Review Flow

This page walks through what happens when the agent makes a tool call, from the moment Claude Code fires a hook to the moment the reviewer's decision is enforced.

The hook pipeline

Zwischen installs hooks for all eight Claude Code hook events:

Hook	When it fires
SessionStart	A new Claude Code session begins
SubagentStart	The agent spawns a subagent
SubagentStop	A subagent finishes
PreToolUse	Before a tool executes
PostToolUse	After a tool returns successfully
PostToolUseFailure	After a tool returns an error
UserPromptSubmit	When you send a message to the agent
Stop	When the agent tries to end its turn

Each hook sends an event to the local daemon, which streams it to the Zwischen server for review.

Pre-execution review

Tools in the pre_review and write_review lists are reviewed before they execute. The flow:

1. Agent calls a tool (e.g., Bash with a shell command)
2. PreToolUse hook fires
3. Client sends the tool name and input to the reviewer
4. Reviewer evaluates and returns a determination
5. If pass: the tool executes normally
6. If block: the tool is denied, the agent sees a block message with the reason

For write_review tools (Edit, Write), the reviewer sees the proposed file diff and can reject it before the file is modified.

Post-execution review

Tools not in the pre-review lists are reviewed after execution. The PostToolUse hook streams the tool's result to the reviewer. The reviewer can still issue a block at this point — the tool already ran, but the agent will be stopped from taking further actions until the block is resolved.

Stop review

When the agent tries to end its turn, the Stop hook fires. The reviewer decides whether the work is actually done. If the reviewer determines the agent stopped too early, it issues a stop block — the agent is forced to continue working, with the reviewer's feedback injected into the session.

Visibility tiers

Not every tool call gets the same level of review. The tools configuration in your settings controls what the reviewer can see:

Tier	What happens
output_visible	Reviewer sees the tool's full output
pre_review	Reviewed before execution; can be blocked
write_review	Like pre_review, but shows the proposed file diff
whitelisted	Passed through without blocking, but still streamed as context
review_exempt	Completely invisible to the reviewer

Tools not in any list are streamed to the reviewer with their name and input parameters only — the output is not shown. See Tool Visibility for configuration details.

Event ordering

Events are processed in strict FIFO order. The reviewer sees them in the same order the agent produced them. This is enforced by a dedicated worker thread per reviewer session — there is no parallelism in the review pipeline for a given agent.