epic: lightweight workflow, skill consolidation, prompt and hook SSOT

Objective

Rebuild OmC's workflow surface around a lightweight, legible, single-source-of-truth operating model comparable in spirit to Gajae-Code: fewer workflow gates, fewer overlapping skills/prompts/hooks, simpler execution paths, and explicit compatibility/deprecation boundaries.

This epic starts with architecture and migration planning. Implementation child issues are created only after the current surface is mapped and the target contract is reviewed.

Current baseline (2026-08-12)

• 41 shipped skills
• 294 files under src/hooks/
• 8 GitHub workflow files
• 19 agent prompt definitions
• 29 command markdown entrypoints
• multiple prompt/document copies (CLAUDE.md, docs/CLAUDE.md, .github/CLAUDE.md, agent/command/skill text, generated coordinator artifacts)
• overlapping workflow/gate surfaces including persistent-mode gating, merge-readiness, prompt prerequisites, task-size routing, skill state, generated-artifact authorization, release boundaries, and duplicated build/shipping closure

Counts are inventory evidence, not automatic deletion targets. Each retained surface must earn its contract.

Owner direction

• remove excessively hard workflow gates;
• consolidate and deprecate overlapping skills;
• simplify prompts;
• simplify hooks and unify them behind an explicit SSOT;
• create the epic and complete planning before broad implementation.

Design principles

1. One obvious path. Routine implementation, review, verification, and continuation should not require stacking multiple skills/gates that restate the same obligation.
2. Advisory by default, hard gate only for irreversible risk. Security, destructive mutation, release/publish authority, credential/privacy boundaries, and proven corruption risks may remain fail-closed. Ceremony-only or duplicative gates must be removed.
3. Single source, generated projections. Prompts, hook registrations, skill aliases, command metadata, and shipped documentation should have one canonical source with deterministic generated views.
4. Compatibility is bounded. Deprecated names receive explicit aliases, warnings, migration docs, and removal milestones rather than permanent duplicated implementations.
5. Measure deletion. The plan must state baseline and target reductions for skills, hooks, prompts, commands, gates, generated artifacts, and maintenance/test surface.
6. No big-bang rewrite. Land reversible phases with parity tests, migration receipts, and clean rollback points.

Planning deliverable

Create one reviewed planning PR targeting dev containing:

• complete registry/call-graph map of skills, commands, agents, prompts, hooks, workflows, gates, generated artifacts, aliases, and configuration entrypoints;
• comparison with the current Gajae-Code lightweight workflow, focusing on reusable design principles rather than blind code copying;
• keep / merge / deprecate / delete classification for every public skill and command;
• hook pipeline map identifying duplicate triggers, state writes, lifecycle overlap, hot-path cost, and proposed unified dispatcher/SSOT;
• prompt SSOT proposal covering canonical source, composition, model/provider deltas, generated projections, digest/versioning, and package/install behavior;
• workflow-gate taxonomy: retain hard, convert to advisory, merge, or delete, with concrete risk justification;
• target architecture and migration sequence;
• compatibility/deprecation policy and telemetry/observability needed to remove aliases safely;
• quantitative before/after targets and acceptance metrics;
• ordered child-issue/PR decomposition, one coherent item per issue, including dependencies and rollback boundaries;
• release/migration strategy for existing users and plugin installations.

Required planning decisions

Skills and commands

• identify synonyms, aliases, wrappers, and multi-step skills that can collapse into fewer primary workflows;
• define the minimal Tier-0 public surface;
• define deprecation warnings, migration mappings, and removal version/timebox.

Prompts and agents

• identify duplicated policy/instructions across CLAUDE.md, docs, agents, commands, skills, stage prompts, and hook injection;
• propose canonical structured prompt sources and deterministic projections;
• shorten prompts by removing repeated procedure while preserving safety and product semantics.

Hooks and state

• inventory every hook registration, event, state root, lock, timeout, and cleanup owner;
• merge redundant pre/post/session/persistent-mode logic into a minimal dispatcher and explicit modules;
• minimize synchronous hot-path work and fail-open noise while preserving security boundaries;
• make registration and cleanup mechanically auditable from one registry.

CI and workflow gates

• remove candidate/base double-gates and generated-closure ceremony where a deterministic build/provenance check can provide the same guarantee more simply;
• retain branch protection and release authority without requiring bespoke authorization PR chains for routine integration;
• simplify exact-head review/CI evidence and release promotion into an understandable state machine.

Non-goals for the planning PR

• no broad runtime behavior rewrite;
• no release, tag, npm publish, protected-main mutation, or compatibility removal;
• no deleting public skills/hooks before the migration plan and parity checks are reviewed;
• no vague "simplify later" roadmap without per-surface classification and measurable targets.

Epic terminal condition

This epic closes only when all approved child issues are terminal, deprecated surfaces have documented migration/removal status, the lightweight workflow metrics are verified on shipped artifacts, and the old duplicated/gated paths are removed or explicitly retained with risk evidence.

— [repo owner's gaebal-gajae (clawdbot) ]