agentheim

Answer a question about the codebase of an Agentheim-managed project — how a feature works, where something lives, what was decided and why, or whether something is built yet — by reading the project's own structure (the index, the bounded-context READMEs, the ADRs, the task boards) and verifying against the actual source. Read-only. Invoke it directly when you want a thorough, code-grounded answer that separates what's decided from what's actually implemented.

Use whenever the user wants to capture an idea, bug, feature request, refinement, or change to an existing bounded context — anything from "the button should be green" to "we need a whole new subscription subsystem". Also use when the user wants to refine existing backlog items or promote refined items to ready-to-work. Triggers on phrases like "I have an idea", "let's model this", "let's do some modeling", "capture this", "add this to the backlog", "refine the auth backlog", "promote X to todo", "we should also", "what if we added", "there's a bug", "change the color of", "the domain needs to handle". Creates task markdown files in the appropriate bounded context with the right status, and can spawn deeper modeling sessions via the orchestrator. Supports six switchable conversational modes (Interrogator [default], Suggestor, Challenger, Storyteller, Facilitator, Synthesizer) during CAPTURE and REFINE — see references/modes.md.

Use whenever the user or another skill needs external information gathered from the web — state of the art on a topic, comparison of approaches, documentation for a library, domain knowledge from outside the codebase. Triggers on phrases like "research X", "find out about", "what's the state of the art for", "look up how others do", "compare options for", "gather info on", and also triggers internally when brainstorm or model hits an "I don't know enough" wall. Produces a research report in .agentheim/knowledge/research/ that can be referenced from tasks, ADRs, and modeling sessions. Every report passes through a fresh-context `research-reviewer` agent (see `skills/research-review/SKILL.md`) that re-verifies its checkable claims against primary sources before the report is citable.

Use whenever the `research` skill is about to ship a researcher's report — between the researcher's return and the report being treated as citable knowledge. Triggers internally from `research`'s post-write gate. Spawns the `research-reviewer` agent with fresh context (no exposure to the researcher's reasoning) to re-verify the report's decision-critical checkable claims against primary sources and either PASS, FAIL with re-dispatch, or label the survivors unverified. Doctrine document; the operational steps live in `research`'s flow.

Use whenever a worker is about to implement a task with observable behavior — features, bug fixes, refactors that should preserve behavior, any change with acceptance criteria you could turn into an assertion. Triggers internally from the `worker` agent at the start of its third action ("do the work"), and externally when the user invokes it explicitly ("do this TDD-style", "red-green-refactor this task"). Doctrine document; readable by humans and consumed by the worker as its execution discipline.

Use whenever the user wants work executed on the todo backlog — running tasks, building features, implementing what has already been modeled. Triggers on phrases like "start working", "execute the todo", "work on it", "build it", "implement the backlog", "let's go", "run the workers", "pick up where you left off", "ship what's ready". Spawns parallel worker sub-agents that resolve task dependencies and claim ready tasks from `contexts/*/todo/`. Workers follow TDD per `skills/test-driven-development/SKILL.md`. Every worker SUCCESS goes through a `verifier` agent (see `skills/verification-before-completion/SKILL.md`) before commit — failed verification re-dispatches the worker up to twice, then escalates to the user. New tasks promoted to todo during the run are picked up automatically as they become ready. Does not do modeling — only executes already-refined tasks.