probe-research (probe SDK/CLI + probe-research plugin)

CLI + SDK client for Probe Research, Probe's experiment-tracking platform. It is a thin client over the v3 ingestion contract (CONTRACT.md in the Probe Research backend). Implemented experiment calls map onto real endpoints.

Two client surfaces

Probe Research exposes experiment tracking through two separate surfaces over the same backend, for two different workflows:

• probe — SDK + CLI (non-agent). A Python library (import probe) and the probe command-line tool for integrating with existing setups and manual experimentation. Drop it into a training script or pipeline to record runs, metrics, spans, and artifacts. No agent required.
• probe-research — plugin: skills + MCP (agent-centric). Installed into a coding agent (e.g. Claude Code). Its skills teach the agent the experiment workflow, its read-only MCP server lets the agent query experiment state, and writes flow through the probe CLI. This is the surface for agent-driven research loops such as Anthrogen.

Same backend, two entry points: humans-in-code reach for the SDK/CLI; agents-in-the-loop use the plugin.

Package boundaries

src/probe/
├── sdk/       # typed client, uploads, local capture, session adapter ABI
│   ├── fluent.py    # probe.init/log/finish — ambient run, contextvar-bound
│   └── analysis.py  # client.compare(): N runs aligned on step
├── cli/       # `probe`: thin shell over the SDK
└── mcp/       # `probe-research-mcp`: strictly read-only tools and resources
skills/
├── start-research-work/
└── track-research-work/

The SDK is the implementation. The CLI, MCP source adapter, future hooks, Python experiments, and passive platform integrations all use it. CLI and SDK therefore have capability parity; they differ only in ergonomics.

Session commands do not upload metrics or experiment outputs. They correlate a coding-agent session with a run and checkpoint redacted transcript metadata. No hooks are installed in this release.

Install

pip install -e ".[dev]"     # from this directory

Auth

probe login       # browser device flow (RFC 8628 + PKCE) — the default; nothing to paste

Air-gap paste path: probe login --token probe_pat_xxxxxxxx (verified via GET /v1/me); probe login --endpoint-only --base-url … saves the endpoint without minting a token. Both write ~/.config/probe/config.json. Or set env: PROBE_BASE_URL, PROBE_TOKEN (user token, /v1), PROBE_INGEST_TOKEN (ingest token, /ingest), PROBE_HMAC_SECRET (optional body-signature secret). SDK-created runs heartbeat every 60s so the server can reap crashed ones; PROBE_HEARTBEAT_SECONDS tunes the interval (<=0 disables).

You can also skip probe login entirely: the first client.run() / probe run start with no token triggers the same browser approval inline (TTY only) and persists the result. Disable with PROBE_AUTO_LOGIN=0; headless/CI keeps the crisp AuthError and should set PROBE_TOKEN.

The MCP server prefers PROBE_MCP_TOKEN, which should be a separately minted read-only token. It falls back to PROBE_TOKEN for local development, but exposes no mutation tools.

On rented compute (RunPod) with no standing config, the /start-research-work skill seeds PROBE_TOKEN at session start.

SDK (agent-driven / interactive)

Two forms, same implementation. The module-level one when you want to log from code that has no handle to pass around:

import probe

probe.init(
    project="folding",
    experiment="dockq-sweep",
    hypothesis="temp 0.7 wins",
)
probe.log({"loss": 0.42, "dockq": 0.71}, step=42)    # from anywhere, any thread
probe.finish()

probe.init() takes everything client.run() does and returns the same Run handle, so with probe.init(...) as run: works and the rest of the API is one attribute away. The binding is a contextvar backed by a process default: a worker thread finds the run (a bare contextvar would not — threads start with an empty context), while a second init() inside a thread or block shadows rather than hijacking the outer one. That is the part of wandb.init()'s global that silently corrupts concurrent runs. probe.active_run() returns the current binding, and a script that exits without finish() is closed at exit as completed / failed / canceled rather than left to the crash reaper.

The explicit form has no ambient state at all:

import probe

client = probe.Client()  # resolves creds from env / `probe login`

# run() resolves by default. `hypothesis=` is the one opt-in to creation, so the
# FIRST run in a new experiment says what you expect to see:
run = client.run(experiment="dockq-sweep", hypothesis="temp 0.7 wins",
                 name="run-1", project="folding",
                 description="DockQ baseline at temperature 0.7",
                 source="runpod", external_id="rp-9931")

# …and every run after that is bare — the experiment already exists, and its
# hypothesis is first-write-wins so reopening never rewrites it:
run = client.run(experiment="dockq-sweep")

# No hypothesis and no experiment? That is a project-direct run, which is the honest
# home for work with none — better than an experiment named after your cwd:
run = client.run(project="folding")

# `name` defaults to a timestamp (the server adds a petname short_id). A slug that is
# a near-miss of an existing one is REFUSED, not created: a warning is invisible from
# a training loop. Creation is SDK-only — `probe run start` never creates, because on
# the CLI the slug is hand-typed every time, which is where typos come from.

client.update_run(run.id, name="DockQ baseline",
                  description="Stable reference run")
child = run.child("retry-1", relation="retry",
                  description="Retry after fixing the data loader")
run.snapshot()                                   # non-disruptive git + deps + GPU capture
run.link(wandb_run_id="abc123", s3_prefix="s3://x/y")

for step in range(100):
    run.log({"loss": ..., "dockq": ...}, step=step)     # POST /v1/runs/{id}/metrics

# Omit `step` and it auto-increments per metric kind, so the bare loop shape
# still produces a curve. `step=None` explicitly means no step at all
# (wall-clock axis) — that is what the CLI and the passive importers pass.
for batch in loader:
    run.log({"loss": loss})                             # steps 0, 1, 2, …

# Values of any type are accepted. Numbers (and bools, numpy scalars, 0-d
# tensors) become metric points and plot; strings, dicts, lists and None go into
# that step's record and read back through the trajectory view.
run.log({"loss": 0.4, "phase": "eval", "cfg": {"lr": 3e-4}})

# Below-run coordinates: `coords` are bounded grouping axes (series identity:
# rank/split/..., never a per-sample id), `labels` per-sample drill-down ids
# (point identity only). Everything logged inside the unit carries them, and
# nested units merge (child wins per key).
with run.unit(coords={"rank": 0}, labels={"sample": 3}):
    run.log({"reward": 0.71}, step=12)   # -> dimensions={"rank": 0}, labels={"sample": 3}

with run.span("rollout", name="rollout-0", step_index=1) as span:  # trajectory span
    span.attributes["reward"] = 0.8      # closes with ended_at + a terminal status,
    ...                                  # `failed` if the body raises; nests inside
run.log_artifact("final.sif", uri="r2://bucket/final.sif", kind="artifact")
run.finish()                                     # flushes spool, sets status+ended_at

Structured knowledge and local process capture use the same SDK:

run.execute(["python", "train.py", "--config", "dockq.yaml"])
client.events.add(run.id, "decision", "Use DockQ scorer v3", evidence_refs=["tool:91"])
report = client.check_run(run.id)

Data writes are fail-open by default: on failure they spool to disk (~/.local/state/probe/spool) and return, never blocking the training loop. run.finish() (or probe flush) replays the spool. Appends and queue rewrites are fsync'd and atomic. On rented compute, put the queue on durable storage with PROBE_SPOOL_DIR=/shared/probe/spool or probe --spool-dir /shared/probe/spool …. Pass strict=True to make a write raise.

Miles tracking backend (probe.connectors.miles)

Drop-in wandb-parity backend for Miles' TrackingManager. Zero miles commits: the registry is a plain dict checked against args flags at init, so activation is two lines in your launcher, before init_tracking(args):

from probe.connectors.miles import register
register(args)          # registers the backend + sets args.use_probe

Every TrackingManager.log() lands with its own step counter (train/step and rollout/step map to step_index per key, the counter entry itself is stripped), values arrive after Miles' DP-rank reduction — exactly what wandb sees — and the run declares its labeled-point plan (num_rollout x rollout_batch_size x n_samples_per_prompt) so later per-sample capture never trips the server's default budget mid-training. Config: PROBE_BASE_URL / PROBE_TOKEN, optional args.probe_experiment / args.probe_run_name (fall back to the wandb names). Fail-open end to end: a broken tracker never costs a training step; non-finite values are dropped per-point. Per-rank and per-sample detail is enabled without replacing aggregate logging by passing:

--custom-rollout-log-function-path probe.connectors.miles.per_sample_rollout_log

The hook logs labeled sample points through the same durable queue as Miles' aggregate tracking.log() calls. Every point carries metric_scope=sample, the Miles sample id, an optional group id, and—when the sample carries Harbor's returned capture external_key—the exact deterministic rollout span. The dashboard can therefore separate aggregate and sample points while resolving sample → trial → trajectory/sandbox without a Miles-core change.

Reward and effective response length are captured by default. Applications can publish arbitrary numeric sample measurements without connector changes by putting an inline dictionary on the sample:

sample.metadata["probe_metrics"] = {
    "agent/input_tokens": input_tokens,
    "agent/output_tokens": output_tokens,
    "quality/custom_score": score,
}

If values already live elsewhere on the Miles sample, declare metric-name to dotted-path mappings on the same args object Miles sends to RolloutManager:

args.probe_sample_metrics = {
    "agent/observed_tokens": "metadata.agent_metrics.observed_tokens",
    "agent/turns": "metadata.agent_metrics.turns",
    "agent/tool_calls": "metadata.agent_metrics.tool_calls",
}

Stock Miles launchers that do not expose arbitrary rollout args can put the same inline mapping in a tiny importable hook module instead:

# my_project/probe_metrics.py
from probe.connectors.miles import make_per_sample_rollout_log

per_sample_rollout_log = make_per_sample_rollout_log({
    "agent/observed_tokens": "metadata.agent_metrics.observed_tokens",
    "agent/turns": "metadata.agent_metrics.turns",
    "agent/tool_calls": "metadata.agent_metrics.tool_calls",
})

Then set --custom-rollout-log-function-path to my_project.probe_metrics.per_sample_rollout_log. This stays entirely outside Miles source while preserving the shipped hook's durable queue, sample labels, and Harbor span linkage.

Missing, non-numeric, boolean, and non-finite values are omitted; an explicit numeric zero is retained. Configured paths override a same-named metadata["probe_metrics"] entry. The run reserves 1,024 sample metric points per sample by default; set args.probe_sample_metric_budget higher when a sample schema intentionally exceeds that.

Per-rank detail otherwise rides the capture-at-source arc (run.unit + capture_trial). Upstreaming a native --use-probe flag into a miles fork is optional polish (registry docstring's own recipe).

SDK (install-once / passive push)

client.ingest(
    project_slug="protein-folding",
    experiment_slug="dockq",
    run={"name": "r1", "source": "temporal", "external_id": "wf-1", "status": "running"},
    metrics=[{"kind": "model", "key": "loss", "value": 0.5, "step_index": 1}],
    batch_id="deadbeef",          # idempotent redelivery
)

One idempotent push (bearer ingest token + optional HMAC), keyed on (customer_id, source, external_id).

SDK (reading runs back for comparison)

comparison = client.compare(experiment_id=exp_id, keys=["dockq"])
aligned = comparison.aligned("dockq")

for label, values in aligned.values.items():
    plot(aligned.steps, values, label=label)     # or aligned.to_pandas()

Name the runs with run_ids=[...] or select them with the filters list_runs takes (experiment_id=, group_id=). One POST /v1/series/query per 50 runs — more than that batches rather than truncating, because silently dropping runs 51+ reads as "these are all of them". Columns are labelled by the server's petname short_id; runs of differing length keep None holes rather than being cut to the shortest, since differing length is usually what is being compared. pandas is optional and only touched by .to_pandas().

There is no separate read client. wandb.Api() is a distinct object because W&B has two transports (a service process for writes, GraphQL for reads); one REST transport does not need the split, so this is a shaping layer on Client.

CLI (probe)

probe project create folding
probe experiment create dockq --hypothesis "temp 0.7 wins" --project folding
RUN=$(probe run start --experiment dockq --name run-1 \
        --project folding --source runpod --external-id rp-9931 \
        --description "DockQ baseline at temperature 0.7")
probe project patch folding --name "Protein folding" --description "DockQ studies"
probe experiment set EXPERIMENT_ID --name "DockQ sweep" --description "Temperature sweep"
probe run set $RUN --name "DockQ baseline" --description "Stable reference run"
probe snapshot $RUN
probe link $RUN --set wandb_run_id=abc --set gpu_job=rp-9931
probe log $RUN loss=0.42 dockq=0.71 --step 42
probe span add $RUN --type rollout --name rollout-0 --step 1
probe artifact add $RUN ./final.sif --kind artifact
probe notes show                      # the project's notes
probe notes write --append ./note.md  # free-text markdown, one file per project
probe exec $RUN -- python train.py --config dockq.yaml
probe run check $RUN
probe run end $RUN --status completed
probe bundle $RUN            # read: run + series + artifacts

Harbor trial capture (probe trial)

Capture a Harbor trial directory into a run, keyed to the training step — the sandbox↔step join (see docs/2026-07-15-harbor-native-ownership-plan.md for status: what's shipped vs parked):

# rollout span + reward metric + labeled CAS file uploads + kind=harbor_trial
# manifest; a recognized trajectory format (ATIF v1.x built in) also expands
# into turn/tool_call spans under the rollout span
probe trial add $RUN jobs/my-job/trials/swe-fix__x1 --step 600 --env-type skypilot-fork
probe trial add $RUN <dir> --step 601 --no-expand      # raw-only capture
# Copy/checksum a host trial tree onto a durable volume without touching the network.
probe trial stage <host-trial-dir> --to /shared/probe/trial-601 \
  --expect result.json --expect lock.json
# Retry one or every Miles bridge request. The descriptor supplies run/step/correlation.
probe trial export /shared/probe/trial-601/export-request.json
probe trial drain /shared/probe/captures
probe trial watch /shared/probe/captures --interval 5
# Bind descriptors produced during an offline run initialization:
probe trial drain /shared/probe/captures --run "$PROBE_RUN_ID"
# retroactively expand a stored trajectory (e.g. after a fork's parser ships);
# deterministic span ids make this idempotent — re-runs upsert, never duplicate
probe trial expand $RUN <manifest-artifact-id> --max-spans 0

Query it back: client.list_run_artifacts(run_id, kind="harbor_trial", step_from=599, step_to=601). Fork parsers plug in via probe.connectors.atif.register_trajectory_parser("their-format", fn); unknown formats are captured raw (never rejected) and expanded later.

probe trial stage and the probe-harbor-export/1 consumer keep an atomic .probe-capture.json beside the durable trial bytes. Its collection status is separate from remote-upload status, so an exporter outage leaves a precise, retryable list of unconfirmed files instead of losing their paths. Stable external keys and span IDs make retries update the same rollout; arbitrary Miles/Osmosis correlation fields are preserved under the harbor_trial manifest's source.context. The exporter additionally promotes sample_id and group_id to the Probe point labels sample and group, respectively, on the reward and harbor_trial manifest. Those labels preserve distinct same-step samples without creating a separate metric series.

Trajectory sources — Harbor-first, by decision. Today trajectories enter Probe through Harbor's on-disk contract: ATIF-supporting agents write trajectory.json into their logs dir, the harness delivers it at <trial_dir>/agent/trajectory.json (the location Harbor's own viewer reads), and parse_trial/capture_trial pick it up from there — raw bytes always stored, recognized formats expanded into turn/tool_call spans, unknown formats expandable retroactively via probe trial expand. Emission is per-agent opt-in upstream (SUPPORTS_ATIF; e.g. the oracle agent emits nothing), so absence is a normal, captured state. "Traditional" trajectory tracking — live SDK/OTel-style span streaming from instrumented agent code, the W&B Weave/MLflow-Tracing model, with no file contract at all — is planned as the second door: the server-side turn/tool_call span rails it needs already exist, so it is an SDK-instrumentation arc, not a schema change. For now the scope is deliberately the Harbor framework.

The completeness claim is intentionally bounded: it covers declared regular files in the host Harbor trial directory. Public Harbor tears down the sandbox before Trial.run() returns, so a post-run SDK consumer cannot know about undeclared state Harbor never materialized. The ledger reports that state as unknown, inventories explicitly declared missing files, and treats hidden files/symlinks as visible skips. A true pre-teardown guarantee requires the producer or environment implementation to invoke durable collection from its lifecycle hook.

The following commands are reserved for future hook configuration and are not part of the normal researcher workflow:

probe hook session attach RUN --session-id SESSION --transcript-path PATH --cwd DIR
probe hook session checkpoint RUN --session-id SESSION --transcript-path PATH --reason pre_compact
probe hook session detach RUN --session-id SESSION --reason session_end

They currently encode session links in run.metadata.agent.sessions[] and transcript checkpoints as redacted local-reference artifacts. Until managed artifact upload exists, transcript portability remains explicitly false.

Read-only MCP server

Run the stdio server with probe-research-mcp. It exposes three tools, plus five deprecated aliases that are removed next release:

research_context, research_search, research_get, research_compare and research_resolve have been REMOVED. They answered as deprecated aliases for one release; that window has closed. Calls to them now fail as unknown tools — use the three above.

Thin harness, fat skills. Coverage grows through get_entity's view and filters parameters, never through more tools. browse_research is the one addition that cleared that bar: it answers a question the others structurally cannot, because search ranks by relevance to a query and therefore needs you to already know what to search for.

get_entity(ref, view=..., filters=..., token_budget=..., cursor=...), where ref is run:<id>, experiment:<id>, artifact:<name>, project:<id>, group:<id>, or a bare id:

card (the default) returns available_views for that entity, so one call tells you what else you can ask for — the matrix above is documentation, not something to memorise. An unrecognised ref kind is rejected outright rather than guessed at: the resolver used to try every getter in turn, so a retired kind surfaced whichever backend parse error came first instead of saying the kind was gone.

Artifacts resolve by NAME, because the reuse check has a name and not an id — and because no GET /v1/artifacts/{id} route exists. The lookup runs against the SHARED, lab-wide level, which is where an official artifact is promoted to and the nearest thing to the tenant-unique names the retired asset registry enforced. get_entity(ref="artifact:<name>", view="versions", filters={"requirement": ">=2"}) is where research_resolve went. A name that does not exist raises not-found; a name that exists with no satisfying version returns state="no_match" with the versions that do exist, so you can see the real ceiling. Requirements match monotonic integers and labels, not semver — ">=2.0" is rejected rather than silently matching nothing. A name carried by more than one shared artifact is a 422 naming both ids: the duplication this check exists to prevent has already happened, and picking one silently would compound it.

trajectory reads a run's spans (the run bundle carries span_type counts only). metrics returns series summaries, and filters={"key": "<key>"} drills to raw points. reproduce resolves env_ref through its execution record. token_budget bounds the row-shaped part of a view and hands back a next_cursor; reproduce is atomic and reports token_budget_exceeded rather than truncating a manifest into something that reproduces nothing.

There is no trace-file tool: no backend trace index has ever existed, so it answered matches: [] to every query, which agents read as "this file has no lineage". To trace a path/URI/hash, use search_knowledge (its exact channel matches artifacts) and follow get_entity(view="lineage").

MCP reads through the Probe Research API—never directly from Postgres or R2. Its logical sources are control identity/tenant scope, the structured experiment store, the artifact/manifest registry, the one-index search door (POST /v1/search: exact SQL channel + the KB engine's semantic channel; search capabilities are discovered against the live backend with one cached probe), and object-store resource pointers returned by the API. W&B, RunPod, Kubernetes, Git, and local transcript paths are not live MCP sources; adapters upload their identifiers and evidence first.

Skills

Two skills, split by moment rather than by entity. probe run start opens a run in an experiment that already exists, or project-direct with no experiment at all; it never creates. Creation on the CLI is always probe project create / probe experiment create.

• start-research-work covers that call: orient against what already exists and what is already running, create the project and experiment explicitly — first, before the scaffold, since the notes hang off a project and have nowhere to go until one exists — resolve artifacts before creating them, and snapshot code + env before launch. It is re-entered through the session rather than run once.
• track-research-work covers what a run produces: metrics, spans, artifacts, version pinning, reading back what actually landed, and closing with the real lifecycle outcome. The project's notes are the exception to the run requirement — free-text prose that needs no run. Its reference.md holds the artifact command syntax, the publication sequence, and project admin.

Reuse hooks are deliberately deferred. start-research-work contains the reuse-before-create rule; deterministic enforcement can be added later without changing the SDK, CLI, MCP, or skill contracts.

What maps to what (v3 endpoints)

v0.4.0.0 ingestion fold-in (Phase 1)

Most earlier gaps are closed by Probe Research v0.4 (PR #13). Now wired:

• Real metric dimensions. log_hw(..., device=3, host="n1") sends dimensions (fold #9); log(..., dimensions={...}). No more key-encoding.
• Presign artifact upload. log_artifact(path=...) runs presign → PUT to R2 → confirm (fold #16), carrying kind/meta so byte uploads are labeled like reference artifacts (Harbor-ownership Phase 0). Fails open to a reference on error.
• Execution records. snapshot() posts a content-addressed execution-record (fold #7); client.execution_record(...).
• Artifact versions. client.list_artifact_versions() + create_artifact_version(). The separate asset registry that used to sit here was folded into artifacts (research-os #143/#144): an artifact is a named thing in a container with a chain of immutable versions, which is what an asset was.
• Experiment versions. client.experiment_version() mints the immutable manifest (fold #6). This replaces the removed run-level promote.
• Lineage edges. client.add_edge() / run.edges() (fold #2).
• foreign_keys. first-class on the ingest path (run['foreign_keys'], fold #8) and surfaced on reads (run.foreign_keys, run.short_id).
• Events read. client.events.list() / for_run() (server-emitted lifecycle log).

Remaining

• MCP semantic/KB search is now wired to POST /v1/search (workspaces+kb fold-in) with an honest keyword fallback on older backends; transcript evidence is not indexed yet. Session hooks remain later work.
• Harbor-native ownership Phases 1–3 (trial capture connector, capture-at-source, platform surface): see docs/2026-07-15-harbor-native-ownership-plan.md.

(Previously listed here and since shipped: RunPatch foreign_keys/env_ref parity, asset materialize, upload kind/meta, and server-side artifact list filters ?kind=&step_from=&step_to=.)

Typed models (generated from the OpenAPI contract)

Request/response models are generated from the backend's OpenAPI schema, not hand-written, so the client cannot silently drift from the contract. The write paths (log/span/log_artifact/ingest/edges/execution-records) build their payloads through the generated models, so a renamed or removed field fails client-side instead of as a server 422. /ingest/v1/runs is now declared in the schema too (Probe Research PR #12), so the passive push is generated and validated like every other path.

• schema/openapi.json - a snapshot of Probe Research's FastAPI schema.
• src/probe/_generated/models.py - generated, never hand-edited.
• src/probe/models.py - the stable import seam the SDK uses.

Refresh when the contract moves:

make regen        # dump-openapi (RESEARCH_OS=../../research-os) + gen-models
# or step by step:
RESEARCH_OS=/path/to/research-os python scripts/dump_openapi.py
python scripts/gen_models.py

RESEARCH_OS points at a local checkout of the Probe Research backend source repo (directory name research-os); it is only used to regenerate the schema snapshot.

CLI grammar note

The CLI is built on typer. Connection flags are global and go before the command: probe --token probe_pat_x log RUN loss=0.1. probe login also accepts them directly (probe login --token ...).

Tests

pytest        # 29 mocked/unit tests + a real-git snapshot test; no live server