Skip to content

ssojux2/high-agency

v0.14.0MIT

Lightweight high-agency coding with impact calibration, adaptive model routing, targeted verification, and bounded autonomous iteration for Codex.

GitHub

high-agency

Version: 0.14.0

Version 0.14.0 bounds repeated verification: a passing check is not rerun on an unchanged snapshot, a failing check gets one focused fix and one retest, and confidence comes from distinct test data in a single run rather than repetitions. The verification hook denies one identical no-new-evidence rerun at PreToolUse; repeating the same command afterwards runs it.

Version 0.13.0 adds explicit assumptions, proportional implementation, focused edits, and observable acceptance criteria to the coding and bounded-autonomy skills. See coding examples and attribution. Existing hooks, model selection, and platform launchers are unchanged.

The 0.12.1 hook fix remains included: whitespace-heavy assistant responses stay within the Impact parser timeout, CRLF and horizontal whitespace are supported, and Codex keeps its native Stop decision format.

Prerequisites

Python 3.10+ must be available on the PATH inherited by Codex.

PlatformRequired interpreter / hook launcher
Linux and macOS (POSIX)python3
Native Windowspython, selected by the hook's commandWindows entry

An activated virtual environment is acceptable if Codex inherits its PATH. Native Windows requires a host that honors commandWindows; a POSIX or WSL run does not validate the native Windows launcher. The plugin does not install Python or edit global PATH/configuration. CI runs native Windows, Linux, and macOS regression suites and real Codex local-installation and model-catalog checks. Authenticated task execution remains a separate acceptance check.

Contents

Lightweight Codex plugin containing:

  • high-agency-coding: direct autonomous coding with explicit assumptions, focused edits, and fresh verification.
  • bounded-autonomy: opt-in bounded continuation via the bundled Stop hook.
  • high-agency-doctor: read-only configuration and routing diagnostics.
  • scripts/model_catalog.py: catalog metadata lookup and family selection.
  • hooks/routing_observer.py: read-only reporting of routing evidence.

Routing and evidence

The current main model stays the integrator. Native subtask delegation is optional skill guidance; plugin code does not force dispatch or launch agents. Resolve exact model IDs and supported efforts from the active host's current catalog rather than a fixed release list.

WorkFamily preference order
Mechanical work / test reportingLuna → Terra → Sol → current main model
BROAD MAPSol → Terra → current main model
WORKHORSE implementationSol → current main model
Hard reasoningAstra → Sol → current main model

Each family means its latest available supported release. Catalog query success, freshness, account entitlement/availability, native dispatch, and served-model evidence are separate checks. A successful lookup does not establish all of them.

From this plugin directory, run python3 hooks/routing_observer.py --report --session-id "<current-session-id>" on Linux/macOS, or python hooks/routing_observer.py --report --session-id "<current-session-id>" on native Windows. Supply the current session ID from the host. --report alone returns capabilities and UNVERIFIED observations; it never chooses a current/latest session automatically. Add --agent-id "<current-agent-id>" for child scope. If PLUGIN_DATA is absent, pass --state-dir "<routing-directory>" only with the host-known routing directory; otherwise state access remains UNVERIFIED.

The scoped report separates observer capability and requested-model, resolved-model, and modelsUsed statuses. Codex PostToolUse events for spawn_agent|Agent support requested-model and dispatch observations. The canonical child resolvedModel/modelsUsed schema is unavailable, so only the served-model metadata capability is unsupported, and served identity remains UNVERIFIED. Do not infer served identity or absence of delegation from an empty report.

State is kept per session under the plugin state directory, without copying prompts, tool output, or credentials. Diagnostics do not run paid probes automatically, install SDKs, or start nested CLI inference sessions.

The 0.12.0 audit changes are checked with regression fixtures and real Codex local plugin installation and model-catalog queries on Windows, Linux, and macOS CI. Authenticated end-to-end evaluation is still needed; token, latency, and cost savings are not established by this patch.

Codex 0.160.1 native exec_command hook responses omit the exit status and effective working directory. Raw output therefore remains UNVERIFIED, including printed success headers. The main agent must inspect the actual tool result; the hook issues bounded reminders and does not certify evidence the host did not provide.

See the repository README, routing scenarios, and runtime validation for installation, usage, and evaluation evidence.