Explanations

Most of a report is what Salidium observed. One part of it is prose, and prose has to be written by something.

Salidium hands your own installed Claude or Codex CLI a short, redacted summary of the session and lets it write that part. It is labelled wherever it appears. Nothing leaves unless you choose one of these explanation calls, and you can switch scheduled calls off without losing anything else on the page.

When it runs

Local only

The full deterministic page, with no model calls. This is the default.

When done

One model call after the session finishes or goes quiet.

Each reply

One model call after every agent reply.

Open Models & Usage to choose when explanations run and which agent writes them. Choose a model opens a short list that adapts to the selected agent; Other model keeps manual entry available as a fallback. The same control is available before the first session exists.

Stopping the daemon does not erase the saved mode. Use salidium explanations off as the persistent no-model-call control; local reports keep working and an explanation already generating is canceled.

What it uses

Models & Usage separates the work agent from the explanation helper in one rail. Session usage belongs to the work being read. Explanation usage is the observed all-time ledger. Both are token counts reported by the agent CLI, never a currency estimate.

Claude defaults to the named Haiku model shown in the panel. When no exact model is chosen, Codex selects its own model. Salidium labels that Automatic instead of exposing CLI terminology or guessing a model name.

What it is given

• The last prompt.
• Up to forty of the agent's statements.
• Fifteen file names, shortened to their last two segments.
• The last six check results.

Personalize Why and How

Choose Personalize in the session toolbar. Once this browser tab has a current personalized version, the same control reads Personalized. It opens a compact composer above Why with one Terms and examples field for what you know, comparisons that help, or terms you use.

The save state stays visible under the field as Not saved, Unsaved changes, or Saved on this machine. Apply saves the terms and makes one explicit model call. When a call cannot be made, including in Local only mode, the button reads Save terms and stores them without generating anything. Delete saved terms removes the owner-only local file and the personalized presentation.

Each report has one browser-only personalized version. Applying again replaces it rather than stacking another. A filled icon-button switch labelled Original and Personalized moves between the two versions. Reloading keeps the saved terms but discards the presentation, so choose Apply again to recreate it. Analogies are labelled In your terms.

The terms can never change Verified, Left or Needs you. Personalizing sends them with the generated Why and How, not transcripts, prompts, commands, diffs or raw records. The selected agent may contact its provider. Deleting saved terms cannot delete a record kept by that agent.

What it cannot do

• It runs with its tools switched off, in a directory of its own.
• Its own run never appears as a session in your list.
• What comes back is validated against a schema before it can appear anywhere.
• It cannot change Verified, Left or Needs you.

When there is not one

The panel says which of these applies rather than leaving a gap, and in every case the observed and derived parts of the report are unaffected.

Explanations are off

Nothing was sent to any agent.

No compatible command

No Claude or Codex CLI that Salidium can run was found, so nothing was sent.

Asked, and nothing usable came back

The agent answered but the answer did not validate, and the next turn tries again.

Not yet

One is written when the agent finishes its next turn.

Turn it off and every observed and derived part of the page stays exactly as it is. Nothing else on a report depends on it, and How we know is what labels it wherever it appears.