Two independently installable plugins provide four normal entry points. The main agent owns the requested outcome, evidence, model selection, and the retry allowance shared across agents and gates for each unresolved problem within one run from user prompt to resolution. Delivery uses complexity-based routing with three initial attempts per task/problem followed by one per higher model tier through Astra; standalone shaping, library, bootstrap and installation retain a three-failure limit. A separate later request, including PR comment review, starts its own budget; reading comments or discovering defects does not consume attempts.
| Plugin | Normal work | One-time setup | Supporting agents |
|---|---|---|---|
| Engineering | shape produces proposals/specs; explicitly invoked deliver implements or repairs through a verified PR. |
bootstrap creates or completes board and forge declarations. |
coder, QA, writer, auditor |
| Library | research investigates and saves cited evidence; ask answers from existing library knowledge without internet requests or library writes. |
bootstrap creates or safely completes the Markdown library; Obsidian is optional. |
researcher, librarian, scribe, clerk |
Both plugins also provide install-subagents. Nontrivial changes require a written spec and fresh validation before production, then a different fresh validator for the candidate. Trivial changes can skip the written spec but still require fresh validation. Production delegation is optional. Missing production roles permit direct scoped work; missing required validators block the affected gate.
Engineering uses a fresh auditor for spec readiness and document acceptance, or fresh QA for code behavior and tests. One adequate final evaluation can include affected documentation and mechanical checks. Library uses fresh clerks for research specs, answers, evidence, and integrity. Production leaves author artifacts and tests; validators report findings and execute checks without repairing the candidate.
Researchers return new-source findings and provenance. Librarians retrieve existing knowledge read-only. The main agent or one designated scribe integrates synthesis and shared metadata after raw-note writers finish. Research follows project conventions and the configured provider; this repository requires webtools for internet research and reports its absence without provider substitution.
Engineering execution specs carry overall and per-task complexity scores (1–10) with rationales. The delivery matrix uses GPT-6 Luna/max for scores 1–4, Sol/medium for 5–6, Sol/high for 7–8 and Astra/medium for 9–10, with Sol medium as the calibration anchor and justified stronger selections recorded. Spec production and implementation tasks keep separate retry histories: a stronger spec author does not impose its model on child tasks. Only supported listed models or host-confirmed aliases receive routing tiers; unknown identities are reported instead of guessed.
Every entry point, including bootstrap and installation, keeps a durable ledger owned by the main agent. It records progress, returned subagent IDs/canonical handles, assignments, gate evidence, and failure history before waits and handoffs. Engineering delivery also records overall/task complexity, intended and actual model/effort, selection rationale and per-tier escalation state. Reuse it across skills and resumptions of that run; a separate later user request gets a new ledger and count, with prior ledgers retained as context. Project ledgers use <temp-folder>/ledgers/<run-id>.md, with the temp folder read from the forge declaration (normally docs/FORGE.md) and defaulting to .tmp/ under the project root. Without a project, use $CODEX_HOME/ledgers/ (default ~/.codex/ledgers/). Before creating new history, search all registered Git worktrees using each checkout’s configuration and legacy ledger directory. Before ledger writes, verify Git ignore protection and add a local exclude rule for the ledger directory when needed. Keep ledgers outside the research library and installed plugin caches, and preserve them through scratch cleanup and completion. Each plugin ships a template: engineering and library.
Project authority lives in docs/BOARD.md, docs/FORGE.md, and library/_meta/librarian.md. Explicit deliver invocation authorizes the task branch/worktree, attributable commits, verified push, and PR creation or update. Implementation requested without deliver may remain uncommitted on master; a proposal does not imply filing a card. Normal work consumes setup without running bootstrap.
Add this repository's marketplace, then install either or both plugins:
codex plugin marketplace add /Users/catty/Workspace/agentic-codex
codex plugin add ca77y-engineering@personal
codex plugin add ca77y-library@personalInstall the selected plugins' custom agents with Python 3.11 or newer:
python3 plugins/ca77y-engineering/skills/install-subagents/scripts/install_agents.py
python3 plugins/ca77y-library/skills/install-subagents/scripts/install_agents.pyRun the same installer after source updates. It embeds each leaf's agents/<role>/AGENT.md and copies conditional references under ~/.codex/agents/.ca77y-engineering/ or .ca77y-library/, so installed references survive checkout/cache removal. Only marked files are updated or removed; unmanaged conflicts are refused and other-plugin files are preserved. Agent definitions contain no fixed model or reasoning settings.
Start a new Codex task after installation to reload skills and the custom-agent catalog. deliver requires explicit invocation, such as $ca77y-engineering:deliver, and is never selected automatically. Other skills can be invoked by their qualified name, such as $ca77y-library:research, or by describing matching work naturally. Engineering dispatches ca77y_engineering_coder, ca77y_engineering_qa, ca77y_engineering_writer, and ca77y_engineering_auditor. Library dispatches ca77y_library_researcher, ca77y_library_librarian, ca77y_library_scribe, and ca77y_library_clerk.
Diagnostics are optional installation utilities: --check inspects source without installing; --check-installed reports read-only drift; --target selects a temporary destination. Checks requested through a skill are delegated to a fresh engineering auditor or library clerk.
Use Python 3.11 or newer with PyYAML available. Run every skill quick validator, then both plugin validators and installer suites:
(
for skill in plugins/*/skills/*/; do
python3 /Users/catty/.codex/skills/.system/skill-creator/scripts/quick_validate.py "$skill" || exit $?
done
python3 /Users/catty/.codex/skills/.system/plugin-creator/scripts/validate_plugin.py plugins/ca77y-engineering || exit $?
python3 /Users/catty/.codex/skills/.system/plugin-creator/scripts/validate_plugin.py plugins/ca77y-library || exit $?
python3 plugins/ca77y-engineering/skills/install-subagents/scripts/test_install_agents.py || exit $?
python3 plugins/ca77y-library/skills/install-subagents/scripts/test_install_agents.py || exit $?
)The installer suites exercise real plugin resources and isolated destinations, including stale cleanup, unmanaged conflicts, drift, and reference survival after source removal. Mechanical checks establish format and packaging validity; bounded scenarios assess instruction behavior separately. The acceptance sources are supporting agents and skill integration , orchestrator ledgers, and delivery complexity routing.