HypervibeDocs
SearchCtrl K Menu

Troubleshoot

How Hypervibe is designed to report refusals, unavailable actions and outcomes that are not confirmed yet, and where to look first when something goes wrong.

Standing
Planned
Release
No public release yet
Last verified
On this page

Hypervibe is designed so that a failure never reads as a success. This page shows how to read refusals, receipts and verdicts, and where to look first.

A refused action says why and stays visible #

A refusal carries one typed code, never a generic sentence. It shows the code in plain words, whether it may be retried, what to do next, how long to wait when the owner sent a time ("no wait given" when it did not), and the attempt and receipt it belongs to. A client that retries on its own says so, and nothing asks you to retry what the client already retries.

An action you cannot take stays on screen, disabled, with its reason: unavailable, not authorized, stale or unsupported. A missing dependency never falls back to a broader credential, a local authority or a sibling service. A refusal never claims an effect that is not known to have happened, and "not found" and "not visible" are one answer, so it never reveals what you cannot see.

Receipts show how far a command got #

Every command carries an operation id, and its receipt is in one of four states. Not confirmed yet is never shown as not done.

ReceiptMeaningWhat to do
… pendingNot finished yetNothing yet; the receipt stays where you can find it
✓ doneThe change is recordedNothing
× not doneIt did not happenRead the refusal; Retry resends the command unchanged
⚠ not confirmed yetThe outcome is unknownLook up the outcome by its operation id before you retry

A retry with the same operation id, the same content and the same action version returns the original result instead of repeating the effect, and anything different under the same id is refused. A timeout is not a refusal: an effect whose outcome is unknown keeps its hold until its owner resolves it.

Verdict, acceptance and standing answer different questions #

A candidate is the exact change a run proposes. Checking it, accepting it and releasing a feature are separate facts.

WordComes fromWhat it tells you
VerdictA verifier that is not the producerPass, fail, inconclusive or error for one candidate; not checked means no verdict exists
DispositionThe receiverAccepted or rejected; a pass alone reads awaiting acceptance
StandingRelease evidenceHow far a feature has come, such as Planned or Qualified
UnknownA value nobody measuredNever shown as zero; unknown cost stays unknown
StaleA source that went quietThe figure stays visible, marked with its age

Common symptoms have a first place to look #

SymptomLikely causeLook first at
Setup refusesAnother operation holds the installation, the release does not match its pin, or the release does not offer the setup you choseThe planned hyper doctor, which compares desired, installed, observed and usable state and changes nothing
A lane seems stuckIt is waiting for capacity or a build, which reads Working with the wait as its detailThe lane's status detail; Idle always shows its reason
You cannot type or answerYour grant is read-only or was revoked, another person holds the input, or the approval went staleThe refusal, which names the act, such as "You can watch this lane but not type into it"; a stale approval is refused, never reopened
Verification failedThe verifier returned fail, inconclusive or error; a timeout reads inconclusiveThe evidence and limitations on the result
A turn will not startThe material it must carry (its objective, acceptance criteria, allowed effects and unfinished obligations) does not fit the model's context windowThe block reason on the turn; that material is never cut to fit
A turn notes that earlier reasoning was not sentThrottle decided the provider could not use it again, for example because the account or the start of the conversation changedThe note on the turn; the receipt records each drop, and the turn continues
A stop has not finishedA stop is recorded as four facts: requested, interruption reported, execution stopped and resources releasedWhich of the four the session's journal shows
A connector action is disabledNo grant, or the connector entry is not defined yetThe reason shown on the disabled action
An imported tool stopped workingIts server's tool list changed, so the action was quarantinedThe reason shown on the action; a person must approve it again
A managed file shows driftA generated file was edited by handBoth versions, then adopt or restore

Setup ends in one of seven outcomes, each with its own exit code, and Install lists them. When the last step is not confirmed yet, the planned hyper setup status checks what happened before anything changes.

A refusal's code leads to its safe repair #

A refusal carries its code as a fact you can copy, such as budget_exhausted. Search these docs for the code to reach its entry below. Each entry says what you see, why it happens, the repair that changes nothing else, and how to tell it worked.

capacity_wait No capacity #

Symptom
Inference is refused, and the refusal reads No capacity.
Cause
No account in the pool is free, or a concurrency limit is reached. Nothing ran.
Safe repair
Wait for capacity to return. A client that retries on its own says so; otherwise the refusal gives the wait, or says no wait was given.
How to verify
The refusal names its cause, scope and retry time. Once the request runs, its record names its route.

budget_exhausted Budget exhausted #

Symptom
Inference is refused, and the refusal reads Budget exhausted.
Cause
Spend reached the budget's limit. The refusal's figure counts settled and held money together.
Safe repair
Wait until the budget resets, or ask an owner or admin of its scope to raise it. Nothing moves to another pool.
How to verify
The refusal names its cause and scope. Once the request runs, its record names its route.

model_not_in_catalog Model not offered #

Symptom
Inside Rocket!, a turn is refused with Model not offered.
Cause
Throttle's capability catalogue does not list the model you chose.
Safe repair
Choose another model with /model.
How to verify
The turn starts, and its request's record names its route.

outcome_unknown Not confirmed yet #

Symptom
A command's receipt reads ⚠ not confirmed yet.
Cause
Whether the command happened is not known, for example after a timeout. A timeout is not a refusal.
Safe repair
Look up the outcome by the command's operation id, then retry unchanged. The same operation id, content and action version return the original result instead of repeating the effect.
How to verify
The receipt reads ✓ done or × not done. Until then, an effect whose outcome is unknown keeps its hold.

Detaching a lane never stops its work #

A lane is one addressable work seat, running an agent or a shell, and it outlives the processes behind it. A Rocket! session is a resumable journal: an append-only record that resumes as the same session after its process is killed. Closing a lane's view detaches it, and detaching never stops the agent; reattaching resumes the same session without repeating an effect. When the program that keeps the lane's terminal exits, the lane reads Holder ended, shown beside its status rather than as a status of its own.

On a team, you can continue from another of your machines. One client types at a time while others watch. Taking input from another person waits for that person's approval, and nothing changes until they give it.

Support bundles are planned to be redacted and bounded #

Diagnostics are planned as an export you choose to make: a bounded, redacted bundle with trace and evidence ids. It carries no credentials, no unrelated workspace data and no automatic replay of effects. Logs carry identifiers and outcomes only, and no screen ever shows a secret value. This manual makes no support promise before a release exists.

Where to go next #

Search the docs