Bridge

Use BridgeChatChat to offer ana IntentRequestRequest and stay with the answer as you inspect other AltarControl PanelControl Panel instruments or applications. Its session, or séance, holds the continuing conversation. Each admitted turn opens its own Circle, whose RunJob RecordJob Record lets you follow that one request through to its result.

Open it with BridgeChat in the top masthead. Resize its divider, choose the left or right side, or collapse it while retaining the conversation. Ordinary instrument navigation does not destroy the selected session, its live reconstruction or a later draft. On a narrow window, the panel is a modal sheet; Escape closes it and returns focus to its opener. The layout's remembered width, side and open state contain no conversation text or request bodies.

After completing Altar setup, open http://127.0.0.1:7134/bridge in the dedicated browser profile on the same host. Existing session, RunJob Record and consent URLs open their exact BridgeChat focus. Keep the listener on literal 127.0.0.1; do not proxy, tunnel, or forward it.

Offer one Intent

Choose New Séance, or select an existing session, and submit your message once. Bridge gives the request an identity that an identical retry reuses, allowing the server to return the same canonical RunJob RecordJob Record. If the response is lost, Retry original offering resends that immutable request. The next offering stays disabled until its outcome is resolved: a later refusal cannot erase an earlier unknown outcome.

You can write a later draft while waiting. If the original offering is definitely refused, BridgeChatChat keeps its text beside that later draft so you can restore or dismiss it.

Drafts and unresolved offerings remain separate for each session when you visit another AltarControl PanelControl Panel instrument. They last while this browser document stays open; reload or closing the tab warns before losing them. This is not storage across browser restarts.

Conversations discloses the session rail; Séance details holds secondary identity and session attention. The conversation remains the main reading area. If admission is unresolved or another RunJob Record is unsettled, the composer explains why Offer is unavailable while keeping the later draft editable; that draft is not queued.

The shell's consent count opens Conversations awaiting consent. Choose a marked session to review its requests; its rail and inspector show that session's count. Requests without an available BridgeChat conversation remain in the global count and are reported in the chooser.

When arriving from AtlasProject BoardProject Board, choose a linked conversation or use the conversation chooser. The project return link stays available, including after opening a new séance. Link this conversation to… opens a prefilled AtlasProject Board form for an explicit save; it does not link or send project context merely because you navigated.

The compact CircleTask BubbleTask Bubble beside the turn shows supplied canonical status and current activity. Expand it for full RunJob Record identity and the supplied PatternWorkflowWorkflow revision, with a link to its declared score. Keep that identity when following the result into another instrument. Older retained replies identify missing metadata explicitly. The Circle guide separates these controls from the fuller workspace still being designed.

If a consent card appears, read the pending request and choose one of its supported decisions. The server validates the request, descriptor, reference and decision together. Your choice applies to that pending request. Each InvocationJob RequestJob Request has its own SigilID BadgeID Badge, ContextBriefing PackBriefing Pack boundary and consequences; permission for another object, wider tools or a later effect needs its own admission. Suggesting an action in conversation cannot authorize it.

Read the result

When the RunJob RecordJob Record settles, follow its strip to Orb for evidence. Failed and cancelled RunsJob Records remain visible even if they produced no settled agent turn. In that case, the conversation can contain your admitted message with no answer beside it.

If text from a failed RunJob Record is still projected live, BridgeChatChat marks it as an incomplete answer. A terminal event may already have supplied a retained failure message; the notice does not identify the text's provenance. When a retained agent turn becomes available in the session snapshot, it replaces that live projection. OrbRun InspectorRun Inspector records the presence of failure while keeping private diagnostic reasons private; it neither authorizes an automatic retry nor performs one.

After reload or restart, a turn may be presented as Retained reply rather than its former live activity card. Failed text can remain in history without retaining every live status badge. Use the exact RunJob Record in OrbRun Inspector for canonical status. The fixed trusted source-browser journey has a separate evidence boundary; it does not qualify every casting, ConsentApproval StepApproval Step interaction or ordinary installed browser profile.

Incoming text preserves your place when you scroll into older work. Latest turn returns to the conversation's end and releases a linked RunJob Record focus without changing your draft.

Follow the question raised by the result: Loom shows the declared score, OrbRun Inspector shows RunJob Record evidence, and Nexus shows physical observations. Each instrument keeps its own controls and owners for any action it offers.

When a turn cannot advance

What you encounter	What it means or what to do
Another RunJob RecordJob Record in this séance is still nonterminal	A different message is refused; an identical retry returns the same canonical RunJob Record. Wait for terminal ledger status before starting a new turn.
The activity stream lags or remains open	Use durable RunJob Record status to establish lifecycle state. An open stream, spinner or advancing clock cannot independently establish that work is advancing.
Cancellation is followed by a failed session refresh	The selected session's consent cards and count have already been cleared. Those actions stay revoked until authoritative session truth is available.

The interface currently shows activity without a structured Why waiting? explanation or an evidence-freshness indicator.

What the next turn remembers

In revisions 1 and 2, BridgeChatChat builds the next turn's history from recent, complete, settled turns. It takes the newest whole turns that fit the configured turn and character budgets. Your message may appear in the conversation earlier, but it joins this history only when its agent turn settles. This path supplies neither full-session replay nor ArchiveLong-term MemoryLong-term Memory retrieval. The opt-in public-note revision below instead answers each question independently, without replaying those turns.

How Bridge retains and restores a turn

Only a settled agent turn appends one complete Pydantic AI history unit. Bridge validates typed messages, normalizes provider hops to their owning LychD RunJob RecordJob Record, and keeps a consent return paired with its original tool call. For a later InvocationJob RequestJob Request it rebuilds the Stable Floor for the newly granted capability. Consent resume preserves the current call chain while fitting older settled turns to that grant's bounds. A policy stop retains results of already completed tools and marks the known refused request as denied. The stopped reply does not claim that earlier actions were undone.

Closed, server-validated GenUI fragments survive a terminal refresh. Older rows retain inert compatibility keys and stay outside the current renderer. Replaying an identical settled outcome makes no change. Reusing a RunJob Record and role identity with different visible content, state or fragments is rejected. An older recovery snapshot cannot replace a newer cursor or generation.

Admission has a durable database delivery outbox. Conversation streaming has a narrower lifetime: semantic events and their reconstruction remain process-local, with no durable token or event delivery. The single-active-Run rule does not create a conversation queue across multiple processes. The existing consent round is supported; general multi-approval flows and notifications remain unavailable. The Bridge delivery record gives the current scope.

Investigate public source

Choose Investigate public source when you have a bounded software question and public material for it to examine. Supply the declared repository, base revision and tree/patch digest, your question, literal named public sources, and explicit criteria. State omissions and non-goals so the investigation's boundary is clear, and confirm that every supplied item is public. You supply the material yourself; this operation acquires no repository or file.

The form shows UTF-8 byte limits and preserves the material exactly, without truncation or silent normalization. The fixed Creation investigation uses no tools: it produces an investigation without editing source, executing a workspace, or promoting a change.

Submit once. If the response is lost, Admission unconfirmed retains the original case UUID and body. Use Check this case to read that identity. A 404 can arrive while submission is still in flight, so it does not establish refusal. Replay identical request explicitly resends the original body and is available only while this document still retains it. Timeouts, reconnects, and navigation never retry POST automatically.

To move on, Start different request opens a consequence review. Keep current request keeps the draft and replay body; Discard draft and replay body clears them but leaves earlier uncertain UUIDs visible. Selecting one valid case within this focus preserves the transient state without a loss prompt or automatic read. A missing, invalid, or duplicate selection leaves the focus and retains its loss warning.

The URL retains only the case UUID. Sources and submitted bodies have no durable browser draft; reload warns before discarding transient material and cannot recover an identical replay body. After reload, explicitly Check the case using its UUID. Opening an investigation does not select a different ordinary séance or discard its drafts and consent. Return to conversation leaves this focus, and the global BridgeChatChat link keeps its ordinary session behavior.

Read the question and source alongside the candidate findings and their citations, then examine the separate verifier verdict. Completion tells you the RunJob RecordJob Record finished; judging the report still requires this comparison. The exact admitted RunJob Record opens in OrbRun InspectorRun Inspector for execution evidence.

Load feedback reads the report's separately protected historical judgments: native verdict, external assessment, owner disposition, and feedback status. Keep the external report, owner reason, and annotation coverage distinct. Native acceptance supplies no owner permission, and coverage does not establish semantic adequacy. A newer assessment needs owner review rather than inheriting an older hold. Reload feedback reads later changes. A refused read clears the feedback; unavailable feedback and an unassessed report remain different states.

The detail read may expire an overdue case; the historical feedback read does not. Use Refresh retained case for later case changes. An unavailable case, missing report, refused read and unconfirmed admission are separate states. Viewing appends no feedback and authorizes no continuation, cancellation or successor.

Public-note answers

Use this opt-in profile for questions about explicitly retained public notes on the same host. It requires the PostgreSQL binding and migrations from Summoning, an exact local chat capability, and the existing loopback AltarControl PanelControl Panel access boundary. It does not automatically read your files or save conversations as memory. The delivery record distinguishes portable tests, dated native operator observations and installed operation.

Select revision 3 in the WeaverWorkflow EngineWorkflow Engine configuration, using the exact capability key observed in Nexus. This example's capability is a placeholder; namespace and owner identify one new public-note collection and must remain stable once material is retained.

[weaver.bridge]
revision = "3"
capability_key = "replace-with-the-exact-local-chat-capability"

[weaver.bridge.archive]
profile = "local-public-notes@1"
namespace_id = "9d340d97-7a12-47ed-af2b-c4742ed690f6"
policy_revision = "local-notes-policy@1"

[weaver.bridge.archive.owner_subject]
identity = "local-notes-owner"

Configuration binds the trusted local caller to that owner; the API cannot supply another subject. Changing the namespace's owner or policy revision does not migrate existing notes. Follow the normal configuration lifecycle when changing the selected revision. An already admitted RunJob RecordJob Record keeps its original profile and capability; changed current authority can refuse its continuation.

Manage public notes

Choose Manage public notes in BridgeChatChat. The Configured profile identifies the collection; Current BridgeChat default separately says whether the actual default requests this exact profile. Unavailable and failed reads stay distinct. This configuration establishes neither storage/model readiness, the next routed RunJob RecordJob Record nor an earlier RunJob Record's binding.

1. Choose New public note, supply source key, source revision, producer, observation time with timezone and complete text, then explicitly declare the material public. Prepare public intake freezes the UUID, observation time, exact submitted UTF-8 digest and body for review. Submit candidate once retains a candidate; Promote this candidate is a separate decision.
2. Manage notes shows bounded metadata across all lifecycles. An exact source-key filter and Load more metadata select further pages; the loaded count is not a total. A record link recovers its current metadata independently of the loaded page. Check exact record refreshes that version without replaying an effect.
3. Stage correction keeps the source key fixed and starts with empty replacement text and a new revision label. Supply the complete replacement and a different revision, then prepare and submit it. The prior promoted revision remains eligible until separate promotion supersedes it. Existing text is not read back or prefilled; revision labels imply no ordering.
4. Review revoke displays the exact namespace, source/revision, record, digest and expected version. Confirm exact revoke removes controlled text and retains the anti-reingestion tombstone. Earlier answers and backups are outside that promise.

For intake, Replay identical intake can resend the frozen original body. Use it deliberately: a metadata lookup cannot establish which intake occurrence committed. A lost response therefore remains unknown even when you can read current metadata.

Lifecycle commands have a different recovery path and no replay action. Check the exact record, then make a fresh explicit decision. An unchanged version, missing read, or later refusal cannot prove rollback. A newer version establishes current state, without identifying the author of the attempted occurrence. Refresh and navigation never send writes automatically, and a late response cannot replace a newer selected focus.

Drafts and frozen bodies remain in this browser document only. Departure/reload warns before losing them; reload recovers authorized metadata, not the original body or identical replay ability. Keep unknown request/source/record identities when needed. Current refusal clears protected material. Ordinary conversation drafts, unresolved offerings and ConsentApproval StepApproval Step remain separate; Return to conversation follows the shell's ordinary BridgeChat selection.

The authenticated API supports the same recovery. These reads reveal no note text, search vectors or embeddings:

GET path	Result
/api/v1/archive/profile	The currently configured public-note profile, namespace, owner and policy. This performs no storage or model work and does not prove that retained material accepts that binding.
/api/v1/archive/bridge-default	The actual admission default and its complete managed-profile match, or unavailable. This reveals no other collection and performs no storage, history, readiness or RunJob Record work.
/api/v1/archive/notes	Metadata for every lifecycle, including candidates, archived revisions and revoked tombstones. limit is 1–50; optional source_key matches exactly. Follow the opaque next_cursor with the same source filter; absence returns an empty page, while changed binding or storage failure refuses.
/api/v1/archive/notes/{record_id}	One retained record's current metadata and lifecycle version. A missing record refuses rather than replaying an effect or restoring revoked content.

After an uncertain lifecycle response, read the exact record, inspect its current version and make a new explicit decision. Do not automatically repeat a stale transition. Listing metadata does not make a note eligible for recall or establish which profile an admitted BridgeChat RunJob Record uses.

The existing writes require the same local access and CSRF protection as other AltarControl PanelControl Panel mutations:

POST path	Body and result
/api/v1/archive/notes	Submit the proposal below; retain the returned reference.record_id and reference.lifecycle_version. Intake creates a candidate, which recall excludes.
/api/v1/archive/notes/{record_id}/promote	{"expected_version": 1} for a fresh candidate; makes it eligible and archives a prior promoted revision of the same source.
/api/v1/archive/notes/{record_id}/archive	Supply the current expected_version; removes a promoted note from recall while retaining its text.
/api/v1/archive/notes/{record_id}/revoke	Supply the current expected_version; removes controlled text and retains a tombstone preventing reingestion of that source revision.
/api/v1/archive/recalls	Supply namespace_id, a new UUID request_id, and query; returns a receipt with selected source references, scores and omissions.
/api/v1/archive/recalls/{receipt_id}/resolve	Supply {"receipt_digest": "the-returned-digest"}; returns the exact selected text only while current authority and source lifecycle still permit it.

For example, this proposal's digest binds its exact UTF-8 text, including punctuation. Use a new request identity for new material; an identical retry reuses the original request identity.

{
  "namespace_id": "9d340d97-7a12-47ed-af2b-c4742ed690f6",
  "request_id": "fe6444d6-4267-44e6-afca-c592f3340f31",
  "source_key": "observatory-hours",
  "source_revision": "1",
  "source_digest": "5ed3a3ecd485ad5b2d58439f0d3ca781979499c5e5d230fa6cd4cd5c7997179f",
  "text": "Example note: the test observatory closes at 21:00 UTC.",
  "producer": "operator-public-note",
  "observed_at": "2026-10-11T00:00:00Z",
  "classification": "public"
}

After explicit promotion, offer observatory in BridgeChat to exercise lexical matching against the example. Retrieval uses PostgreSQL simple word matching, not semantic similarity or synonym expansion. A question is bounded to 2,000 characters; a note to 16,000. Bridge recall selects at most five notes within an 8,000-character text budget, omitting whole notes that do not fit. The receipt reports those omissions; database failure is never reported as an empty match.

The answer profile has no model tools and does not replay prior answers. Each request receives the exact selected notes as attributed data, with room reserved in ContextBriefing PackBriefing Pack. The source envelope is transient; receipts retain identifiers and digests. The model is instructed to cite source keys and revisions, but its generated claims still need judgment.

The 19 October observation follows the same question through candidate, promotion, correction and revocation on a real local model. The answer changes with the eligible source revision and abstains when none remains. That one terse lexical case does not establish general question answering or semantic retrieval.

To correct a note, submit a new source_revision and explicitly promote it. Prior receipts then refuse the superseded revision. Revocation also prevents future resolution, including before a pending answer settles, but cannot erase output already generated or streamed. A timeout after a mutation may leave it committed: retry exact intake with its original identity; a stale lifecycle version refuses rather than repeating the transition.

Prepare a promoted note for semantic recall

Open the exact record from Manage public notes. Preparation works on this promoted public revision and the configured generation; source text stays on the server. The exact embedding runtime must already be warm through its Orchestrator owner. Configuration alone proves neither registration, source preparation nor readiness.

1. Check preparation reads metadata for the exact generation and source reference. It performs no embedding. Distinguish not checked, missing generation/derivation, prepared, running, expired, complete, failed and invalidated from a refused or unavailable read.
2. Register configured generation is a separately named mutation of the shared generation. Inspect the configured profile/reference before acting. It performs no embedding; an exact explicit repeat can observe retained active/retired state, but cannot revive retirement.
3. Review this source key/revision, namespace, complete source reference and generation, then choose Prepare this promoted revision. The request may acquire a local embedding lease and derive that exact note. It does not promote the note, activate a model, start a RunJob RecordJob Record or submit a recall. Oversize or changed sources refuse; no truncation or automatic correction occurs.

Read a derivation UUID as the identity of a durable row. Its row and attempt versions, together with server times, describe the latest state. They do not identify a particular browser HTTP call or account for all compute spent. A successful Check therefore leaves a lost preparation response uncertain.

Prepared, failed, or DB-expired state can admit new work even for the same body. Choosing to prepare again may repeat compute; it never retries automatically. Observing expiry neither stops earlier work nor proves its transport has settled. Only the server can establish an eligible completed replay that performs no new work.

Keep the frozen generation and exact source reference while recovering an uncertain call. After a lifecycle change, the newer current reference may not locate historical status; reload cannot reconstruct an older lifecycle version. Leaving, clearing or aborting a read proves no rollback or cancellation. Current authority/profile changes clear protected display and fence late responses. Source preparation and intake/lifecycle writes are locally serialized.

Complete means a retained derivation for that exact row. It promises neither corpus coverage, future eligibility nor a selected answer. Open RecallMemory SeedMemory Seed public sources separately to choose a query and threshold; completion never launches it automatically.

Recall semantic sources

Choose RecallMemory SeedMemory Seed public sources in BridgeChatChat, or open /bridge?notes=recall. This separate view returns attributed public-note sources; it does not generate an answer or admit a RunJob RecordJob Record. Ordinary revision 3 public-note answers remain lexical. Opening the view reads only the authorized configured profile and generation. These metadata do not prove that semantic recall is enabled, that the generation is registered, that notes have eligible derivations, or that the exact model is warm.

Prepare those prerequisites separately through the existing Archive derivation and semantic contract. Intake and promotion alone do not derive a note. This view never registers a generation, derives notes or activates a model. Installed runtime preparation and general retrieval quality retain their delivery boundary.

1. Enter a public query and an explicit minimum cosine similarity between −1 and 1. Similarity locates context; it is not answer confidence or a truth score. Review the result and source character bounds. Query and load exact sources once freezes the exact query, namespace, generation, bounds and a fresh request UUID, then requests a selection and resolves its exact receipt digest under current authority.
2. Read coverage and whole-source budget omissions alongside the ordered sources. Each source identifies its key, immutable revision, record/lifecycle version and similarity. Selection time describes the retained snapshot; source resolution time describes this browser's later read. Partial derivation coverage and an empty selection are distinct from an unavailable model or refused request. Source text is data and grants no instruction or effect authority.
3. If the query response is lost, Check exact request metadata only inspects the same frozen request without embedding or loading text. No receipt yet does not prove that an in-flight query failed. A recovered receipt requires Load exact current sources; correction, revocation or changed authority may refuse the old selection. No automatic query retry occurs.

Query drafts, frozen bodies and resolved text stay in this browser document. Departure warns before losing them; they do not enter the URL or browser storage. Clearing or leaving does not cancel model work, undo a selection or prove rollback. A later explicit submission uses a new UUID and may incur new work. Current refusal and profile changes clear protected display; late responses cannot replace a newer focus. Return to conversation preserves its separate drafts, unresolved offerings and ConsentApproval StepApproval Step.

Answer from a semantic generation

The separate revision 5 uses the configured semantic generation to supply sources to a native answer workflow. Its delivery record separates component, PostgreSQL and actual model acceptance. Enable it deliberately through the normal configuration lifecycle; opening RecallMemory SeedMemory Seed public sources does not change conversation routing.

Use the existing independent archive.profile, archive.derivation and archive.semantic = true selection. Keep the retained namespace, owner and policy unchanged. When moving from revision 3's legacy weaver.bridge.archive configuration, move that same profile to archive.profile; revision 5 does not accept weaver.bridge.archive. Generation registration and promoted-note preparation remain explicit management operations.

[weaver.bridge]
revision = "5"
capability_key = "replace-with-the-exact-local-chat-capability"

[weaver.bridge.semantic]
min_cosine = 0.5
max_results = 5
char_budget = 8000

The threshold is an explicit operator choice, not a confidence calibration. CompositionAppApp requires PostgreSQL, local access and declared exact chat and embedding bindings; it starts no model or preparation. Bridge's composer identifies the configured semantic default without claiming current readiness. Each admitted RunJob RecordJob Record retains its bindings; an older RunJob Record keeps its own revision. There is no per-message mode selector.

Ask about the promoted public notes. LoomWorkflow ViewerWorkflow Viewer shows recall, answer and settlement; OrbRun InspectorRun Inspector follows their actual RunJob Record events. A managed dedicated model can enter bounded hardware waiting through the existing OrchestratorStage ManagerStage Manager. Missing generation, revoked sources and service failure refuse the RunJob Record; they are not successful empty searches. Returned context can be irrelevant, and coverage can be partial. Inspect the answer's source keys, revisions, uncertainty and conflicts rather than treating a similarity score as proof.

If generation reaches its output limit, the semantic request fails with answer_output_limit rather than publishing a normal settled answer or new SDK message history. Text already streamed is provisional and may remain in RunJob Record evidence. Open the linked OrbRun Inspector to inspect the failure; review the configured generation budget before making a new offering. There is no automatic retry or budget increase, and replaying the same admitted request does not create another inference. When a successful current recall selects no notes, the workflow returns a labelled deterministic recall result without calling the answer-generation model. It explains which boundary applied: no notes were searchable, none met the threshold, or the retrieved notes did not fit the character limit. The result includes the receipt's coverage at recall; partial coverage cannot establish an exhaustive search.

Fresh query embedding still consumes its admitted work, and this visible outcome adds no SDK message history. Source or service refusal produces a failed RunJob Record. Even successful completion cannot certify answer truth or universal provider completeness.

This revision has no tools and does not replay earlier conversation history. It resolves the same receipt again before outbound model work and settlement. The full escaped source envelope has a separate 16000-character ceiling; overly large projections refuse rather than truncate. Source text stays transient, while the generated answer and ordinary messages may quote it and persist. Revocation blocks future use without erasing earlier influence. The existing matching ArchiveLong-term MemoryLong-term Memory inspector remains revision-3 selection evidence; this mode does not create a stronger capture or transmission receipt.

Public-page answers

When ScoutWeb FetcherWeb Fetcher is the configured BridgeChatChat default, the composer shows its command instructions and a two-line example. These instructions follow the current default, even in an older conversation; they do not establish that a request is authorized or ready to run. Use Shift + Enter to put the question on the next line:

/read-public https://example.com/page
What does this page claim?

With the opt-in Scout profile, offer /read-public followed by one exact public HTTPS URL, then a newline and your question. Review the complete URL, resolver disclosure, limits and local answer capability on the consent card before approving. The answer may quote the page and is retained in this conversation.

This profile answers one page independently, using no earlier conversation history or model tools. It does not search for URLs, follow redirects or retry a lost acquisition. A refused or expired review grants no contact. LoomWorkflow ViewerWorkflow Viewer shows the exact bridge_chat@4 score; OrbRun InspectorRun Inspector keeps the RunJob RecordJob Record outcome.

Consent history direction

The planned consent history lets you follow a proposed action through review, revisions and its reported result. Related proposals appear as branches, with your decision shown separately from what the executor actually did:

Configuration change
└─ Proposal A
   ├─ Decision: denied
   └─ Proposal B: revises A
      ├─ Review: exact base and diff
      ├─ Decision: approved
      └─ Execution receipt: applied

A file proposal links the exact reviewed diff and artifact revision; a page acquisition shows its destination and limits. Each type supplies its own review details. You can inspect a refused version and a later alternative without approval carrying between them. A missing receipt means the result is unavailable, not that an approved action succeeded. Withdrawing approval cannot undo an effect that already happened.

This tree and its revision/explanation controls are not available in the current BridgeChatChat. The current consent card still judges one exact request; denial does not automatically ask an Agent to revise it. State of Work records the remaining storage, recovery and interface work.

Two crossings still being designed

Pin and Ask would begin a new IntentRequestRequest from an authorized typed reference. Before admission, you would review which reference is being carried, what material would be included or summarized, what is unavailable or redacted, and what permission is involved. The VesselEngine RoomEngine Room would then reauthorize the request. Pinning itself would grant neither retrieval nor mutation authority.

Propose in LoomWorkflow ViewerWorkflow Viewer would turn selected conversation material into an attributed, inert charcoal ScrollWorkflow VersionWorkflow Version candidate against a specified base PatternWorkflowWorkflow revision. A future SpellweaverWorkflow EngineWorkflow Engine draft contract would have to validate it and publish a new revision before it could become executable work. The selection must become a typed handoff; copied prose or canvas coordinates alone cannot supply it.

Neither action is available yet. Follow their development through Loom and the delivery record.

Reading direction

For interface review and design, use the following composition and acceptance case.

Give the conversation the widest readable area so a reader can follow one offered text through its exact RunJob RecordJob Record status to a reply, an explicit missing reply, consent, or unresolved admission. Keep each permission request's actual target and scope beside its decision controls. Conversation selection stays accessible while secondary metadata unfolds on demand.

Judge the composition by what survives a detour: returning from OrbRun InspectorRun Inspector should recover the originating RunJob Record while preserving the later draft and the reader's place. Project origin, saved membership, and material in model ContextBriefing PackBriefing Pack must remain distinguishable. Circle's links should carry the reader to the owning view; its fuller workspace remains a separate design.

Acceptance case: lose an offering's admission response, write a later draft, inspect another instrument, and attend to consent in a different session. Return to the original session with the same request identity and original text, the later draft intact, and correctly scoped attention. Test unknown admission and accepted-Run consent independently; a mock must not suggest two unrelated active offerings are permitted in one session. Browser restart recovery needs a persistence contract beyond the current open-document state. These are presentation targets, not a new queue, memory, or recovery service.