v0.5.0 — ARMED handoff and codex collaboration

• ARMED-handoff transport: the codex wrapper launches codex in the background and ends its turn; the orchestrator owns the waiting through a background watcher (DONE / DIED / TIMEOUT) and messages the wrapper to relay. The wrapper busy-wait loop is gone.
• Pid-file liveness: the launch command records the codex pid and the watcher probes it with kill -0 — command-line pattern matching proved unreliable under host hooks that inject text into every shell's argv (falsified live before shipping).
• Codex collaboration dispatch: codex exec sessions expose collaboration.spawn_agent; codex reviewers dispatch auxiliary personas for real and record child ids. Execution-tier coverage still never satisfies final-fresh-exit — codex-side ids reach the orchestrator only as self-reported relay text, so the design-tier gate keeps its verifiable-dispatch rationale.
• Researcher personas rewritten as contracts (lens + negative space + output shape) in place of inherited phase choreography; the repository's own git history is repo evidence at every fresh-review boundary; adversarial review depth keys on consequence, not changed-line counts.
• cvg-work failing-first evidence now checks against the slice's parent commit — mechanically verifiable and compatible with one atomic commit per slice.
• Hybrid-transport smoke procedure documented (five legs, including the DIED path) after being run live end-to-end; wrapper-compliance eval rewritten two-phase, re-baselined 5/5.

v0.4.0 — Contractual boundaries and the loop ledger

• Permission and sandbox management removed from the hybrid loops: convergo is a developer tool; codex sessions run with whatever the user's own codex configuration grants. The loop's protection is contractual — review gates, git reversibility, phase-gated external side effects.
• Gate 5 loop ledger: round counting and adjudications move from in-context memory to a per-loop file; the cap check and the ratchet read the ledger.
• Two-arm adjudication ratchet: the evidence arm is never exempted; the classification arm exempts final-fresh-exit reviewers (an exit gate blocks on its own judgment, citing the prior classification).
• Codex resume re-pins the model: codex exec resume drops the first run's -m pin and re-resolves from config — every resume now carries -c model= and -c model_reasoning_effort=, with the eval:repin ground-truth canary asserting the compensation holds across codex upgrades.
• Execution tier pinned to gpt-5.6-sol at high reasoning effort, with a config-default fallback when a pinned id is retired; hybrid loops echo resolved bindings before the first spawn.
• Frontier alignment: correctness AND testing reviewers always-on, additional lenses chosen by the diff's risk surface with the rationale recorded; aux reviewers get a side-effect boundary (targeted tests allowed) instead of a literal write ban; cvg-work TDD enforcement moves from step choreography to reviewer-checkable evidence.
• Five dangling persona lens references fixed; a new conventions test scans persona prose for deferrals to nonexistent lenses.

v0.3.9 — Hybrid split-engine loops

Highlights (0.3.0 → 0.3.9)

Two new Claude Code-only skills route the existing loops across a
high-taste Claude design model and cheap Codex CLI sessions:

• /cvg-hybrid-plan-loop — the design-tier model (default: Fable) drafts and
  revises the plan; Codex CLI sessions gate it as fresh plan reviewers.
• /cvg-hybrid-build-loop — Codex implements and takes every first-pass
  review round; a clean cheap-tier review promotes to a design-tier
  final-fresh-exit review instead of ending the loop. Only a clean
  design-tier pass exits.

Every mechanic in the codex-backed specialist wrapper pattern was verified
against live runs on codex-cli 0.142.3 and hardened release by release:

• wrappers stay in-turn (a stopped background specialist is never re-invoked
  when its background command exits) with a SendMessage recovery path
• sandbox re-pinned on every resume (an unpinned resume re-resolves from
  local config and can silently escalate to full access)
• reviewers run workspace-write from a scratch working directory: repo
  OS-level read-only, /tmp audit artifacts still land (read-only blocks
  them entirely)
• workers in linked git worktrees get the git common dir via --add-dir, so
  per-slice commits work
• the -o file is the source of truth, not the codex exit code
• cross-repo plans: one loop per repo, uncovered repos are blocking known
  gaps (also applied to the base cvg-build-loop)

Three behavioral eval layers under evals/wrapper-behavior/ back these
rules: a sandbox ground-truth matrix against the real codex CLI (also the
canary for codex upgrades), wrapper-compliance scenarios against a stub codex
with per-rule ablation, and orchestrator cross-repo scoping scenarios.

Full details: README §Hybrid Loops and evals/wrapper-behavior/README.md.