Orb
In plain words: Run Inspector
Guide to the Run Inspector (Orb) at /orb: browse Job Records, zoom into stations and records, and read the retained evidence of one run with clear labels for gaps and capture limits.
- 1The Job Record owns status; the Inspector only shows evidence and cannot override it.
- 2Evidence is either process local (lost on restart) or durable best effort (may have gaps).
- 3Up to three live run views at once, 100 records per page.
Partial Works on current evidence; native durable evidence service is Designed.
Words on this page, decoded click one for its card
Run Inspector Orb
Look inside one job: what happened, what was recorded, and what is not known.
It's like: A flight data recorder viewer.
Click a dot to open that part's card.
A result raises two questions: which RunJob Record produced it, and how much of that RunJob Record can still be examined? The OrbRun Inspector opens its ordered structural evidence and keeps gaps visible as you read. This act of looking through the evidence is scrying.
Open /orb for the parallel RunJob Record field, or choose Look into the OrbRun Inspector at a BridgeChat result to
inspect /orb/{run_id}. The shell’s OrbRun Inspector tab opens the parallel field; exact RunJob Record and event
links preserve the account being reviewed.
Read the parallel field
The field starts with a bounded page of RunJob Record identities. Load more retained RunsJob Records extends the catalogue; the number shown is the loaded count. Its newest page refreshes every five seconds while visible. Refresh RunsJob Records rereads it immediately, preserving already opened RunJob Record lenses. The canvas can display several workflows together, with at most three live RunJob Record lenses open. Close a lens before unfolding another; its loaded RunJob Record group remains in the field.
Zoom out to see whole RunJob Record groups. Zoom in to disclose Stations, supplied Occurrences, and Records, or choose those same levels directly. Zoom changes detail controls this automatic disclosure. Each lens retains at most 100 records, and its coverage notice distinguishes that window from older retained history. Current manifests supply no nested subworkflow identities; an occurrence or delegated job is not an invented child workflow.
Select a RunJob Record to read its feed beside the canvas. Follow selected RunJob Record follows its projection; manual camera interaction pauses following. Connection state and recorded RunJob Record status have separate labels. Open exact RunJob Record and history enters the selected account below. The semantic RunJob Record list offers the same identities and entry points without requiring canvas gestures.
Establish the account you are reading
Confirm the RunJob Record id, Latest fetched RunJob Record status, and capture class in the header. The RunJob Record owns ledger status. The OrbRun Inspector shows the bounded trace its sources can supply and cannot replace RunJob Record, consent, grant, effect, or artifact truth.
Oculus names the body’s evidence service; OrbRun Inspector is
its AltarControl Panel surface. Native durable Oculus storage is not delivered. Current evidence is available
through this running VesselEngine Room and is labeled process local or durable best effort.
process local evidence disappears when the VesselEngine Room process ends. With durable best effort,
PostgreSQL retains structural records, but failed writes and omitted token deltas can leave
sequence gaps.
Then read the coverage strip:
| Label | Boundary |
|---|---|
| Read at | Response assembly time. Run status, evidence and jobs are read separately; this is not an atomic historical state. |
| retained through | The retained ledger head. |
| loaded through | How far this page has loaded. |
| live updates | ledger poll: updates come from retained evidence, subject to its capture limits. |
Refresh requests a new bounded snapshot. The first page contains up to 100 evidence records; Load more retained evidence requests another page while more remain. If that read fails, the evidence already loaded stays visible; use Retry loading retained evidence to retry the same page read.
The live feed collects newly retained structural records and metadata in bounded batches. On reconnect it resumes from the delivered cursor; if the account resets, a fresh snapshot replaces it. The feed settles once the ledger reports terminal status and the currently retained account has drained. This tells you the available account has caught up, without promising a complete trace.
The live window holds the latest 100 received records. Older retained evidence remains separately pageable. Token text, prompts, output bodies, and private errors are absent. Refreshing or opening a transport connection leaves RunJob Record status unchanged.
Locate the event
Use Browse loaded evidence to choose a Recorded subject. Station observations contain NODE records; Capability dispatches contain DISPATCH records. A station and a capability with the same name remain separate. Counts describe loaded records, not complete subject membership. Show all loaded evidence restores the account. Filtering leaves reported gaps visible; sequence jumps between filtered rows do not imply missing evidence.
For a station, choose a Station attempt to read records with that exact supplied correlation.
No recorded attempt keeps unassigned records visible. The raw occurrence_id field means
station-attempt correlation here, not a scheduled trigger Occurrence. Repeated attempts remain
separate, and an ID recorded under multiple station subjects is flagged without merging them.
Select a row to open Selected historical evidence. Its stable review URL is
/orb/{run_id}?event={event_id}. The inspector shows recorded sequence, producer timestamp, kind,
subject, phase, station-attempt correlation, and capture class. Its historical observation leaves
the header's latest fetched canonical RunJob Record status unchanged.
Previous loaded record and Next loaded record follow RunJob Record sequence across all loaded records, including when timestamps repeat or move backward. Stepping outside a subject or attempt filter restores the full account. The controls stop at the loaded boundaries; load another page explicitly to continue. Sequence and timestamps establish neither elapsed work nor causality.
The subject and attempt selectors and record buttons use native keyboard controls. Opening a record focuses the inspector; Close returns to its row when available, or to Ordered evidence if filtering removed the opener. Changing a filter retains the selected record and labels it when outside that filter. Refresh preserves the selected identity without moving the reader's focus; if the record cannot be found at that read, the requested identity stays in the URL.
Load more pages when a linked event lies beyond the current page. If you exhaust the retained pages without finding it, OrbRun Inspector reports that absence at this read. To investigate a wait, examine its recorded status, phase, subject, station attempt, and transition correlation. The cause may still be unknown; OrbRun Inspector supplies no Why waiting? analysis.
Follow Stasis and return for the recorded wait's owner and return boundary. If the VesselEngine Room has died, continue through Reanimation.
When delegated work exists, open Delegated jobs below the evidence and selected-record detail.
The disclosure shows at most the newest 32 summaries, each with
its AgentJob identity, runtime, CoffinSandbox profile, station, status, bounded-result presence,
artifact-reference count, and up to the latest 64 lifecycle events. Older omitted jobs and
truncation are explicit. A linked job absent from this bounded snapshot remains visibly unverified;
OrbRun Inspector does not choose another job in its place. Prompts, output text, and private errors are not exposed.
Follow a relation only where it was recorded
Bridge returns to the owning conversation and focuses the exact RunJob Record’s turn when available. An unavailable turn is reported. Exact PatternWorkflow opens Loom only when the RunJob Record's pinned manifest validates and equals the entire registered revision. The graph uses that same exact declared score when available. Its RunJob Record and selected-event query preserves the return destination; neither changes PatternWorkflow identity.
The fixed trusted source-browser journey checks this exact RunJob Record→PatternWorkflow→RunJob Record relation and retained failed/useful history across a normal reload. It does not establish complete trace capture, every evidence interaction or general browser/accessibility acceptance.
A selected consent record with a retained reference offers Open exact consent in BridgeChat. The reference records the historical wait; it does not establish that the request is still pending or that you may judge it. Bridge focuses only the exact request for that RunJob Record in its current conversation snapshot. A missing, settled, or unavailable request stays explicit; another request is never substituted. Return to OrbRun Inspector event returns to the same RunJob Record and selected record.
A selected event with exact transition correlation offers Open transition in NexusModel Status Board and carries the selected event id. The link supplies a relation worth inspecting; viewing it requests no mutation. Nexus keeps any lifecycle decision separate.
Selection, folding, and pagination leave RunJob Record state unchanged. Open /orb to compare another RunJob Record,
or follow it from BridgeChat. The current OrbRun Inspector offers no annotation, retry, cancellation, approval,
publication or transition action.
Revisit a recorded moment
In an exact RunJob Record, Record cursor, Previous record, and Next record move through loaded structural evidence. The graph shows observations through that sequence; future observations and AnimatorModel Runtime selections disappear. Entering playback freezes that loaded lens, so incoming events and live-window eviction cannot rewrite the moment being inspected. LIVE or Follow latest returns to the latest loaded account. Refresh reads a fresh account and leaves playback.
The AnimatorsModel Runtimes rail names only identities supplied by retained dispatch records. Its stable colours match dispatch nodes, and a glow marks a grant at the viewed cursor. A grant proves which AnimatorModel Runtime was selected, not how long it was running or whether it remained available. Older records without this identity remain unknown; current NexusModel Status Board observations never fill in historical detail. The parallel field shows these same supplied identities for its selected live RunJob Record.
Hover or keyboard focus previews a graph node's meaning without changing selection. Escape closes the preview. Glow marks the latest record or playback cursor, not continuous execution. The RunJob Record status above the graph remains current. Record spacing represents sequence, not elapsed time; gaps, missing records, unretained tokens, and unloaded pages still limit the account. This is a record cursor, not full-workspace wall-clock replay or a configurable retention service.
Inspect a matching Archive selection
In an exact RunJob Record's ArchiveLong-term Memory selection inspector, choose Read ArchiveLong-term Memory selection. Nothing is fetched until that gesture. Refresh ArchiveLong-term Memory selection reads again; a failed read clears its protected result and offers Retry ArchiveLong-term Memory selection. Run navigation aborts and fences old results independently of graph updates or playback. Read time and receipt creation time remain separate. Long identities and digests open through native disclosures.
The separate same-host API read is
GET /api/v1/archive/runs/{run_id}/recall. Use the authenticated local API under the configured
ArchiveLong-term Memory management authority; permission to view a RunJob Record alone does not disclose its source metadata.
The retained RunJob Record must identify the supported exact BridgeChat ArchiveLong-term Memory PatternWorkflow and its admitted profile.
The response distinguishes not_applicable, not_recorded and recorded. A recorded receipt may
have zero selected notes; that successful empty selection differs from a failed or denied read.
It retains ordered source keys/revisions, exact record/lifecycle identities, scores, character and
result bounds and omitted_for_budget. These describe that selection: its character bound covers
selected note text, not the complete model ContextBriefing Pack or provider token accounting.
The capture class is matching_retained_request: the retained receipt matches the request's
identities. Because RecallMemory Seed accepts caller-supplied identities, that match can exist before the
workflow reaches recall. It therefore establishes a selection account, without proving consumption
or provider transmission. The read neither reconstructs the released ContextBriefing Pack cache nor exposes
the prompt, source body, answer, or provider payload. Source names and digests remain protected
local metadata.
A later correction, archival or revocation does not rewrite the old receipt. Current source eligibility requires a separate authorized check; inspecting history neither resolves former content nor makes that revision eligible again. Storage and binding failures remain errors rather than an empty selection. Memory owns receipt authority and Observability owns the capture boundary.
Witness without possession
A richer view can reveal more of the retained account. It cannot fill the missing part of a history. Frontend's folding law keeps groups, expansion, and sequence faithful to what their sources supplied.
Reading direction
For interface review and design, use the following composition and acceptance case.
Place the selected account's qualifications where the reader judges it: Run identity, canonical status, capture class, snapshot time, retained and loaded boundaries, and omissions stay beside the graph and records. Summaries show the supplied subject; selection opens exact identity and fields. Run-record timestamps include a date and use local time. A verified linked job opens its disclosure and receives focus after the evidence and selected record, where a long job collection cannot bury the account's limits.
The diagram must distinguish declared permission from supplied record membership and observed traversal. Sequence and equal spacing establish order, not elapsed time or causality. Unknown subjects and gaps stay visible; job events keep their own sequence unless the server supplies a RunJob Record correlation. Fold counts describe loaded membership. Deeper expansion needs supplied identities under the folding contract.
Acceptance case: open an event beyond the first page, distinguish not-yet-loaded evidence from an unretained interval or failed read, inspect the exact score or correlated physical transition, and return to the same selected event using only the keyboard. A narrow display should start with a readable list. Efficient direct event seeking needs a separate retrieval contract; the current path remains loading subsequent pages. The shared renderer's cross-engine lifecycle qualification and wider accessibility acceptance remain recorded separately in State of Work. A graph and live feed establish neither complete capture nor execution control.