Skip to content

Harden adaptive reasoning router lifecycle - #3

Draft
blackjose007-stack wants to merge 1 commit into
mainfrom
agent/adaptive-router-v2
Draft

Harden adaptive reasoning router lifecycle#3
blackjose007-stack wants to merge 1 commit into
mainfrom
agent/adaptive-router-v2

Conversation

@blackjose007-stack

Copy link
Copy Markdown
Owner

What changed

  • Add a deterministic v2 work-order classifier for MEDIUM, HIGH, and EXTRA_HIGH.
  • Keep HIGH as the stable default; require every low-risk condition before selecting MEDIUM.
  • Route architecture, migration, security, release/signoff, destructive work, owner decisions, contradictory evidence, and repeated failures to EXTRA_HIGH.
  • Distinguish requested host effort from effort that the host actually confirmed and applied.
  • Enforce valid START, transition, BLOCKED, RESUME, and END lifecycle states.
  • Add v2 hash-chain checkpoints, tail-truncation detection, restrictive file permissions, and credential redaction.
  • Preserve readability of existing v1 decision records.
  • Document the guarded local-loop-council / Qwen routing handoff without expanding local-model authority.
  • Update the README, skill instructions, agent metadata, and community documentation.
  • Expand unit coverage for classification, lifecycle validation, compatibility, integrity, permissions, and redaction.

Why

The original skill described stable reasoning allocation but left several runtime boundaries implicit. This update makes classification deterministic, prevents invalid or oscillating lifecycle transitions, avoids claiming a host effort change that did not actually occur, and strengthens the audit trail without adding a daemon or database.

User and developer impact

  • Small mechanical tasks can select MEDIUM only when all required conditions are explicit.
  • Normal engineering work remains on the stable HIGH default.
  • Release, migration, security, architecture, or other high-consequence work receives EXTRA_HIGH policy and broader validation.
  • Existing v1 logs remain usable; new entries use the stricter v2 schema.
  • Private logs continue to avoid raw prompts, source code, and credentials.
  • The skill can generate a bounded local Qwen workflow when local-loop-council is installed, while Codex retains acceptance authority.

Validation

  • python3 -m unittest discover -s tests -v — 14 tests passed.
  • PYTHONPYCACHEPREFIX=/private/tmp/adaptive-router-pycache python3 -m compileall -q skills/adaptive-reasoning-router/scripts tests — passed.
  • git diff --check — passed.
  • Installed skill source matched the repository source; only an ignored local __pycache__ directory differed.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants