agent

Page size.

Opaque cursor from the previous page's next_cursor.

Page of sessions.

Missing or invalid credentials.

Request failed validation.

What the agent should achieve, in plain words.

owner/slug of a robot repo the goal is about. Validated like Recipe.robot: missing, private and malformed are one 422.

A campaign (C3) the session works inside, by id or slug. C3b: it must be an active campaign of the session's owner that the creator can see (422 otherwise), and it is resolved on every spending call: a round the agent proposes INTO this campaign is auto-approved only when its cap is at or under the campaign's per_run_limit_usd and fits in spend.remaining_usd; anything else waits for a person. Spends that are not this campaign's (a round of another campaign, a standalone run, a coach call) are judged by the workspace's rules, because the campaign's budget does not pay for them.

A total a person pre-approves for the session's spending tools (GPU training on a priced executor, a Modal rollout, a Modal endpoint, a plan, a coach call), on top of the per-run limit: a call is auto-approved only while every approved call's cap, this one's included, fits in it. Null: no session total — the per-run limit alone decides. 0: every spending call asks. Not the model or sandbox spend (that is cap_usd and the allowance).

This session's per-call auto-approve limit, which may only LOWER the workspace's (AGENT_AUTO_APPROVE_PER_RUN_USD, $3.00): a spending call whose cap is above it asks a person. Inside a campaign the campaign's own rules apply instead. Null: the workspace's.

Cap on the session's own model + sandbox spend (AGENT_DEFAULT_CAP_USD, $10, when null). The next model call is refused at it.

The model the loop runs on (AGENT_MODEL, claude-sonnet-5, when null). On the free allowance no other model is offered (422).

Scopes the session's keys carry — the session acts with the highest one named, which may not exceed what the creating credential was granted (403 otherwise). Null: the creator's own.

Who the session is billed to: the creator (null) or an org handle the creator is a member of. The allowance and the working-session cap (AG1) are per owner.

Offer the agent a sandbox shell (when the deployment configures one).

K3 — hub objects the session starts with. Each is checked as the caller (422 at context[i] for missing, private or the wrong kind — never an existence oracle) and summarised for the model's first turn from the hub's own reads: a run's status, gate, config groups and checkpoint gates; a model's card, contract and the run that trained it; a robot's card and interface; a dataset's meta; a rollout's verdict.

Session created; its first turn is queued.

Missing or invalid credentials.

The account's monthly allowance for agent model + sandbox spend is used up.

Authenticated but not allowed (visibility, membership or scope).

State conflict (duplicate handle/slug, wrong repo kind, terminal job, ...).

Request failed validation.

A service this endpoint needs is not configured in this deployment (M0's boot guarantee: the API boots with zero secrets, and an endpoint that needs one says so instead of returning a stack trace).

Agent session id (asess_...).

The session.

Missing or invalid credentials.

Resource not found (or hidden from the caller).

Agent session id (asess_...).

Message recorded; a turn is queued.

Missing or invalid credentials.

Authenticated but not allowed (visibility, membership or scope).

Resource not found (or hidden from the caller).

State conflict (duplicate handle/slug, wrong repo kind, terminal job, ...).

Request failed validation.

Agent session id (asess_...).

Keep the stream open until the session ends.

Resume strictly after this event id (the Last-Event-ID header wins when both are sent).

The id of the last event the client saw (sent by EventSource on reconnect).

SSE event stream.

Missing or invalid credentials.

Resource not found (or hidden from the caller).

Request failed validation.

Agent session id (asess_...).

Agent spend approval id (aappr_...).

Shown to the agent with the decision.

AG1, approve only — "approve up to $X for the rest of the session": this call runs, and from now on compute calls whose caps fit in what is left of $X (this call's cap included) run without asking, even outside the plan; the session's budget_usd is raised to cover it. Must be at least this call's cap_usd (422 otherwise).

The decided approval.

Missing or invalid credentials.

Authenticated but not allowed (visibility, membership or scope).

Resource not found (or hidden from the caller).

State conflict (duplicate handle/slug, wrong repo kind, terminal job, ...).

Request failed validation.

Agent session id (asess_...).

The canceled session.

Missing or invalid credentials.

Authenticated but not allowed (visibility, membership or scope).

Resource not found (or hidden from the caller).

State conflict (duplicate handle/slug, wrong repo kind, terminal job, ...).

Agent session id (asess_...).

The ended session.

Missing or invalid credentials.

Authenticated but not allowed (visibility, membership or scope).

Resource not found (or hidden from the caller).

State conflict (duplicate handle/slug, wrong repo kind, terminal job, ...).

Agent session id (asess_...).

The plans.

Missing or invalid credentials.

Resource not found (or hidden from the caller).

Agent session id (asess_...).

Agent plan id (aplan_...), one version of the session's plan (AG1).

The plan.

Missing or invalid credentials.

Resource not found (or hidden from the caller).

Agent session id (asess_...).

Agent plan id (aplan_...), one version of the session's plan (AG1).

Shown to the agent with the approval.

The approved plan, with the rounds it created.

Missing or invalid credentials.

Authenticated but not allowed (visibility, membership or scope).

Resource not found (or hidden from the caller).

State conflict (duplicate handle/slug, wrong repo kind, terminal job, ...).

Request failed validation.

Agent session id (asess_...).

The decisions.

Missing or invalid credentials.

Resource not found (or hidden from the caller).

Request failed validation.

Agent session id (asess_...).

Lines about what happened after this instant (default the viewer's last_seen_at).

Move the viewer's last_seen_at to now after reading (default true).

The digest.

Missing or invalid credentials.

Resource not found (or hidden from the caller).

Request failed validation.

Agent session id (asess_...).

The objects.

Missing or invalid credentials.

Resource not found (or hidden from the caller).

What the agent should achieve, in plain words.

owner/slug of a robot repo the goal is about. Validated like Recipe.robot: missing, private and malformed are one 422.

A campaign (C3) the session works inside, by id or slug. C3b: it must be an active campaign of the session's owner that the creator can see (422 otherwise), and it is resolved on every spending call: a round the agent proposes INTO this campaign is auto-approved only when its cap is at or under the campaign's per_run_limit_usd and fits in spend.remaining_usd; anything else waits for a person. Spends that are not this campaign's (a round of another campaign, a standalone run, a coach call) are judged by the workspace's rules, because the campaign's budget does not pay for them.

A total a person pre-approves for the session's spending tools (GPU training on a priced executor, a Modal rollout, a Modal endpoint, a plan, a coach call), on top of the per-run limit: a call is auto-approved only while every approved call's cap, this one's included, fits in it. Null: no session total — the per-run limit alone decides. 0: every spending call asks. Not the model or sandbox spend (that is cap_usd and the allowance).

This session's per-call auto-approve limit, which may only LOWER the workspace's (AGENT_AUTO_APPROVE_PER_RUN_USD, $3.00): a spending call whose cap is above it asks a person. Inside a campaign the campaign's own rules apply instead. Null: the workspace's.

Cap on the session's own model + sandbox spend (AGENT_DEFAULT_CAP_USD, $10, when null). The next model call is refused at it.

The model the loop runs on (AGENT_MODEL, claude-sonnet-5, when null). On the free allowance no other model is offered (422).

Scopes the session's keys carry — the session acts with the highest one named, which may not exceed what the creating credential was granted (403 otherwise). Null: the creator's own.

Who the session is billed to: the creator (null) or an org handle the creator is a member of. The allowance and the working-session cap (AG1) are per owner.

Offer the agent a sandbox shell (when the deployment configures one).

K3 — hub objects the session starts with. Each is checked as the caller (422 at context[i] for missing, private or the wrong kind — never an existence oracle) and summarised for the model's first turn from the hub's own reads: a run's status, gate, config groups and checkpoint gates; a model's card, contract and the run that trained it; a robot's card and interface; a dataset's meta; a rollout's verdict.

Why a failed / done session stopped: session_cap, allowance, max_calls, refusal, model_error, credential_gone, canceled; ended (a person ended it) and complete (the agent declared the task complete) since AG1; A1's expired and idle remain on old rows only.

AG1 — the session's title, set by the agent from the goal (e.g. "Omni walk · R4 lateral feed-forward"); null until it sets one.

AG1 — the status line the agent writes whenever it acts (the Home board's second line): what it last did and what waits. Set with its set_summary tool, else the first sentence of its last message in that turn — always the agent's own words.

AG1 — when the session last went to sleep; null while it is awake.

AG1 — when the caller last read the session's digest (per viewer); the digest's default since.

AG1 — the session's latest plan version (approved or not).

AG1 — the plan version in force; null until a person approves one.

AG1 — decisions waiting for a person (plan, spend, shortlist).

K3 — the hub objects the session was started with (AgentSessionCreate.context), as stored.

anthropic or recorded; a replayed session must never read as a real one.

Who pays for the model: platform — Lucen, at list price with 0 % markup, drawn from the allowance. org_key / user_key arrive with bring-your-own-key (A4).

What the session's keys are granted (the highest one is used).

The session's pre-approved total for spending tools; null — the per-run limit alone.

The per-call auto-approve limit in force (the workspace's, or lower for this session).

Where streamAgentEvents resumes after (send it as Last-Event-ID).

AG1: no longer a kill. The end of the session's current window (AGENT_SESSION_HARD_CAP_S, 24 h, from its start or its latest wake): a spend asked for in it expires then, and the sandbox cannot outlive it. A message or a wake opens the next window.

Shown to the agent with the decision.

AG1, approve only — "approve up to $X for the rest of the session": this call runs, and from now on compute calls whose caps fit in what is left of $X (this call's cap included) run without asking, even outside the plan; the session's budget_usd is raised to cover it. Must be at least this call's cap_usd (422 otherwise).

1 for the first proposal; each edit is the next version.

The plan card's title, e.g. "R4, and R4b if needed".

The agent's own paragraph about the plan, quoted as written.

The session's campaign, whose rounds the plan's rounds become; null without one.

"Then" — what follows a PASS (e.g. the 20-seed band scan, a shortlist of up to 2, the robot handoff by consent).

The session-level stop rules in words (its budget, its model cap), hub-generated.

The most the plan can spend on compute — every round's cap, contingent rounds included.

The expected compute cost of the rounds that start at approval, where the planner models them.

The session's budget_usd when the plan was proposed; the plan's cap_usd must fit in it.

A rough figure for the agent's own model turns across the plan (a flat per-checkpoint wake cost), from the free allowance.

The plan version that replaced this one (edited).

The plan decision that asked a person for it.

Shown to the agent with the approval.

Whether this read moved the viewer's last_seen_at to until.

Always the four groups, in the order Needs you, Running, Ready for review, Done.

owner/slug for a model, robot or dataset repo; the id for a run or a rollout.

model_usd + sandbox_usd — what counts against cap_usd and the allowance.

A model call's reservation while it is in flight; 0 between calls.

Model calls made, each one a ledger row.

The header's third line: GPU dollars of the runs this session started — a finished run's cost_usd, a running priced run's cost so far (elapsed seconds x its billed rate, live). Billed on the runs, not on this session; not counted against cap_usd or the allowance.

approval — a spend a person must answer (approval_ids); run / rollout / approval_request / plan / endpoint — a hub resource the agent asked to be woken about (ref, until); user — the agent answered and waits for a message; decision (AG1) — decisions wait for a person (decision_ids); when the agent is also waiting on a hub resource, event names its kind and ref / until / timeout_at describe it.

AG1 — the pending decisions (kind: decision).

AG1 — with kind: decision, the hub resource the agent also waits on (run, rollout, …).

The resource id waited on.

What wakes the agent: finish (the resource ended), gate (a gate verdict on the run, a tripwire, or its end), decision (a person decided a device request), claim (a server claimed the endpoint).

Modal's id for the current container.

When the current active interval began.

Seconds billed so far over every interval.

The sandbox's 24 h hard cap, from its first start.

The only hosts the sandbox's network may reach — the hub's.

owner/slug of the campaign's robot.

The round's short label inside the plan, e.g. R4, R4b; unique per plan.

The round's name, e.g. "lateral feed-forward".

A contingent round's condition in words, e.g. "only if R4's vy±0.10 is still over the band at the scan". It never starts at approval; the agent starts it by key when it judges the condition met, and its reading is quoted on the round (note).

The checkpoint it resumes from; null trains from scratch (or from the campaign's root).

Instead of parent: the key of an earlier round of this plan whose checkpoint it resumes from (the step is named when it starts, and C3's resume-from-PASS rule applies then).

at_approval — created when the plan is approved; later — contingent, or resuming from another plan round.

The gate while training as it will run — the recipe's, else the campaign task's default (L1's poster c_matrix for walk/omni).

estimateRun's own arithmetic on the recipe (with the spec's iteration cap written in).

What the round is expected to cost where the hardware planner can model the template (recommendRun's cost_usd_est, on a priced executor); null otherwise — the cap is then the only number.

The planner's expected wall-clock hours (recommendRun's est_hours), where it models the template.

The most the round can cost — estimate.cost_usd_max (0 on an unpriced executor). Approving the plan pre-approves it.

What to call the parent checkpoint's run (its display_name, e.g. laika-omni-c2-700).

C3's one-variable arithmetic, planned when the plan was proposed (the recipe against the parent's resolved config) — null for a round whose parent is not known yet. A plan whose round breaks a rule is rejected before a person sees it.

What C3 noticed and would not refuse (a stand round with no push gate, a key that cannot be compared).

The round's stop rules in words, from its tripwires and its cap (hub-generated).

The campaign round it became (C3), once created.

Its training run, once created.

The agent's reading of a contingent condition (quoted), a skip's reason (quoted), or why the hub refused it.

One sentence, hub-generated; the agent's words appear only inside quotes.

P1 — the same line for a phone, at most 90 characters: the same facts with the identifiers, the quotes and the subordinate clauses dropped (docs/API.md "The digest's short form"), e.g. "R4 finished 11:06 — 1,600 it in 29 min, $1.60." Hub-generated like text, never the agent's words.

True while the run is running and its cost is still accruing.

The object's id (a checkpoint's is <run_id>:<step>, a plan round's <plan_id>:<key>).

E.g. "R4 c4r4_side_ff", "step_1600", "Render video".

One line of facts, e.g. "9 / 13 · vx+0.30 93 % · video ready".

The state as a word for the row (Succeeded, rendering, pending, not needed).