Workflow

Themis runs one workflow definition, THEMIS_WORKFLOW in @aegis-skills/themis-core, in both modes. It has seven steps.

Steps

Step	Kind	What happens
validate	deterministic	Cleans the task (control characters removed, at most 4000 characters, no long base64 blocks, no <script or javascript:) and checks the context (at most 20 environments and 20 attack-surface tags).
decompose	reasoning	Plans the task as sub-tasks, each assigned to one Aegis skill. Unknown skills are dropped, at most 8 sub-tasks are kept, and dependency cycles are broken.
schedule	deterministic	Works out which sub-tasks are ready, given their dependencies.
fanout	reasoning	Runs the ready sub-tasks and returns findings with a confidence of high, medium or low.
guardrail	reasoning	Gives each result a verdict: PASS, FLAG or BLOCK. A result containing a secret is always BLOCK, whatever verdict is given, and an unknown verdict counts as BLOCK.
synthesise	reasoning	Writes the report from the results that were not blocked.
audit	deterministic	Completes the run once a report exists.

Deterministic steps are plain code and behave the same in both modes. Reasoning steps are done by your assistant (plugin) or by the configured model (standalone).

Transitions and guards

From	To	Guard checked before entering the next step
validate	decompose	The task is not empty after validation.
decompose	schedule	The plan has at most 8 sub-tasks.
schedule	fanout	At least one sub-task is ready. Chosen while sub-tasks are ready to dispatch.
schedule	guardrail	Every sub-task has finished. Chosen when nothing is left to dispatch.
fanout	schedule	The plan has at most 8 sub-tasks, so dependent sub-tasks run in later rounds.
guardrail	synthesise	Every result has a guardrail verdict.
synthesise	audit	A report exists.
audit	end	The run completes.

If a guard fails, the run ends as failed with the reason. If a reasoning step's output has the wrong shape, the call is rejected and the run is unchanged, so the step can be retried.

Run states

A run is running, awaiting_approval, completed or failed. The standalone app sets awaiting_approval while a tool action waits for a human, and the run resumes when the last pending approval is decided.

One definition, two runtimes

The plugin runs the definition directly: it executes the deterministic steps itself and hands your assistant the prompt for each reasoning step. The standalone app maps each step onto a LangGraph node, so the graph is built from the same definition.