Skip to content

Agents

The skills delegate heavy or noisy work to three subagents so the main session’s context stays light — the same context frugality principle the generated harnesses enforce.

AgentModelAccessRole
transcend-stack-detectorhaikuread-only (Read, Bash, Glob, Grep)Detect the project’s stack profile
transcend-generatorinheritread-write (Read, Write, Edit, Bash)Materialize the harness files
transcend-auditorinheritread-only (Read, Bash, Glob, Grep)Inspect and critique an existing harness

Runs the read-only detector scripts (core/detectors/detect.sh), sanity-checks the result against the manifests actually present (package.json, pyproject.toml, go.mod, Cargo.toml, lockfiles), loads the matched stack profile, and returns a compact stack-profile JSON with resolved command variables. Invoked by /transcend-init.

Materializes a harness into the target project’s .claude/ from a fully resolved set of choices and variable bindings — CLAUDE.md, path-scoped rules, settings.json hooks, copied hook scripts, the handoff scaffold, catalog wiring, and the manifest. It never re-runs the interview or changes choices.

Runs in one of two modes:

  • generate — from /transcend-init: write the harness exactly as specified.
  • merge — from /transcend-audit and re-init/upgrade: apply safe fixes (create / regenerate / append / settings-merge) while re-hashing each target at write time, so hand-edited and untracked files are never overwritten.

Read-only inspector for /transcend-audit. Runs the manifest drift verifier, then critiques across the audit dimensions: provenance/drift, missing pillars, over/under-enforcement, stale handoffs, size/import health, portability, and catalog drift. Returns findings JSON with machine-applicable fix plans the generator’s merge mode can execute.

For how the three fit together, see Architecture.

The three above are transcend’s own machinery. The harness can also generate bespoke agents into your project — written to .claude/agents/ and owned by your repo, not the framework. The Delivery Pipeline pillar generates four: pm, architect, coder, and research, each carved to your stack and rules. The Review & Quality pillar’s agent-review-gate option generates a project-local reviewer. These are invoked as Task subagents by the generated orchestration skills.