v2026.1104.0

v2026.1104.0

Released: 2026-11-04.

Agent profiles now connect approved responsibilities, task mappings and bounded evidence checks. Run completion, task acceptance, project progress and business outcomes remain separate records. This payload also corrects the public workspace-request guidance, feedback-bundle authorization and known automatic wake refusals after accepted task edits.

Highlights

• Define and approve responsibilities. Edit success criteria in an agent's profile, request approval, then apply the approved revision. Suggested work-family checks remain editable, and advanced settings are preserved. (#1126)
• Map work beside its recorded criteria. Task Reviews links a task to an applied responsibility and shows mapping history, criteria and profile revision. The mapping does not itself establish acceptance. Mapping-history and profile/performance controls meet the existing 44-pixel phone target minimum. (#1126)
• Check persisted output. Optional exact-document checks compare server-observed persisted content and revision with the owner's expectation. Required automated criteria need evidence; a DONE claim or an old passing harness label is insufficient. (#1126)
• Keep outcomes distinct. Successful execution, criterion evidence, owner acceptance, project progress, use and business outcomes do not stand in for one another. Correct approval blocks, missing work and unverified results stay distinct. (#1126)

Privacy and provenance

• Historical results retain their provenance. Stored scorecards keep the scoring/formula version and evidence provenance captured for that result. Viewer access is checked before evidence projection; old receipts do not bypass current task or project restrictions. (#1126)
• Feedback bundles authorize against database identity. A valid agent UUID can resemble a phone number and be redacted for display. Bundle access now checks the canonical same-company source-run identity before loading logs and transcripts, preserving the intended 404 for unauthorized readers. Phone redaction and document-enabled transcript withholding remain active; unavailable source identity fails closed while document access is enabled. (#1127)

Fixes

• Hosted workspace requests come first. The public MCP page shows the existing hosted HTTP endpoint and its four request tools before self-hosted installation. It distinguishes requesting a workspace from later setup, keeps email verification, claim and access approvals explicit, and explains the model/API-key prerequisite and separate provider usage costs. Signup and pricing disclose that prerequisite, pricing uses the published contact, and invitation help is searchable in the cloud quickstart. (#1125)
• System-created hires reuse configured models. Hires reuse an explicit same-runtime company model, with provenance or a fallback warning. Configuration inheritance does not establish provider-account compatibility or successful model execution. Prompt writes also tolerate a worker closing its input early. (#1126)
• Known automatic wake refusals no longer make an accepted edit fail. A status-only PATCH to /api/issues/:id can keep an existing paused assignee while committing the requested field edit. The canonical typed 409 refusal for paused, terminated or pending-approval agents is now recorded as a withheld automatic wake instead of an unresolved HTTP 500. The agent is not resumed or approved and no run is implied. Direct/manual invocation, assignment restrictions, budget conflicts, unrelated 409s and unexpected-error recovery remain unchanged. (#1128)
• The earlier security note is corrected. v2026.1031.0 retains the high/critical advisory exclusions and their source-level rationale, and states that lower-severity findings remain outside that gate. It does not claim that all advisories were eliminated.

Evidence and limitations

• The isolated deterministic native suite distinguishes correct output, convincing false completion, legitimate approval blocks, missing work and unverified evidence. The fixed rubric passed 47 checks, failed 5 under a controlled evaluator mutation, then passed 47 after repair without changing the rubric. These are bounded orchestration and observable-state checks, not a language-model quality benchmark.
• At code commit c624a240, two isolated upgrade/rollback fixture baselines and database-recovery replay passed. Status-only edits returned HTTP 200 while the assignee stayed paused and no heartbeat run was created. Those receipts apply to that code head; they do not by themselves qualify a different source build.
• Independent agent code review and local automated checks are technical evidence, not human review or calibration. Genuine human labels, semantic quality, useful-work/customer acceptance, external-harness equivalence, production load and causal business outcomes remain unverified.
• Publication is confirmed by the matching stable tag and GitHub Release.
• Reviewer auto-hire follow-through and performance-ledger read scaling remain follow-ups. No production model-readiness or business-performance claim is made.

Upgrade Guide

• No new dependency or database migration relative to stable v2026.1102.1. Back up and rehearse the exact selected payload under the existing release procedure.
• Verify both the published GitHub Release body and the built in-app Changelog. A release-control note added after a payload commit does not prove that the installed source includes that note.

Contributors

Release changes: @thetangstr.

AgentDash is built on Paperclip. Thank you to the Paperclip maintainers and contributors for the inherited control-plane foundation. This credit describes that foundation; the changes above are AgentDash contributions.

Container image

Built from this release's tagged commit for linux/amd64 and linux/arm64.

Tag	ghcr.io/thetangstr/agentdash:v2026.1104.0 (also ghcr.io/thetangstr/agentdash:2026.1104.0)
Digest	sha256:b693399e28c9afe9e1021a71519ef2297fe3310d77d9f34123f93375648c8256

Deploy by digest, which cannot move:

docker pull ghcr.io/thetangstr/agentdash@sha256:b693399e28c9afe9e1021a71519ef2297fe3310d77d9f34123f93375648c8256