Task Evidence Working Sets

Branch status: intended post-integration behavior for issue #913; this page does not claim Shipped behavior on develop.

The work-board task inspector recalls evidence within its selected goal. Refresh creates a private, bounded packet of citation references and digests in the canonical workspace. The packet contains no copied source bodies. Generic task events retain neutral packet revision/digest and source exclusion identifiers; they do not return claim text, source spans or private paths.

The private task evidence API re-reads canonical source records and artifact files before returning claims. Source spans carry their original owner, version/digest, line location, confidence when known and freshness. Page/row locations remain explicitly unknown when the source does not provide them. Correction and forget use the existing canonical memory APIs. Excluding a source affects this task packet, while canonical deletion/tombstones prevent its return through refreshed packets. A source change invalidates the stored span rather than silently replacing the exact evidence revision previously used.

Refresh and exclusion always create local-only packets. After the spans render, the separate adoption action requires that exact packet revision and digest plus the current task revision. It revalidates every cited source, without retrieving new spans, before recording neutral adoption evidence. Refresh, exclusion and source drift require a fresh review; reset removes model-context adoption.

Sources include operator-approved canonical facts and signed accepted task memory, verified browser extracts, research/document outputs, local encrypted Mail reply drafts and Calendar meeting preparation artifacts. Artifact reads require the owner/goal/task/attempt/run relationship, artifact identity/digest, independent verified readback and a bounded no-follow filesystem read. Mail and Calendar use their existing capability-specific source/connection/consent and message/event checks. Revoked, expired, deleted, unsupported or unavailable sources remain absent with a visible blocked/degraded receipt. This feature does not fetch new provider source data or decrypt credentials/cookies.

GoalSnapshot output belongs to its deterministic child run, while the board attempt links the parent. Evidence resolves only the parent’s canonical board_child_readback to that exact child, using the dispatcher’s named-service, root, fence, goal-owner, authority, input and typed readback contracts. The shared output-artifact selection resolver uses the same proof. Descendant scanning does not authorize outputs.

Stable ownership recovery permits exact selected historical sources to be inspected locally. A fresh task can recall a historical source only when the identity recovery journal explicitly selected it and linked that fresh task. Output artifacts require their own output_artifact selection; selecting a task does not select all of its outputs. Original owner principal/root pairs remain immutable. Recovered task packets are read only; old sessions never become execution principals.

Specify and Decompose place the exact adopted packet in their existing governed task prompt. The packet revision/digest is revalidated inside the provider-contact transaction and recorded as neutral task-use evidence. Local read permission does not grant cloud egress: the operator must explicitly adopt a reviewed packet for task model context. Private or recovered sources remain blocked for the generic strategist purpose; excluding those sources is the available recovery until a reviewed source-purpose egress contract supports that use. Existing capability-specific reply/preparation model consent does not cover this new purpose.

Retrieval uses the existing memory lexical scoring on the CPU and reports lexical_degraded with remote_embeddings_not_used. It introduces no inference provider, mandatory embedding request, separate memory authority or scheduler. Limits are 16 claim spans, 32 artifact candidates/exclusions, 128 bounded canonical fact candidates and 64 KiB per source/private packet. Packet mutation uses task and packet revision checks, and running/archived tasks reject it. Retrieval and inspection report no_learning; only explicit canonical operator correction/forget changes memory.

Focused validation is owned by backend/tests/test_task_evidence.py and frontend/src/components/cockpit/TaskEvidencePanel.test.tsx. They exercise real canonical DB/artifact/API reads, correction/tombstones, exclusion/CAS, cross-owner and symlink/drift rejection, private derived source consent/revocation, exact task prompt binding and inspector controls. Identity-continuity integration requires the matching #900 scope helpers and real recovery tests; branch-local helper absence is visibly blocked. No live provider usefulness or paid inference receipt is implied by these mechanical checks.

The focused snapshot test executes the actual local dispatcher and registered workflow get_goals/write_file steps, then reads its durable child artifact through the evidence API and exact output selection helper. Other source-specific seeded completion receipts remain complementary permission and drift checks.