AIHAUS FLOW v1.11.0: One run per task

What changed

A task now runs from start to finish without stopping between phases. It stops
only when a decision you have not settled is open. Under a standing workflow
opt-in such as Claude Code's ultracode, analysis, review, and execution no
longer split into separate runs.

• The confirmation stop applies only to a non-empty list.
  • If the agent assumed nothing, it says so and continues.
  • Ambiguities that the spec review finds also go on the list.
  • Only your answer to the list it showed you confirms it. A general
    go-ahead does not.
• Assumed decisions: accept. in a request accepts the agent's
  assumptions in advance, for that request only. The agent records them on the
  task, puts them in the final review brief, and reports them with the result.
• Host runs. A standing host opt-in counts as asking for a workflow run.
  • A host run never selects or changes the mode.
  • A host run carries the mode's sequence, gates, and limits.
  • A mode's consecutive steps run in one host run. The run ends early only
    for one of the reasons listed in the harness, for example an open
    decision, a reached limit, a safety pause, or the first remote write.
• normal and pure.
  • In normal, the session writes the change in its own turns, not through
    workflow agents or forks. The host run, if any, starts after that
    implementation and covers the review, the fixes, and the rechecks.
  • pure uses no host run, fork, or subagent.
• Forks and independence.
  • An agent that inherits the author's conversation is never a separate
    context. This covers a fork, the author continued by message, and an
    advisor.
  • A fresh agent started on the committed candidate SHA is a separate
    context, even inside the same run.
  • An agent that inherits the session's model fills only roles whose
    profile model is the session's.
• Fix loops.
  • Continuing the author's conversation, by message or fork, counts as
    resuming it.
  • A fix and its recheck may share a run.
  • An amended spec needs a new spec review.
  • Spec and fix cycles count across runs. Past the limit, the work pauses
    until you direct otherwise.
• README.
  • Hosting a CLI executor or reviewer: wait for the process's own exit code;
    a stopped or timed-out call yields no verdict; on Windows, the hosting
    agent commits for a Codex executor.
  • Claude Code notes: ultracode, forks and CLAUDE_CODE_FORK_SUBAGENT=0,
    and the per-prompt ultracode keyword to keep cost bounded.

There is no new role, room, tool, hook, or contract file. Project memory
templates are unchanged.

Migration note

• Delegated implementation in normal. Under a standing host opt-in,
  normal no longer delegates implementation to workflow agents. To delegate,
  request Mode: workflow.
• Standing opt-ins. A standing host opt-in now counts as asking for a host
  run. To keep host runs per request, turn the opt-in off. In Claude Code, that
  means not turning ultracode on for the whole session, and typing the
  ultracode keyword in the prompts that should use a workflow.

Install or update

Run from the root of the repository that should use aihaus:

npm exec --yes --package=https://github.com/overdrive-dev/aihaus-flow/releases/download/v1.11.0/aihaus-flow-v1.11.0.tgz -- aihaus init --target . --check --json
npm exec --yes --package=https://github.com/overdrive-dev/aihaus-flow/releases/download/v1.11.0/aihaus-flow-v1.11.0.tgz -- aihaus init --target . --json

Project memory is preserved, including your Assignments and Autonomy
sections. Customized installations should run the --check preview first
and follow the runbook in
INSTALL-VIA-LLM.md.
The .sha256 asset supports an independent archive integrity check.

This release was built in normal mode with delegated execution:

1. A read-only workflow of 7 agents diagnosed the problem from the aihaus
   contracts, the Claude Code docs, and field transcripts.
2. The owner decided two open questions.
3. Codex gpt-6.1-sol xhigh implemented the change.
4. A separate gpt-6.1-sol xhigh review found two P2 findings.
5. A fresh executor fixed both, and a recheck ended with ship.

CI ran 107 contract tests on Linux, macOS, and Windows. In headless ultracode
lab runs, workflow mode used one workflow instead of one per phase, and
normal implemented in the session before a single review run. Those lab runs
were stopped before they finished.

What's Changed

• Note execution modes and maintainer profile in runbook by @overdrive-dev in #65
• Note task intake, branch deletion, and CI queue in runbook by @overdrive-dev in #66
• One run per task under host workflow engines (v1.11.0) by @overdrive-dev in #67

Full Changelog: v1.10.0...v1.11.0