump.

Added

MetricSink trait + WandB

• xrun_core::metric_sink::MetricSink (async via async-trait):
  name / open_run / log_metrics_batch / log_artifact / finalize.
  Backends are independent crates; adding one is a single
  impl MetricSink for X plus a registry entry in MetricFanOut::new.

• OpenRunCtx<'_> carries everything a tracking server needs
  (run_id, experiment, run_name, vendor, instance_id, config, tags),
  borrowed so callers don't clone hashmaps per tick.

• RemoteRunHandle { sink_name, remote_run_id, remote_url } — the opaque
  token. remote_url lets xrun show / xrun metrics --mlflow-url
  build deep links without rebuilding from raw config (which broke when
  the user later edited mlflow.url).

• MetricSinkError granular enough to retry on Network / Server,
  fail loud on Auth / Config, swallow on Disabled.

• crates/xrun-wandb/ — new crate. WandbClient wraps three operations
  on api.wandb.ai:

  • viewer { entity } GraphQL probe for default-entity resolution.
  • mutation upsertBucket to create / reuse a run row (idempotent on
    run name, so a daemon restart doesn't double-create).
  • POST /files/{entity}/{project}/{run}/file_stream to append to
    wandb-history.jsonl and signal complete + exitcode on finalize
    (same path the official wandb Python SDK uses internally).

• WandbSink: MetricSink groups MetricPoints by step before posting
  so multi-key updates within one training iteration land on the same
  wandb x-axis tick. Tracks an AtomicU64 history offset per run for
  retry-safe appends.

Poller fan-out

• MetricFanOut (was MlflowMirror) holds Vec<Box<dyn MetricSink>>.
  start() opens runs on every enabled sink in sequence; log_metrics
  fans batches out concurrently within one block boundary, so total
  latency is max(per_sink) not the sum. Per-sink warned latches
  surface the first failure and swallow the rest.
• MetricSinksConfig { mlflow: Option<MlflowSubConfig>, wandb: Option<WandbSubConfig>, … } — toggle each sub independently.
• Sinks are gated by both [metrics] sinks = […] (config) and
  credential presence; a sink listed without creds warns and is dropped
  (no fail), so a partial setup still gives the user whatever sinks
  remain — including local-only.

Persistence

• Schema migration 005_sink_run_ids.sql:
  • mlflow_run_url (resolved at sink-open time).
  • wandb_run_id, wandb_run_url (paired write — never half-populated).
• Store::set_mlflow_run_url, Store::set_wandb_run setters.
• MetricFanOut::start routes the new RemoteRunHandle back to the
  right setter via a sink-name dispatch.

CLI

• xrun init --wandb-key - flag, mirrors --vast-key / --kaggle-token
  with - to read trimmed line from stdin (one stdin read max per
  invocation).
• xrun config probe --vendor wandb — lands in the existing probe
  surface, mirroring vast / kaggle / mlflow shape. Reads
  XRUN_PROBE_WANDB_KEY from env so the wizard / TUI can validate a
  pasted key without writing it to disk first. Hits viewer { entity },
  returns 401 cleanly on a bad key.
• xrun init-manifest --vendor <X> --sink <Y> — generate a manifest
  skeleton for any (vendor × sink) combination. Every editable spot
  is marked TODO_<field>; grep TODO_ <path> lights up everything
  needing review before launch. Vendors: vast, kaggle, local,
  ssh. Sinks: mlflow, wandb, none. Closes the loop on the
  user-facing goal — Claude Code can produce a working manifest
  without knowing the schema.
• xrun init --sink wandb now accepted (was a v0.8 rejection).
• xrun config show includes wandb.api_key (masked unless
  --secrets); xrun config set wandb.api_key … works through the
  schema-driven setter.

TUI

• New screens/sinks.py (Python Textual). Cards: MLflow, WandB,
  Comet ([v0.8] disabled). Status pill with five distinct states:
  EMPTY / PAUSED / CHECK / READY / ERROR — PAUSED makes the
  "key set but not in metrics.sinks list" mistake visible at a glance.
  Actions: edit / test / toggle-default / revoke. Routing: g m chord
  • Go: Sinks palette entry.
• SinkEditScreen with masked inputs: 4 fields for MLflow (url goes
  to global config, auth fields to credentials.toml), 1 for WandB
  (entity is auto-probed at first launch).
• Wizard catalog: wandb available_now flag flipped to True. Wizard
  step still treats wandb as "tick to add to default list, configure
  key on Sinks screen" — full wizard form is a follow-up.

Schema

• WandbCredentials { api_key } in Credentials.
• Run carries mlflow_run_url, wandb_run_id, wandb_run_url
  (Option<String>).

Changed

• mlflow_mirror.rs → metric_fanout.rs. MlflowMirror →
  MetricFanOut, MlflowMirrorConfig → MetricSinksConfig.
  Poller::with_mlflow → Poller::with_metric_sinks. The wiremock-
  based poller integration tests still validate the same wire-level
  MLflow calls (create / log-batch / update-run) — they don't care
  that the path goes through a trait now.

Tests

• 11 new unit + wiremock tests in xrun-wandb covering viewer-probe,
  upsertBucket, file_stream batching, finalize exit code, 401 auth
  surfacing.
• 8 unit tests in init_manifest covering the (vendor × sink) matrix
  and the TODO_ token contract.
• 1 multi-sink integration test in xrun-poller runs the poller
  end-to-end against two wiremock servers (one MLflow, one wandb) and
  asserts both received the metric batch.
• Full workspace: 404 passing, 0 failures across 50 suites.

Live smoke

• Run 01KQWJSX0VDNEEMBQGPK8KH6HN — quickstart.yaml with
  metrics.sinks = ["wandb"]. WandB run created, finalized, page
  exists at https://wandb.ai/data_force/quickstart/runs/xrun-…
  (HTTP 200), DB row populated with wandb_run_id + wandb_run_url.
• xrun init-manifest --vendor local --sink mlflow parses cleanly
  through xrun doctor --manifest <path>.

Not yet shipped

• WandB artifacts API — log_artifact returns Disabled for now.
  Needs the separate manifests + S3-presigned PUT surface; deferred
  to v0.7.x patch or v0.8.
• Wizard form for wandb (just api_key input) — checkbox enables
  the sink in metrics.sinks but credential entry still goes through
  the Sinks screen.
• Comet ML — slot in TUI catalog, no impl. v0.8.

---

xrun v0.5.3

Field-feedback sweep: closes the eight items in ISSUES.md from the
arborust evening session, plus one latent Store::open bug surfaced
while wiring live telemetry. End-to-end live metrics + events on Kaggle
now work via the existing MLflow side-channel.

Added

• xrun pull is no longer a stub — resolves run → vendor adapter →
  adapter.pull(), defaults destination to runs/<id>/artifacts/,
  supports --into, and reports matching files for --ckpt best|latest|all. (#4)
• Live telemetry on Kaggle: xrun_hook's log streamer now also tails
  events.jsonl / metrics.jsonl and pushes them as MLflow artifact
  chunks alongside stdout. The Kaggle adapter ingests new chunks every
  poll tick — events and metrics appear in xrun events/xrun metrics
  while the kernel is still running. Requires mlflow.url configured.
  (#8)
• Streamer also mirrors metric records to MLflow's native
  /api/2.0/mlflow/runs/log-batch endpoint so the MLflow UI's Metrics
  tab plots them. The artifact JSONL is still the source of truth for
  the local poller; the native logging is purely for the human
  dashboard. (Without it, the UI shows "No model metrics recorded".)
• Kaggle dataset version pinning: after datasets status = ready, the
  adapter resolves the dataset's currentVersionNumber via the REST
  API and rewrites slugs to owner/name/N, so kernels never mount a
  stale snapshot due to the kernel-creation cache lag. Slugs that
  already carry an explicit /N are left alone. (#1)
• TUI Artifacts screen: walks the run's local artifacts dir and lists
  real files (was previously a xrun pull not yet implemented stub).
  Empty state hints press \a` to pull; a` runs the real CLI and
  auto-refreshes. (Surfaced after #4 unblocked it.)

Fixed

• xrun launch --detach no longer hangs after kaggle kernels push
  finishes. The push subprocess's pipes are now drained on background
  threads, so try_wait returns instead of waiting on a full OS pipe
  buffer. (#3)
• Duplicate running:start events after a poll-daemon restart. The
  Kaggle adapter rehydrates its last kernel state from the DB on the
  first poll so already-emitted transitions don't re-fire. (#6)
• TUI ⚠ stale warning for healthy runs whose poll-daemon died
  mid-session: the app now runs auto-resume every 60 s, so a poller
  killed by a binary upgrade self-heals without user action. (#7)
• xrun_hook install path on Kaggle: the wheel is base64-embedded in
  the kernel main.py and bootstrapped before user setup, so the
  resolution-atomic pip install no longer drops siblings when
  xrun_hook isn't on PyPI. (#2/#5)
• Latent: Store::open was being called on the data directory in
  three Kaggle adapter callsites (post-pull ingest, kernel-state
  recovery, live telemetry ingest), failing silently with
  unable to open database file. New db_path() helper appends
  runs.db; without this the live-telemetry ingest never persisted
  any events or metrics.

Changed

• crates/xrun-kaggle/src/log_stream.rs exports
  parse_chunk_seq_with(stem, ext, path) so adapters can reuse the
  same chunk reassembly path for logs/, events/, and metrics/
  prefixes.

---

xrun v0.5.1

Release 0.5.1

xrun v0.5.0

Added

• xrun-local adapter — run manifests directly on the host as a first-class
  vendor (vendor: local). Full lifecycle parity with vast (provision →
  upload → execute → tail → pull → destroy), cross-platform (bash/sh on Unix,
  pwsh/powershell on Windows), idempotent destroy via PID file.
• xrun-ssh adapter — same lifecycle against your own server / NAS / VPS over
  SSH; reuses the local execution model.
• xrun init — first-run wizard. TTY → spawns xrun-tui --wizard (4 steps:
  local capabilities → vendors → logging mode → recap with live xrun doctor).
  Non-interactive flags for the Claude skill / CI: --probe-local --json,
  --non-interactive --mark-completed --sink mlflow. Credential flags
  (--vast-key, --kaggle-token, --kaggle-username, --kaggle-key) accept
  - to read one stdin line, so secrets stay out of shell history.
• xrun doctor categorized output: env / vendors / manifests groups with
  per-category counts, --json for skill consumption.
• xrun config probe --vendor <name> — validate credentials read from
  XRUN_PROBE_* env (used by the wizard).
• xrun metrics --per-key — emit one chart per metric key instead of overlay.
• xrun_hook.metrics(values: dict, step: int) — batch shortcut: writes one row
  per key sharing a single timestamp. Avoids N separate metric() calls in the
  training loop.
• Kaggle live log streaming end-to-end via xrun_hook → MLflow chunked
  artifacts (the public Kaggle API has no live-log endpoint).
• exp/templates/ — starter manifests + train.py for common ML tasks:
  quickstart (zero-config smoke test), classification
  (loss/acc/f1_macro/precision/recall), regression (loss/mae/rmse/r2).
  Templates run end-to-end without torch so they smoke-test the structure
  before adaptation.
• TUI Vendors screen: brand-coloured cards (vast orange #ff6b35, kaggle cyan
  #20beff via border-left thick accent), status pills
  (READY/CHECK/ERROR/EMPTY), pulse animation on the connectivity probe dot,
  double-click opens edit.
• TUI Vast edit screen: "Region filter" section reusing the existing
  CountryExcludeScreen — pulls and saves search.exclude_countries via
  xrun config, surfaces current exclusions inline.
• TUI country picker: pills tinted by continent (EU blue, AS red, NA green,
  ME orange, AF violet, OC cyan, SA amber) so country codes render and group
  visually in any terminal — including Windows Terminal which doesn't compose
  regional-indicator codepoints into flag glyphs.
• TUI u binding on Vendors screen — opens the vendor's quota/billing page
  in a browser (kaggle.com/settings, cloud.vast.ai/billing).
• TUI wizard split into a sub-package with image_view / metrics_view /
  report_view widgets for reuse in run detail.

Changed

• Wizard rebuilt for keyboard-first UX: Checkbox widgets (Tab/Space toggle),
  o opens API-key page of focused card (works before selecting), Esc-skip
  now requires Y/N confirmation, probe shows a loading indicator, Recap runs
  xrun doctor --json and prints ✓/⚠/✗ per check. Toggling no longer rebuilds
  the body — pasted keys keep focus.
• TUI Kaggle card surfaces free-tier limits (CPU ∞ 24h/session,
  GPU 30h/wk 12h/session, TPU 20h/wk 9h/session) instead of an opaque
  "competitions visible: N" probe artifact.
• TUI vendor cards: flat Horizontal rows replaced with Vertical cards;
  vendor-row/vendor-dot/vendor-name-col CSS retired in favour of
  vendor-card/vendor-card-head/vendor-card-foot. Old class names removed.
• xrun gc filters non-vast records — local cleanup is via xrun stop.

Removed

• xrun init --vendor flag. It was informational-only (echoed in JSON, never
  wrote anything). The wizard now relies on --sink and the credential flags
  for non-interactive setup.

Configuration

• [ui] wizard_completed: bool (default false) — TUI auto-launches the
  wizard when false; finishing the wizard sets both this flag and
  [metrics] sinks via the CLI.
• [metrics] sinks: Vec<String> (default ["mlflow"]) — editable through
  xrun config set. WandB and Comet sink checkboxes are visible but disabled
  with a [v0.8] badge.

---