Wukong Code v0.0.16

Wukong Code 0.0.16 — Converge Safely

Released: 2026-07-21

Distribution: free/BYOK native binaries for macOS, Linux, and Windows.

0.0.16 makes the existing Loop converge around a fixed user-owned target. It
does not add payment, hosted execution, managed model credits, PR publishing,
Team features, or a multi-reviewer quorum.

Loop convergence

• A new Loop freezes the user objective and optional completion criterion in an
  append-only local contract. Writer, reviewer, and strategist cannot rewrite it.
• Each review receives the previous blocking findings and must mark them
  resolved, still_open, or replaced_by before it can pass.
• Blocking findings must map to a frozen criterion. Stable keys and fingerprints
  exclude prose, absolute paths, and line numbers.
• Two consecutive identical blocker sets trigger one fresh-context, read-only
  strategist. If the same blocker remains afterward, Loop pauses with
  NEEDS_WORK/no_progress instead of spending the remaining iterations blindly.
• Invalid structured reviewer output retries once on the same model. An explicit
  reviewer model that remains invalid or unavailable falls back to the writer
  model once. Cancellation never triggers fallback.

Permission Mode v2

• Manual remains the interactive confirmation mode.
• Auto no longer asks questions or silently grants them. It allows ordinary
  autonomous work but blocks workspace-external writes, sensitive targets,
  high-risk Bash, and unclassified custom/MCP tools. An active Loop pauses
  with NEEDS_WORK/permission_required.
• YOLO is an explicit dangerous approval bypass, not an unlimited capability.
  User deny rules, safety hooks, Plan guard, profile allowlists, and read-only
  child-agent ceilings still apply.
• Headless prompts default to guarded Auto. Dangerous headless execution requires
  --yolo --yes; --yes by itself no longer selects YOLO. The legacy
  --auto-approve flag emits a migration warning.

Resume, Roles, and TUI

• Welcome may show one opt-outable current-workspace suggestion from Codex,
  Claude Code, or Cursor, bounded to 30 days and 50 candidates per source.
• The Wukong session picker can search user prompts from the latest 100 sessions
  using bounded, cancellable local reads. System, tool, and synthetic content is
  not indexed.
• Role Profiles can declare input, output, and capability contracts. These
  contracts only narrow the base profile and tool set.
• Planner, reviewer, strategist, and read-only explorer built-ins cannot inherit
  Bash or mutating tools from a more permissive parent.
• The footer uses one compact Task Plane count; /tasks remains the place for
  status, duration, cancellation, and detailed output.

Compatibility and privacy

• Existing Goal/session/Role records, LoopResult states, exit codes, completion
  outbox, and lifecycle APIs remain compatible.
• Legacy strategy records are normalized in memory without rewriting history.
• New convergence, permission, capability-ceiling, and Resume-suggestion metrics
  contain only bounded categories and counts. They exclude objectives, blocker
  keys, paths, commands, custom tool names, session IDs, prompts, and transcripts.

Validation completed so far

• Agent-core full suite: 3,307 tests passed, one todo, one skipped at the recorded
  development checkpoint.
• CLI full suite: 2,367 tests passed, three skipped at the recorded development
  checkpoint.
• Real Node, Python, Rust, and Go fixtures each cover a failing project check,
  an actual file fix, and a subsequent PASS.
• Codex, Claude Code, and Cursor each cover suggestion, source picker, editable
  Loop objective, Goal creation, and terminal PASS in the TUI integration suite.
• The bundled CLI, docs site, and local macOS arm64 SEA binary build and pass
  their smoke checks.
• The public release contains six native ZIP files and six matching SHA-256
  files. A fresh 0.0.16 install and a real native 0.0.15 → 0.0.16
  wukong upgrade both verify the checksum and run as 0.0.16.
• The public update manifest, version-targeted announcement, download page,
  changelog, Today article, and production health endpoint are live and report
  0.0.16.

The release remains free and BYOK. Payment, managed model credits, PR Gate
publishing, Team features, hosted execution, and multi-reviewer quorum remain
outside 0.0.16.