Agent Lens
Agent Lens turns a live agent pane into something you can scan in seconds instead of scrolling. It is a second presentation mode over a terminal pane whose PTY is already running — not a second pane, and not a transcript viewer bolted on the side.
Session and Terminal are the same pane
A pane running a supported agent offers two modes:
- Session — one chronological narrative of what the agent is doing.
- Terminal — the raw libghostty surface, exactly as it always was.
Switching between them may detach the terminal renderer, but it does not replace the terminal surface, restart the foreground process, replay input, or discard scrollback. You are looking at one pane two ways. Terminal is always available as exact re-entry.
What Session shows you
The default projection is built to answer four questions fast:
- What is this agent trying to accomplish?
- What is it doing right now?
- Does it need my judgment?
- What materially happened, in order?
Everything lives on one evidence-ordered timeline: conversation, tool lifecycles, subagent work, interaction requests and diagnostics. Session does not split "chat" from "activity", because the thing you are reconstructing is a single sequence of events.
Some deliberate choices in how it condenses:
- Instructions are context, not events. System, developer, project and skill instructions stay collapsed in the inspector unless you ask for them. They explain the agent; they are not things it did.
- Finished work folds, it does not vanish. Adjacent tool, plan and change facts collapse into one quiet row, and expanding it returns every original fact with its stable ID. Nothing is deleted to make the view tidy.
- Pending questions and approvals are promoted into the session summary. Those are the rows that need you.
- Evidence is one action away, and the timeline itself shows only what you need to scan safely: authority and freshness.
- Reported and observed are distinguished. An agent saying it ran the tests and a test run Tenon watched happen are not the same claim, and Session does not flatten them into one.
A narrow pane reflows Markdown tables into labelled fields rather than making you scroll sideways — a supervision surface that needs horizontal scrolling is not doing its job.
Answering a question from Session
When the agent is showing a question, you can answer it in Session. Tenon sends one provider-specific selection frame — Claude receives the option hotkey plus Return, Codex receives only the committing hotkey — into the live TUI. It is the same keystroke you would have typed, delivered to the same process.
Clicking a file the agent cited
When an agent writes a path in its prose, that path is resolved and clickable: the click opens the file in a pane beside it. This is the shortest version of the product's whole argument — the claim and its evidence in the same window, one click apart.
When it cannot bind a session
Agent Lens needs an authoritative binding between the pane and the provider's own session record. It will not guess.
Specifically, it rejects a stale process, a child-agent fact, a mismatched process group, or a rotated terminal-surface token. When binding is unavailable it degrades explicitly: it tells you it is degraded and hands you Terminal, which was the exact evidence path all along.
This is the correct failure. The newest transcript in a directory, picked by modification time, is a guess — and a supervision tool that guesses which session you are looking at is worse than one that admits it does not know.
Codex needs its hook installed
For Codex, check that the additive hook was installed in the active CODEX_HOME, that the provider approved it, and that the transcript is a current-user regular JSONL file under CODEX_HOME/sessions. See Troubleshooting.
What Agent Lens is not
It is not a complete transcript renderer, and it does not replace the terminal as the source of truth. It is a projection whose value is that it is shorter than the transcript while still resolving to it.
It is also not the Attention Inbox. Agent Lens works on one session at a time. The cross-session inbox — explicit states like needs_input, approval, failed, ready_for_review across every workstream, with a "since you last looked" capsule — is described in the vision and is not built yet.
Agent Lens adds no public plugin path, no core intent, no audience and no control-plane operation. It is host-internal, which is why a plugin cannot reach into it.