The human-in-the-loop review journey in Cadora: monitor a run, read the documents it produced at a gate, ask, request a revision, or decide — and the run continues.

Cadora · human-in-the-loop

Reviewing agent work, without leaving the room

Cadora drives a coding agent through a gated workflow and pauses at the steps a human should sign off on. This is the journey a reviewer takes — from a run parked at a gate to a decision that moves it forward.

Surface · local dashboardRole · reviewer / operatorCaptured · fixture run, aidlc-hitl
01

A run is in flight

You start a gated run — cadora run … --hitl --review-file — and open the dashboard. It's the operator's map: every node, its cost, and its state. Nothing asks for you yet; the agent is working.

localhost:8768/runs/pr1-verify
pr1-verifyfixture · aidlc-hitl
running
requirements · analyst
design · architect
construction · engineer
02

A gate opens

The first node finishes and the run parks. An amber callout rises to the top of the page: a node is waiting for your decision, and it hands you exactly the documents that node produced — as links, not a wall of raw JSON.

localhost:8768/runs/pr1-verify
Review requiredrequirements is waiting for your decision
Comments — required to request changes
ApproveRequest changesAbort

The gate is fail-closed and patient. Started with --review-timeout 0, it waits indefinitely — walk away, come back tomorrow, the work is preserved.

03

You read the work

Click a document to open the rendered file, or preview it inline without leaving the panel. The reviewer sees exactly the diff this stage produced — scoped, not the whole tree.

localhost:8768/runs/pr1-verify
Review requiredrequirements is waiting for your decision
ApproveRequest changesAbort
04

You interrogate the work

Before deciding, talk to the work. Pick a document, then ask a question or describe a revision. The parked run drives the same agent that wrote the draft — scoped to that document — and answers in place: a question is answered, a revision rewrites the draft so you can see it before you accept it.

localhost:8768/runs/pr1-verify
Review requiredrequirements is waiting for your decision
execution-plan.md ▾What does step 2 of this plan actually do?Ask ↩Revise ↩
Fixture answer to “What does step 2 of this plan actually do?”. Based on the document: it pauses for human review when the topology marks the node review: true. A real executor would reason over the full document and reply here.

Each ask or revision is a real turn on the agent — you're questioning the model that wrote the draft, not a summary of it. A revision rewrites the document in place, so approving keeps the version you just read.

05

You decide

Three ways forward, each honest about what happens next. Approve releases the next node. Request changes re-runs this stage with your comments woven into its prompt. Abort stops the run. Your comment travels with the decision.

localhost:8768/runs/pr1-verify
Review requiredrequirements is waiting for your decision
Looks good — proceed to design.
ApproveRequest changesAbortsubmitting…

While you're deciding, the page stops auto-refreshing — it will never wipe a half-typed comment or move a button out from under your click.

06

The run continues

Your decision reaches the live run. requirements turns green — clean, approved — and the panel advances to the next gate: design. The comment you left is already in the design node's prompt.

localhost:8768/runs/pr1-verify
Review requireddesign is waiting for your decision
ApproveRequest changesAbort
requirements
cleanapprove
design · in review
07

The pack is complete

Every gate cleared, the run finishes and its evidence pack is sealed. The DAG reads at a glance — three green nodes, each stamped with its gate and review outcome. The audit trail records who decided what, and when.

localhost:8768/runs/pr1-verify
pr1-verifyfixture · aidlc-hitl · 3 nodes
completed
requirements
cleanapprove
design
cleanapprove
construction
gate passed

The screens above reproduce the live Cadora dashboard exactly — rendered as editable HTML rather than raster captures, so they can be lifted into Figma or iterated directly. Every gate is fail-closed, patient, and recorded: the human stays in the loop, and the evidence pack remembers who decided what.