0.202610091645

Action needed

• claude plugin update for agent-roles and sdlc.

What this release is

The Claude Code auto mode classifier treats a subagent's first prompt as the user's own words and every later message as the coordinator's. A subagent briefed read-only and then resumed to write was denied most of the time, because the classifier refuses to let a coordinator lift a boundary the first prompt set. The orchestrator role and the sdlc skills now brief a subagent with the user's request quoted and every later write named up front, so resuming it for those writes passes the classifier.

agent-roles 0.4.0

• The orchestrator's subagent brief quotes the user's request word for word and names each later write by action and target: "This turn: …, then stop. A later message will ask you to on ; that is part of the request above." It never says "read-only", "change nothing" or "post nothing".
• A resume message names the next step of the task from the first prompt and claims no approval. Anything new the user said since then is quoted word for word.
• When the first prompt forbade a write or did not name it, the orchestrator does that write itself or briefs a fresh subagent.
• On a classifier denial, the orchestrator reports the exact action to the user and does not retry it on standing permission, neither itself nor through another agent.
• Task workers stop at a classifier denial and report the command and reason. They read production data through the read-only client a skill names, unless the task quotes the user asking for another route. They leave a worktree's .venv, node_modules and build directories alone, and never decode or print a token to check authentication.

sdlc 0.21.0

• ticket-attach-docs and wrap-up: the inventory subagent's first prompt names the uploads, the wrap-up comments and the move to Done, and the subagent is resumed for them, instead of being briefed read-only and lifted later.
• mr-babysit: the actor's first prompt quotes the user's invocation and names force-push, fixes, thread replies, resolves and marking ready. Later batches only refer back to it. A command the classifier blocked is not retried by the orchestrator or by a fresh actor. references/actor-brief.md matches.
• write-plan: the implementation subagent, which later becomes the babysit actor, has the babysit writes named in its first prompt.
• DESIGN.md records the decision.

Versions

Plugin	Version
agent-roles	0.4.0
ai-tool-use-validator	0.2.0
code-review	1.16.0
diagrams	0.1.1
gemini-cli	1.2.1
git	0.3.2
glab	1.3.5
migrate-to-uv	1.2.0
noisy-tools-in-subagent	0.6.0
prose	0.3.1
rabbitmqadmin	0.2.0
sdlc	0.21.0
searxngcli	0.3.2
skill-keyword-reminder	1.2.1
teamwork	0.1.0
web-researcher	1.1.1