dsh-agent-swarm v0.1.5 — Team recovery and cancellation safety

This recovery patch restores the usable v0.1.3 feature set after the regressive v0.1.4 integration, and retains the verified task-cancellation safety fix. Existing Git history and the v0.1.4 tag remain available; unfinished work is preserved for separate integration.

Restored and fixed

• Click a member card to open its profile. The Send message action opens that member's official chat, with an entry to return to the original team group chat. Separate profile buttons below member cards are removed.
• Public-message history and bounded message reading are restored.
• The optional Skills manager and its package entry are restored, including immutable candidates, Captain approval, approved-version assignment and actual-load provenance. It remains disabled in the default Bundle.
• Cancelling a task is rejected before mutation if another non-cancelled task still depends on it. This protects the persisted dependency graph; it does not cascade cancellation or repair an already-invalid graph.

Validation and limits

The frozen package was independently compared with the published v0.1.3 tarball. All 610 package files remain present; client.js and skills.mjs are byte-identical. The runtime adds only the cancellation guard. The new guard was reproduced RED and then passed all 9 domain/runtime tests. All 17 model-selection and cold-continuation tests passed after adding the official persistence flush before a disk-read assertion, without weakening its expectations.

The owned 3099 Profile was backed up, checked idle, switched to this exact package and opened in the real browser. Existing group chat, member profile, the correct member chat and return to the same group were verified without sending new model messages. Earlier v0.1.3 real Qwen Skills loading evidence remains applicable to the unchanged Skills artifact; this patch is not a new evaluation of model quality or skill benefit.

The initial recovery CI failure in a warm-consumption assertion was caused by reading before the official asynchronous persistence flush. That test synchronization was fixed. An intermittent Skills-manager supplement failure in the initial local full run is retained in the recovery record; its focused rerun passed. An external P0 controller receipt is NOT_CONFIGURED. Engineering checks and scoped real-Profile verification do not establish the entire planned architecture.

Member-append provisioning, full C2/A1 collaboration and perception, M3 evidence quality, G1 source constraints, and automatic evolution remain incomplete. Restoring this baseline does not declare those workstreams delivered. The damaged workspace and separate feature handoffs remain preserved. Other running Profiles were not switched.

Install or update

Use the supported DSH 0.1.5-rc.2 baseline and verify the attached SHA256SUMS.txt. Stop an idle target Profile and retain its current package, configuration and data before using the official plugin replacement command. See the English README for the exact command and supported runtime.

The optional swarm-skills.example.yml must be configured with an existing provider/model and exact team scope. Its global prompt override supports a Skills manager without a working directory. It contains placeholders, not credentials. Migration of newer experimental schema data and lossless automatic rollback are not established; do not replace live data with an old snapshot.

dsh-agent-swarm v0.1.3 — Approved Skills management

dsh-agent-swarm v0.1.3 — Approved Skills management

This incremental release adds an optional Skills manager to the existing group-chat, member-session and private-memory foundation. It targets DeepSeek Harness 0.1.5-rc.2.

Included

• A separate dsh-agent-swarm/skills Host plugin with its own configured model and durable Session. Captains submit durable requests; the manager investigates only Host-authorized Team scope and captures immutable text candidates.
• Independent Captain review of the exact captured body. An author cannot approve its own candidate. Approved releases bind content, provenance and the reviewed capture.
• Exact approved-version assignment, revision checks and selection of an earlier approved version for rollback. The Team allow-list continues to constrain access.
• The content actually loaded and its provenance use the same frozen version. Changing an assignment does not replace a skill already loaded in the current task attempt, including after official Activation release/resume. The next valid attempt adopts the new approved selection.
• Compatible composition with current-task private-note recall. Member profile navigation and continued private text chat from v0.1.2 are retained.

Install and use

Download the .tgz and SHA256SUMS.txt, verify the digest, and follow the English installation guide. Let active work finish before updating a Profile; preserve the previous package, configuration and persistent state.

The default Bundle does not enable the Skills manager. Add the optional Skills configuration, supplying the exact workspace scope, Team ID and an existing model route. Its manager has only scoped investigation and candidate-proposal tools.

The attached swarm-skills.example.yml combines the optional module and the required global-template override below. Replace its placeholders, then run dsh --profile web --patch ./swarm-skills.example.yml --dump-config before starting that Profile with the same patch.

Required for the standard rc.2 base configuration: its global persona suffix uses {{cwd}}, but the dedicated manager has no workspace. Add this override to the same optional Skills patch so the global template also supports that manager. A workspace-aware agent preset may retain its own working-directory prompt. This changes prompt composition and grants no filesystem or tool access:

- id: system-prompt
  name: '@deepseek-ai/dsh-system-prompt'
  config:
    personaSuffix: Work only within the capabilities granted to this agent.

Captains use agent_swarm_skills_request, agent_swarm_skills_status, agent_swarm_skills_candidates, agent_swarm_skills_review and agent_swarm_skills_assign. Stored assignment is distinct from actual loading and from demonstrated benefit.

For a member's first assignment, bring that member online through a legitimate active task before assigning the approved version. Binding the first assembly to a cold member Session is refused explicitly. Read the current assignment revision when selecting a later version.

Version assignment does not widen the member's original recruitment-time assignedSkills subset. If that subset excludes the skill, the actual load is denied even when an approved version is stored for that member. Keep those boundaries intact and select a member whose existing capability scope permits the skill.

Verification

All four PR checks passed. The complete engineering run passed 1,910 tests, with one existing skipped test, and completed build, artifact and package-install checks. The first engineering attempt had a Windows EPERM during an official JSON rename in one manager test (1,909 passed, one failed); its focused local rerun and one CI rerun passed. The original failure remains recorded as flaky evidence; its external locking cause has not been established.

Independent original QA covered source and built same-attempt Activation recovery, next-attempt selection and rollback. The real acceptance Profile additionally ran a GLM Captain request and a dedicated Qwen manager with exactly two scoped tools, captured a pending candidate and acknowledged the bounded activity batch. The same manager identity and durable pending request resumed across the configuration restart. These are bounded runtime observations; they do not establish every recovery or skill-benefit claim.

The Captain then independently read and approved that candidate and assigned its exact version. An allowed real Qwen member successfully called the official skill tool, received the approved body and produced the required output. A second member with a narrower existing skill subset was correctly denied; its permissions were preserved.

Remaining boundaries

• This slice supports skill text with an empty resource tree. Complete resource bundles, full cross-team observation, automatic evolution and demonstrated improvement on real work remain separate work.
• Selective collaboration C2, related perception A1, memory-quality M3 and authorized observation G1 are not complete in this release. Track follow-up work through the existing issues and roadmap.
• Real A/B review outputs included remaining false findings; the evaluated Skill was not automatically promoted. Completing an evaluation task does not approve its candidate.
• Whole-Host recovery of every in-flight skill, complete upgrade/rollback and migration matrices, long-duration stability and real-model image-chat acceptance are not claimed. The existing real-model private-chat evidence covers text.
• External P0 managed-Team proof ingestion is NOT_CONFIGURED; engineering CI and independently retained runtime observations are separate evidence.

The previous v0.1.2 package remains available as the stable comparison and rollback artifact. This is a usable incremental release, not completion of the entire refactor.

dsh-agent-swarm v0.1.2 — Member chat and personal memory

dsh-agent-swarm v0.1.2 — Member chat and personal memory

This release restores continued conversations with individual team members and adds private-note maintenance with optional recall for the member's current task. It builds on v0.1.1's team group chat, profile navigation, and public-message evidence tools.

What changed

• Continue a member conversation. Open a member's profile, select Send message, and write directly to that same official member Session, including when its managed Captain is unavailable. Failed sends retain the draft. Text and image input are supported; ordinary file attachments are explicitly refused.
• Maintain personal experience. A member can add, revise, invalidate, or replace notes in their own private partition, with task provenance when available, retry deduplication, and preserved history.
• Recall relevant notes for a task. With the Profile owner's explicit configuration, the member receives a small selection of relevant notes for its unique active task and attempt. Recall defaults to disabled; enable it through the Agent Swarm Host setting privateMemoryRecall: active-task. Revocation, task changes, invalidation, and unload clear the current contribution.
• Use the English repository entry point. The README now explains installation, normal usage, release boundaries, and documentation links in English. The repository instructions require English as the primary language for public documentation; existing architecture documents retain their established language.

Personal notes remain private. Relevance, an active note, or a task reference does not prove that an experience is correct. This release does not claim autonomous learning or automatic Skill evolution.

Validation

• The integrated candidate passed all four CI checks: engineering, compatibility, isolation, and policy. Engineering tests: 1,863 passed, one skipped.
• Build and package checks passed. Integrated member-composer, runtime, and private-memory service checks passed 46/46.
• On the official DSH 0.1.5-rc.2 runtime, real browser checks followed group chat → member profile → the exact member Chat. A Qwen member completed two direct human-message replies, including after cold Host startup and after involved Sessions became idle. The same official Session retained each user message exactly once; team task and public-message counts were unchanged.
• Unchanged personal-memory code retains its real Qwen warm/cold, compaction, recall, invalidation, and revoke/unload evidence. These are bounded checks, not a full model or long-load matrix.
• Automated P0 product-proof ingestion remains NOT_CONFIGURED. CI success alone does not establish full product acceptance.

Remaining work

Selective collaboration and shared communication budgets, the independent Skills approval/version workflow, and broader capability/growth acceptance continue in the refactor milestone. Evidence-quality controls and bounded idle memory maintenance are tracked separately in M3 #291. These are not included as completed features in this release.

Automatic upgrade, data migration, lossless rollback, and long-duration stability are not guaranteed. Actual image-model acceptance is not established by the text-chat checks above.

Install or update

Use an official DSH 0.1.5-rc.2 installation and a configured Profile with Storage, Session persistence, Subagents, and a working model provider. Download the .tgz and SHA256SUMS.txt assets from this release.

$swarmPackage = (Resolve-Path './dsh-agent-swarm-0.1.2.tgz').Path
Get-FileHash -Algorithm SHA256 $swarmPackage
Get-Content './SHA256SUMS.txt'
dsh plugin --profile web add --workspace-root $swarmPackage
dsh --profile web --dump-config

Confirm the hashes match before installation. To update an existing Profile, let active work finish, retain its package/configuration/data, stop it, and use dsh plugin --profile web add --workspace-root --force $swarmPackage. Restart and verify the installed version, existing team records, and member chat. See the README for the full flow.

Keep the previous artifact and matching configuration for rollback, and check current data remains readable before resuming. Do not replace newer Session data with an old backup.

Artifact

• Tag: v0.1.2
• Frozen build commit: 3f522cf1adaf437663f795d196267cb7080def5c
• Integration: PR #292
• Package SHA-256: 3B54D1704265EBA4859F723187483AECC0E083C14EF9E64765E507DF4EE2FF12

The versioned tarball is the frozen, verified artifact. Later work is published under a new tag and package.

dsh-agent-swarm v0.1.1 — Group chat and public evidence

dsh-agent-swarm v0.1.1 — Group chat and public evidence

This release delivers team group chat, member-profile navigation, and public-message reading for DeepSeek Harness. It is a usable increment in the ongoing refactor; the complete foundation is still being developed.

Delivered capabilities

• Keep each team's group chat and member roster separate. Select a member to view their profile, open their exact official Session history, and return to the group.
• Let members page through public chat history, read a message by ID in bounded sections, and reply to an actual public message.
• Retain the existing tasks, permissions, goals, membership, and lifecycle behavior. Official Sessions remain the conversation authority.

The frozen candidate passed 1,751 tests with one skip, plus build and installation checks. A real official DSH Profile and browser verified the new flow: a Qwen member read paged and exact public evidence, posted a public reply, and received a GLM Captain's task review. Automated P0 product-proof ingestion remains NOT_CONFIGURED; engineering CI and these bounded runtime checks are separate evidence.

Known limitations and follow-up work

• Continued member chat is incomplete in this package. A member's Chat can become read-only when its managed Captain is unavailable. Profile navigation and transcript viewing work; #286 tracks continued messaging.
• This tarball does not include later personal-memory maintenance and recall, selective-collaboration quotas, or the independent Skills request, assignment, and revision workflow.
• Automatic upgrades, data migration, lossless rollback, and long-duration load stability are outside this release's accepted scope.
• Remaining features are tracked in the refactor milestone. Integration defects stay with their original feature owners.

Installation

The supported DSH baseline is 0.1.5-rc.2. Download dsh-agent-swarm-0.1.1.tgz and SHA256SUMS.txt from this release and compare the package's SHA-256 before installing it. The target Profile must already supply official Storage, Session persistence, Subagents, and a working model provider.

$swarmPackage = (Resolve-Path './dsh-agent-swarm-0.1.1.tgz').Path
Get-FileHash -Algorithm SHA256 $swarmPackage
Get-Content './SHA256SUMS.txt'
dsh plugin --profile web add --workspace-root $swarmPackage
dsh --profile web --dump-config

For an existing Profile, let active work finish, retain the current package, configuration, and persistent data, and stop the Profile. Use add --workspace-root --force to replace the package, then restart and verify the actual installed files and existing team records. No team recreation is required.

For rollback, stop new admission and let in-flight work settle. Restore the retained package and matching configuration, then verify current data is readable before resuming. Replacing current data with an old backup can discard later work; automatic lossless migration or rollback is not provided.

Release identity

• Tag: v0.1.1
• Integrated commit: 4fd47856372942b495fbfeb6f9fe4e1ced7477b4
• Frozen build: d76330eccc41db9fcc3723220ed0d44bbebf9a02; its complete Git tree matches the integrated commit.
• Integration: PR #285
• Package SHA-256: E155A63AEFFFF68A0B3AA36E18F1C30CC8CA62062F10A8C9E9D46C901FF7803F

The accepted package bytes and tag remain unchanged. This edit translates the release description into English; later fixes are delivered in separate versions and assets.

dsh-agent-swarm v0.1.0

First GitHub-installable Team Workbench release. Built from protected main commit ddf7acd after verify, policy, and isolation passed. Install with the official DSH plugin command using the immutable release asset URL.

GitHub main d1ebef2

Fixes #157 and #160: plugin settings registration and compact header team switcher. Asset SHA-256: c81901454b5eebf7427ed40665bcdbb9aeadb32df68b58e6ca7ba0b7ed876141.

E2E package 6cff69b

GitHub main package for the fresh DSH Settings verification. SHA-256: 812a42d3ace62be8f38de5ef7c29930d2532c506bc8fbbf53faabd481ff354b2

E2E GitHub main 02f2987

Immutable acceptance artifact for E2E-GITHUB-CATCAFE-001. Built from a fresh GitHub clone of main commit 02f2987 after green CI.

E2E GitHub main 01e6edc

Immutable package for E2E-GITHUB-CATCAFE-001. Source commit: 01e6edc. SHA-256: e6103090449ad80bfe6ff30af4d63e4482ec8913924c8aff44f4d09ce161029d.