triad-plus 1.9.0 → 1.11.0

package/skills/triad-loop-reviewer/SKILL.md

@@ -30,6 +30,22 @@ mismatched, or unreported required repository skill is a `blocked` result until
 the Orchestrator creates a valid assignment; external context never substitutes
 for the repository skill policy.
 
+Before inspecting the candidate, prove the actual Reviewer runtime context by
+running this command from the activation cwd:
+
+```bash
+node /absolute/path/to/control-workspace/.triad-runtime/triad-runtime-context.mjs \
+  --project /absolute/path/to/control-workspace \
+  --assignment /absolute/path/to/control-workspace/.loop/runtime/assignments/<assignment-file>.json
+```
+
+Record packet `cwd`, process `cwd`, `pwd`, `git rev-parse --show-toplevel`, and
+each repository-skill path, resolved path, and SHA-256. Compare the actual cwd
+and Git root with the assigned product worktree, and resolve every relative
+skill path under that worktree only. On `NotFound`, hash mismatch, mixed Git
+roots, or cwd inconsistency, return `blocked` / invalid runtime context; never
+fall back to a copy in the control workspace or a global skill.
+
 Use native `read`, `grep`, `glob`, and `list` primitives for simple reads when
 the host provides them. Use shell/Bash for builds, tests, git, scripts, and
 system commands, not as the default filesystem API.

package/skills/triad-model-configuration/SKILL.md

@@ -0,0 +1,98 @@
+---
+name: triad-model-configuration
+description: Inspect and safely configure Triad+ role models through the canonical team configuration and the selected host adapter.
+---
+
+# Triad+ model configuration
+
+Use this skill when the owner asks to inspect, explain, or change the models
+used by Triad+ roles in a control workspace. Keep the operation small,
+deterministic, and reversible.
+
+## Source of truth
+
+1. Identify the requested project-control workspace. Do not guess a workspace
+   from the current directory when more than one is visible.
+2. Read `.triad-runtime/adapter.json` and `.triad-plus/team.json` from that
+   workspace. The team file is the canonical source of role configuration.
+3. Validate the schema version, role IDs, and model/reasoning value types before
+   proposing or applying a change. A malformed team file is a blocker; do not
+   repair it by inference.
+4. Show the current values and the adapter binding mode before changing them.
+
+Never edit generated host files as the first step. Preserve every field in
+`team.json` that the owner did not request, including display names, personas,
+interaction settings, Evaluator+ enablement, and policy fields. Do not put
+tokens, API keys, credentials, or private deployment data in the team file.
+
+## Requested changes
+
+Interpret only an explicit owner request. A request may set a role `model`, a
+role `reasoning_effort`, clear either value to `null` (host default), or ask for
+an inspection without changing anything. Do not invent model IDs from a
+marketing name or from memory. If the host exposes a reliable native model
+list, present only IDs from that list; otherwise ask for an explicit ID or keep
+the host default.
+
+Reject unknown fields, malformed values, ambiguous role names, and requests for
+a host-native field that the selected adapter cannot materialize. Explain the
+limitation instead of silently substituting another model or pretending that a
+field was applied.
+
+When a change is authorized, update only the requested keys in
+`.triad-plus/team.json`, preserving its JSON structure and all other values.
+Use the existing managed binding path to materialize the result (normally an
+`upgrade --apply` for the selected host). Do not duplicate frontmatter/TOML
+parsing in a shell command or in this skill. Preview the change first when the
+host workflow offers a dry run, then apply it only after the requested values
+are clear.
+
+After applying, reread `team.json`, verify that only requested fields changed,
+and inspect the generated host assets. A failed materialization is a failed
+operation; do not claim success from the team file alone.
+
+## Adapter binding modes
+
+Read the adapter metadata rather than branching on a host name:
+
+- `global-profiles`: the adapter writes supported role values to its native
+  user-level profiles (for example Codex). Verify the resulting profile files.
+- `project-frontmatter`: the adapter writes only the fields listed by its
+  metadata to managed project agent definitions. OpenCode supports `model` and
+  its native `variant` field (the canonical `team.json.reasoning_effort` value
+  is materialized as that variant); Copilot supports `model` and
+  `reasoningEffort`; Claude Code currently materializes `model` only for its
+  supported roles. A field not listed by the adapter remains host/session
+  managed and must be reported as such.
+- `team-record`: the adapter records the owner's intent in `team.json` but has
+  no native model file to update. Report “recorded in team.json; host-native
+  materialization unavailable” and do not fabricate an agent binding.
+
+For every requested role/field, classify the result as **host-native applied**,
+**recorded only**, **host/session default**, or **unsupported**. Null or blank
+values mean host default and should not create artificial frontmatter/profile
+values. Reasoning levels are not portable across hosts; never translate a
+level or model ID from another provider. For OpenCode, inspect the host's
+available model variants (for example with `opencode models`) before requesting
+one; a role profile's `variant` is separate from the primary OpenCode session
+default.
+
+OpenCode's managed binding is the only writer for `.opencode/agents/*.md`.
+Never patch `reasoningEffort` into those profiles by analogy with Copilot. The
+same binding is reapplied after `upgrade --apply`, while null values remove
+stale managed fields and restore the host default.
+
+## Safe completion report
+
+Report:
+
+- control workspace and selected adapter;
+- fields inspected and fields changed (or “no changes”);
+- binding mode and host-native fields supported;
+- generated assets checked and their result;
+- any unsupported/recorded-only limitation;
+- confirmation that unrelated `team.json` fields and secrets were preserved.
+
+This skill configures role models only. It does not change Triad authority,
+retry policy, verification, Evaluator+ behavior, host routing, or workflow
+state, and it does not perform releases or publish packages.

package/skills/triad-model-configuration/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Triad Model Configuration"
+  short_description: "Inspect and safely configure Triad role models"
+  default_prompt: "Use $triad-model-configuration to inspect or change the requested Triad+ role models in a control workspace."