triad-plus 1.0.0

package/CHANGELOG.md

@@ -0,0 +1,9 @@
+# Changelog
+
+## 1.0.0 — 2026-08-24
+
+- Public Triad+ engineering loop with Orchestrator, Developer, and Reviewer.
+- Optional fresh post-run Evaluator+.
+- Codex, Claude Code, OpenCode, Antigravity, and Hermes Agent adapters.
+- Hash-bound control-plane verification evidence and package installer.
+- Public setup, runtime, architecture, verification, and contribution guides.

package/CODE_OF_CONDUCT.md

@@ -0,0 +1,9 @@
+# Code of Conduct
+
+Be respectful, constructive, and precise. Discuss code and ideas rather than
+people. Harassment, discrimination, threats, and disclosure of another person's
+private information are not acceptable.
+
+If a discussion needs moderation, use the repository's private reporting path
+described in [SECURITY.md](SECURITY.md) or request a private contact channel from
+the project owner.

package/CONTRIBUTING.md

@@ -0,0 +1,44 @@
+# Contributing to Triad+
+
+Thank you for contributing. Triad+ is intentionally small: preserve the Core
+roles and keep the host agent in charge of the loop.
+
+## Local setup
+
+Use Node.js 20 or later, then run:
+
+```bash
+npm test
+npm run pack:check
+```
+
+The repository is organized as follows:
+
+- `skills/` — role and bootstrap contracts.
+- `runtime/` — verification and evidence utilities.
+- `adapters/` — host-specific assets and metadata.
+- `docs/` — public usage and architecture documentation.
+
+## Pull requests
+
+Keep each pull request focused. Include tests for changed behavior, update public
+documentation when user-visible behavior changes, and avoid private paths,
+tokens, deployment configuration, or generated smoke artifacts.
+
+Triad+ is feature-frozen around its public Core: Orchestrator, Developer, and
+Reviewer, with optional post-run Evaluator+. Do not add workflow engines,
+schedulers, automatic repair, additional Core roles, or runtime-specific Core
+branches without prior design agreement.
+
+## Adapters
+
+To add or modify an adapter, update its descriptor in `adapters/registry.mjs`,
+its adapter assets, and tests. The descriptor should express only genuine host
+differences: detection, asset locations, entry point, capabilities, and model
+binding. Do not add `if <runtime>` behavior to Core installer or verifier code.
+
+## Review expectations
+
+Explain the user-facing outcome, provide test output, and call out runtime
+limitations. Maintainers may ask for a clean package smoke when install assets
+change.

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Giovanni Delbono
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/README.md

@@ -0,0 +1,80 @@
+# Triad+
+
+Triad+ is a lightweight engineering loop for coding agents with three explicit roles.
+
+> **Developer proposes. Reviewer challenges. Orchestrator governs.**
+
+It is not a workflow engine or autonomous development platform. The coding-agent host stays in control; Triad+ provides role contracts, independent review, and verification evidence.
+
+```text
+                Orchestrator
+                /           \
+          Developer  ←→  Reviewer
+
+Triad approved → optional fresh Evaluator+
+```
+
+## Why Triad?
+
+A single agent can implement, self-check, and declare success in one conversation. Triad separates those responsibilities: the Developer changes the candidate, the Reviewer looks for defects, and the Orchestrator decides rework, approval, or escalation. `triad-verify` supplies environment-derived gate evidence, distinct from an agent claim.
+
+## Supported coding agents
+
+| Runtime | Orchestrator | Developer | Reviewer | Evaluator+ | Verification dispatch |
+| --- | --- | --- | --- | --- | --- |
+| Codex | Yes | Yes | Yes | Yes | Explicit fallback; async hook when validated |
+| Claude Code | Yes | Yes | Yes | Yes | Explicit fallback; hook when validated |
+| OpenCode | Yes | Yes | Yes | Yes | Explicit dispatch |
+| Antigravity | Yes | Yes | Yes | Yes | Explicit dispatch |
+| Hermes Agent | Yes | Yes | Yes | Yes | Explicit dispatch |
+
+Hooks are optional optimizations; they write evidence and never make a Triad state decision. See [runtime details](docs/runtimes.md).
+
+## Quick start
+
+Requirements: Node.js 20+ and one supported coding-agent host.
+
+```bash
+npx triad-plus init --host codex --control /absolute/path/to/triad-control --global
+npx triad-plus doctor --host codex --control /absolute/path/to/triad-control
+```
+
+`init` creates a separate control workspace and refuses overwrites. Open it in the selected host, then start Triad with an absolute PRD path:
+
+| Host | Entry point |
+| --- | --- |
+| Codex | `/prompts:triad /absolute/path/to/prd.md` |
+| Claude Code | `/triad /absolute/path/to/prd.md` |
+| OpenCode | `/triad /absolute/path/to/prd.md` |
+| Antigravity | `/triad /absolute/path/to/prd.md` |
+| Hermes Agent | `/triad /absolute/path/to/prd.md` |
+
+The Orchestrator presents feature cards before implementation. Follow [Getting started](docs/getting-started.md) for the complete copy-paste walkthrough.
+
+## Configure agents
+
+The wizard or reviewed `team.json` configures display name/persona, model, and supported effort/options for each role. Use neutral names. The selected adapter/runtime is project-level; per-role model binding is applied only where the host supports it. [Configuration](docs/configuration.md) explains the exact contract.
+
+## Optional Evaluator+
+
+Evaluator+ is not a fourth member of Triad. After Reviewer approval, invoke the native command with `--evaluator` where supported:
+
+```text
+/triad --evaluator
+```
+
+It evaluates the completed result freshly and independently. A `FAIL` does not reopen Triad or start repair. See [Evaluator+](docs/evaluator-plus.md).
+
+## Architecture and limits
+
+Triad+ consists of role skills, adapters, and a verifier. Core has no runtime-specific behavior; adapters own installation, entry-point, lifecycle, and model-binding differences. The verifier validates binding, runs declared control-plane gates, detects mutation, and writes atomic evidence. It does not govern the conversation or replace the Orchestrator.
+
+Triad+ does not provide a daemon, scheduler, dashboard, shared memory, automatic model routing, multi-reviewer voting, or evaluator-driven repair. Read [architecture](docs/architecture.md), [verification](docs/verification.md), and [troubleshooting](docs/troubleshooting.md).
+
+## Contributing
+
+Start with [CONTRIBUTING.md](CONTRIBUTING.md), run `npm test`, and keep runtime-specific behavior inside an adapter.
+
+## License
+
+[MIT](LICENSE).

package/SECURITY.md

@@ -0,0 +1,12 @@
+# Security policy
+
+Please do not publish suspected vulnerabilities, credential exposure, or unsafe
+verification behavior in a public issue before maintainers have assessed it.
+
+Use GitHub's private security-advisory reporting feature for this repository if
+it is enabled. If it is not available, open a minimal public issue requesting a
+private contact channel and do not include exploit details, tokens, or sensitive
+paths. The project owner will establish the appropriate channel.
+
+When reporting, include affected version, reproduction steps, impact, and any
+safe mitigation you have identified.

package/adapters/antigravity/.agents/agents/triad-developer/agent.md

@@ -0,0 +1,11 @@
+---
+name: triad-developer
+description: Implement one bounded Triad+ feature card with tests and precise evidence.
+subagent: true
+---
+
+Load `triad-loop-developer` at the start of every assignment. Implement only
+the declared feature card in its worktree, run required local checks, and report
+exact commands, results, changed files, metrics, risks, and blockers. Do not
+approve, review, commit, push, publish, or change loop state unless explicitly
+assigned by the Orchestrator.

package/adapters/antigravity/.agents/agents/triad-evaluator/agent.md

@@ -0,0 +1,11 @@
+---
+name: triad-evaluator
+description: Freshly evaluate a verified Triad+ Gauntlet candidate against its quality bar.
+subagent: true
+---
+
+Load `triad-loop-evaluator`. Use a fresh context and only the evaluation packet,
+quality bar, candidate observation instructions, and valid verification summary.
+Return `candidate_wins`, `bar_wins`, or `indeterminate`; a `bar_wins` decision
+has exactly one evidence-based largest gap. Do not edit source, change loop
+state, commit, push, or review delivery.

package/adapters/antigravity/.agents/agents/triad-orchestrator/agent.md

@@ -0,0 +1,19 @@
+---
+name: triad-orchestrator
+description: Govern a Triad+ Engineering Loop, delegate normal work, and make evidence-based workflow decisions.
+---
+
+Act as the Triad+ Orchestrator. Load `triad-loop-bootstrap` for a new project
+and `triad-loop-orchestrator` for an initialized one. Read
+`.triad-plus/team.json` if present and verify exposed role models against its
+contract before starting. If the host cannot expose a model identity, say so and
+ask the owner to confirm it; never invent a match. Keep the project-control workspace separate
+from product repositories; delegate normal development, fresh Gauntlet
+evaluation, and independent review to the corresponding Triad agents.
+
+Detect the runtime with `.triad-runtime/triad-runtime-capabilities.mjs --host
+antigravity`. Use explicit verification dispatch unless the recorded capability
+snapshot proves a supported asynchronous route. Govern ordinary disagreements
+from evidence and record the decision. Do not implement or review routinely,
+and do not start demos, create pull requests, publish packages, releases, or
+force-push without owner authorization.

package/adapters/antigravity/.agents/agents/triad-reviewer/agent.md

@@ -0,0 +1,10 @@
+---
+name: triad-reviewer
+description: Independently review one verified Triad+ feature attempt and return an evidence-based recommendation.
+subagent: true
+---
+
+Load `triad-loop-reviewer`. Review the declared card, diff, valid verification
+evidence, metrics, risks, and applicable evaluation result independently from
+the Developer. Return `approved`, `rework`, or `blocked` with concrete evidence.
+Do not routinely implement fixes, commit, push, or alter project policy.

package/adapters/antigravity/.agents/skills/triad/SKILL.md

@@ -0,0 +1,30 @@
+---
+name: triad
+description: Start or resume a Triad+ Engineering Loop in Antigravity from a PRD source, target repositories, and measurable delivery goals. Use when the owner invokes /triad or asks Antigravity to orchestrate a Triad+ project.
+---
+
+# Triad+ workflow
+
+Operate the current conversation as the Triad+ Orchestrator for the owner's
+request. Read `.triad-plus/team.json` when present and use its language, owner
+address, display names, personas, and model contract; technical role IDs remain
+unchanged.
+
+Before work, compare every role model that the host exposes with the recorded
+contract. If Antigravity does not expose a model identity, say that it cannot be
+verified and ask the owner to select or confirm it; never claim a model match
+without evidence. An unavailable required role stops the loop rather than being
+silently substituted.
+
+For a new project, load `triad-loop-bootstrap`; for an initialized project,
+load `triad-loop-orchestrator`. At bootstrap and resume, run
+`.triad-runtime/triad-runtime-capabilities.mjs --adapter .triad-runtime/adapter.json` and follow
+its selected verification route. Use explicit verifier dispatch unless a future
+Antigravity lifecycle adapter is actually detected and validated.
+
+Delegate normal implementation to `triad-developer` and independent review to
+`triad-reviewer`. Invoke `triad-evaluator` only after Triad approval when the
+owner requests Evaluator+; its independent report never reopens Triad. Continue
+through declared cards and normal pushes after all gates pass. Escalate only the
+decision types defined by the Triad+ skills. Do not start or stop a demo without
+an owner instruction.

package/adapters/antigravity/.agents/skills/triad/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Triad+ Workflow"
+  short_description: "Start or resume a Triad+ engineering loop."
+  default_prompt: "Start Triad+ for this project."

package/adapters/antigravity/README.md

@@ -0,0 +1,26 @@
+# Triad+ for Antigravity
+
+The Antigravity adapter uses native workspace paths: `.agents/skills/` for the
+five shared Triad+ skills and `/triad` workflow, plus `.agents/agents/` for the
+four technical agents. Start Antigravity from the project-control workspace and
+invoke:
+
+```text
+/triad <PRD path or project request>
+```
+
+Install with the Triad+ CLI:
+
+```bash
+npx triad-plus init --host antigravity --control /path/to/project-control
+```
+
+Add `--global` to install the same assets under `~/.gemini/config/`. The
+installer stores the selected role/model contract in `.triad-plus/team.json`.
+Antigravity model selection remains a host setting; select a model through its
+native controls. `/triad` compares the contract with any host-exposed model
+identity and asks for owner confirmation when the host cannot expose one.
+
+The initial adapter uses explicit verification dispatch. Background subagents
+are useful for role separation, but they are not treated as proof of a verified
+lifecycle hook until a versioned Antigravity control-plane adapter exists.

package/adapters/antigravity/install.sh

@@ -0,0 +1,72 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+usage() {
+  cat <<'EOF'
+Usage:
+  ./adapters/antigravity/install.sh --project <control-repository-path>
+  ./adapters/antigravity/install.sh --global
+
+Install the Triad+ Antigravity adapter without overwriting existing agents,
+skills, workflow, or project runtime. --project installs into
+<control-repository-path>/.agents and also installs .triad-runtime. --global
+installs agents, skills, and /triad workflow under $HOME/.gemini/config.
+EOF
+}
+
+if [[ $# -eq 2 && "$1" == "--project" ]]; then
+  target_root="$2/.agents"
+  project_root="$2"
+  project_install=true
+elif [[ $# -eq 1 && "$1" == "--global" ]]; then
+  target_root="$HOME/.gemini/config"
+  project_root=""
+  project_install=false
+else
+  usage >&2
+  exit 2
+fi
+
+adapter_root="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
+repository_root="$(cd -- "$adapter_root/../.." && pwd)"
+agent_names=(triad-orchestrator triad-developer triad-evaluator triad-reviewer)
+skill_names=(triad-loop-bootstrap triad-loop-orchestrator triad-loop-developer triad-loop-evaluator triad-loop-reviewer)
+collisions=()
+
+for agent_name in "${agent_names[@]}"; do
+  [[ -e "$target_root/agents/$agent_name/agent.md" ]] && collisions+=("agents/$agent_name/agent.md")
+done
+for skill_name in "${skill_names[@]}"; do
+  [[ -e "$target_root/skills/$skill_name" ]] && collisions+=("skills/$skill_name")
+done
+[[ -e "$target_root/skills/triad" ]] && collisions+=("skills/triad")
+if [[ "$project_install" == true && -e "$project_root/.triad-runtime" ]]; then
+  collisions+=(".triad-runtime")
+fi
+
+if [[ ${#collisions[@]} -gt 0 ]]; then
+  printf 'Installation aborted; existing paths would be overwritten:\n' >&2
+  printf '  %s\n' "${collisions[@]}" >&2
+  printf 'Remove or rename those paths deliberately, then run the installer again.\n' >&2
+  exit 1
+fi
+
+mkdir -p "$target_root/agents" "$target_root/skills"
+for agent_name in "${agent_names[@]}"; do
+  cp -R "$adapter_root/.agents/agents/$agent_name" "$target_root/agents/$agent_name"
+done
+for skill_name in "${skill_names[@]}"; do
+  cp -R "$repository_root/skills/$skill_name" "$target_root/skills/$skill_name"
+done
+cp -R "$adapter_root/.agents/skills/triad" "$target_root/skills/triad"
+if [[ "$project_install" == true ]]; then
+  cp -R "$repository_root/runtime" "$project_root/.triad-runtime"
+  cp -R "$repository_root/schemas" "$project_root/.triad-runtime/schemas"
+  cp "$adapter_root/runtime.json" "$project_root/.triad-runtime/adapter.json"
+fi
+
+printf 'Triad+ Engineering Loop installed in %s\n' "$target_root"
+if [[ "$project_install" == true ]]; then
+  printf 'Control-plane runtime installed in %s/.triad-runtime\n' "$project_root"
+fi
+printf 'Start Antigravity in the target project and use /triad followed by a PRD source or project request.\n'

package/adapters/antigravity/runtime.json

@@ -0,0 +1,6 @@
+{
+  "schema_version": 1,
+  "id": "antigravity",
+  "binary": "agy",
+  "lifecycle": null
+}

package/adapters/claude-code/.claude/agents/triad-developer.md

@@ -0,0 +1,19 @@
+---
+name: triad-developer
+description: Implement exactly one bounded Triad feature card with focused tests and precise evidence. Use only when the orchestrator assigns a ready or quality-rework card.
+tools: Read, Edit, Write, Bash, Glob, Grep, Skill
+permissionMode: default
+skills:
+  - triad-loop-developer
+---
+
+Implement only the assigned feature card in its declared worktree. Read the
+card, PRD excerpt, repository instructions, allowed surface, prior attempt, and
+required gates first. Run useful local checks but do not claim a `control-plane`
+gate passed: finish the attempt and let the selected verification route create
+authoritative evidence.
+
+On a quality repair, use only the current largest gap, its direct evidence, and
+the bounded repair scope. Do not seek prior evaluator narrative. Do not change
+queue, run state, assignments, evidence, evaluations, scope, or policy; do not
+approve, review, commit, push, publish, or release unless explicitly assigned.

package/adapters/claude-code/.claude/agents/triad-evaluator.md

@@ -0,0 +1,17 @@
+---
+name: triad-evaluator
+description: Freshly compare one verified Gauntlet candidate against a snapshotted quality bar and return one bounded largest remaining gap.
+tools: Read, Bash, Glob, Grep, Skill
+permissionMode: plan
+skills:
+  - triad-loop-evaluator
+---
+
+Use a fresh context for every evaluation. Read only the evaluation packet, real
+candidate artifact, quality-bar snapshot, observation instructions, and valid
+verification summary. Do not request or use developer reports, reasoning,
+history, or prior evaluator/reviewer findings.
+
+Return `candidate_wins`, `bar_wins`, or `indeterminate`. A `bar_wins` result has
+exactly one evidence-based largest gap with one bounded repair scope. Do not edit
+product source, change state, commit, push, or approve delivery.

package/adapters/claude-code/.claude/agents/triad-reviewer.md

@@ -0,0 +1,14 @@
+---
+name: triad-reviewer
+description: Independently review one complete Triad attempt for engineering delivery readiness after verification and optional Gauntlet evaluation.
+tools: Read, Bash, Glob, Grep, Skill
+permissionMode: plan
+skills:
+  - triad-loop-reviewer
+---
+
+Independently verify the card, PRD, actual diff, candidate fingerprint, external
+verification evidence, gates, metrics, scope, risks, and, when applicable, the
+quality-bar/evaluation trail. Return `approved`, `rework`, or `blocked` with
+severity-ranked findings. Do not edit source, transition state, commit, push,
+publish, or release.

package/adapters/claude-code/.claude/commands/triad.md

@@ -0,0 +1,26 @@
+---
+description: Start or resume a Triad+ Engineering Loop from a PRD source, target repositories, and measurable goals.
+---
+
+Operate a Triad+ Engineering Loop for this owner request:
+
+$ARGUMENTS
+
+Load `triad-loop-bootstrap` for a new project or `triad-loop-orchestrator` for
+an initialized project. At bootstrap and resume, run
+`.triad-runtime/triad-runtime-capabilities.mjs --adapter .triad-runtime/adapter.json` and record
+the result. Follow its selected verification mode: `hook_dispatch` only when the
+installed Claude Code SubagentStop hook has been configured and verified;
+otherwise explicitly invoke the verifier after Developer completion.
+
+If `.triad-plus/team.json` exists, load it before replying. Use its interaction
+language, owner address, display names, personas, and model contract in communication;
+technical role identifiers and authority remain unchanged. If the active
+Orchestrator model cannot meet the recorded contract, say so before work starts.
+
+Delegate implementation to `triad-developer` and review to `triad-reviewer`.
+Invoke fresh `triad-evaluator` only after Triad approval and only when requested;
+its report never reopens the completed run. Continue
+autonomously through declared cards and normal branch pushes once all gates pass.
+Escalate only the decision types defined by the Triad skills. Do not start or
+stop a demo without an owner instruction.

package/adapters/claude-code/README.md

@@ -0,0 +1,60 @@
+# Triad+ Engineering Loop for Claude Code
+
+This adapter installs the shared five Triad skills, three isolated Claude Code
+subagents, the `/triad` command, and—on a project installation—the Node control
+plane. It does not modify global Claude settings or automatically activate a
+hook.
+
+## Install into a project-control repository
+
+```bash
+./adapters/claude-code/install.sh --project /path/to/project-control-repository
+```
+
+It creates:
+
+```text
+.claude/agents/triad-developer.md
+.claude/agents/triad-evaluator.md
+.claude/agents/triad-reviewer.md
+.claude/commands/triad.md
+.claude/skills/triad-loop-*/
+.claude/triad-hooks.json
+.triad-runtime/
+```
+
+The installer refuses all collisions before copying. `.claude/triad-hooks.json`
+is an inert, reviewable fragment: merge it into the trusted Claude Code settings
+only after replacing its paths and confirming it through `/hooks`.
+
+## Start and verify
+
+Launch Claude Code in the control repository, then use:
+
+```text
+/triad <PRD location or concise project request>
+```
+
+The command delegates Developer, Evaluator, and Reviewer to fresh subagents.
+Confirm them with `/agents`; confirm skills with `/skills`. The project control
+plane detects Claude Code and Node at bootstrap. It selects `hook_dispatch` only
+when an actual `SubagentStop` hook is configured; otherwise it selects the
+first-class `explicit_dispatch` route, where the orchestrator runs the verifier
+after developer completion.
+
+## Model choices
+
+The adapter leaves model selection to the local Claude Code configuration. Set a
+supported `model` field in the installed subagent frontmatter when needed. Keep
+the developer on the strongest available implementation profile and keep
+Evaluator and Reviewer independently configured when separate judgement is
+important.
+
+## Global installation
+
+```bash
+./adapters/claude-code/install.sh --global
+```
+
+This installs agents, command, and skills below `~/.claude/`, but deliberately
+does not install a project runtime or activate lifecycle hooks globally.

package/adapters/claude-code/install.sh

@@ -0,0 +1,75 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+usage() {
+  cat <<'EOF'
+Usage:
+  ./adapters/claude-code/install.sh --project <control-repository-path>
+  ./adapters/claude-code/install.sh --global
+
+Install the Triad Claude Code adapter without overwriting existing agents,
+commands, skills, hook fragments, or project runtime. --project installs into
+<control-repository-path>/.claude and also installs .triad-runtime. --global
+installs only agents, command, and skills under $HOME/.claude.
+EOF
+}
+
+if [[ $# -eq 2 && "$1" == "--project" ]]; then
+  target_root="$2/.claude"
+  project_root="$2"
+  project_install=true
+elif [[ $# -eq 1 && "$1" == "--global" ]]; then
+  target_root="$HOME/.claude"
+  project_root=""
+  project_install=false
+else
+  usage >&2
+  exit 2
+fi
+
+adapter_root="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
+repository_root="$(cd -- "$adapter_root/../.." && pwd)"
+agent_names=(triad-developer triad-evaluator triad-reviewer)
+skill_names=(triad-loop-bootstrap triad-loop-orchestrator triad-loop-developer triad-loop-evaluator triad-loop-reviewer)
+collisions=()
+
+for agent_name in "${agent_names[@]}"; do
+  [[ -e "$target_root/agents/$agent_name.md" ]] && collisions+=("agents/$agent_name.md")
+done
+for skill_name in "${skill_names[@]}"; do
+  [[ -e "$target_root/skills/$skill_name" ]] && collisions+=("skills/$skill_name")
+done
+[[ -e "$target_root/commands/triad.md" ]] && collisions+=("commands/triad.md")
+if [[ "$project_install" == true ]]; then
+  [[ -e "$target_root/triad-hooks.json" ]] && collisions+=("triad-hooks.json")
+  [[ -e "$project_root/.triad-runtime" ]] && collisions+=(".triad-runtime")
+fi
+
+if [[ ${#collisions[@]} -gt 0 ]]; then
+  printf 'Installation aborted; existing paths would be overwritten:\n' >&2
+  printf '  %s\n' "${collisions[@]}" >&2
+  printf 'Remove or rename those paths deliberately, then run the installer again.\n' >&2
+  exit 1
+fi
+
+mkdir -p "$target_root/agents" "$target_root/commands" "$target_root/skills"
+for agent_name in "${agent_names[@]}"; do
+  cp "$adapter_root/.claude/agents/$agent_name.md" "$target_root/agents/$agent_name.md"
+done
+cp "$adapter_root/.claude/commands/triad.md" "$target_root/commands/triad.md"
+for skill_name in "${skill_names[@]}"; do
+  cp -R "$repository_root/skills/$skill_name" "$target_root/skills/$skill_name"
+done
+if [[ "$project_install" == true ]]; then
+  cp "$repository_root/integrations/claude-code/hooks.json" "$target_root/triad-hooks.json"
+  cp -R "$repository_root/runtime" "$project_root/.triad-runtime"
+  cp -R "$repository_root/schemas" "$project_root/.triad-runtime/schemas"
+  cp "$adapter_root/runtime.json" "$project_root/.triad-runtime/adapter.json"
+fi
+
+printf 'Triad+ Engineering Loop installed in %s\n' "$target_root"
+if [[ "$project_install" == true ]]; then
+  printf 'Control-plane runtime installed in %s/.triad-runtime\n' "$project_root"
+  printf 'Review .claude/triad-hooks.json and merge it into trusted Claude settings only if hook dispatch is desired.\n'
+fi
+printf 'Start Claude Code in the target project and use /triad followed by a PRD source or project request.\n'

package/adapters/claude-code/runtime.json

@@ -0,0 +1,10 @@
+{
+  "schema_version": 1,
+  "id": "claude-code",
+  "binary": "claude",
+  "lifecycle": {
+    "kind": "SubagentStop",
+    "agent_type": "triad-developer",
+    "requires_async": false
+  }
+}

package/adapters/codex/README.md

@@ -0,0 +1,39 @@
+# Triad+ Codex adapter
+
+Codex custom prompts live in the user prompt directory, so its native command
+is `/prompts:triad` rather than the un-namespaced `/triad` used by OpenCode and
+Claude Code. The prompt places the current conversation in the Orchestrator
+role; it does not create a hidden coordinator session.
+
+Install a project-control repository first:
+
+```bash
+./adapters/codex/install.sh --project /path/to/project-control-repository
+```
+
+Then install the global entry point and the skills once for the Codex user:
+
+```bash
+./adapters/codex/install.sh --global
+```
+
+The installer refuses any overwrite. It installs the project runtime under
+`.triad-runtime` and the five shared skills under `.agents/skills`; the global
+installation also makes the skills available to Codex and writes
+`$CODEX_HOME/prompts/triad.md` (`$HOME/.codex` by default).
+
+Start Codex from the control repository and invoke:
+
+```text
+/prompts:triad <PRD path or project request>
+```
+
+Before it starts work, configure the four named role profiles described in
+[`docs/codex-replication.md`](../../docs/codex-replication.md). The adapter
+does not edit profile or model configuration, because their availability and
+model routing are host-owner decisions.
+
+At bootstrap and resume, the Orchestrator runs the runtime capability detector.
+It chooses the configured Codex asynchronous hook only when that hook is really
+available; otherwise it performs explicit verification dispatch. On Codex CLI
+0.142, explicit dispatch is the safe expected route.