Context manifest

“What did the model actually see?” is the first question after a surprising result. Azhi records the answer for every turn of every agent step.

What a manifest holds

• The items that went into the context: system prompt, instructions, inputs, retrieved excerpts, tool results.
• The source of each item.
• Token counts per item and per turn.
• What is unobservable. If an executor compacts context itself (OpenCode does), Azhi says so rather than pretending to know the final prompt.

npx azhi inspect <run-id> --context analyse

Usage next to it

Each turn also records usage: input, output, cache read, cache write and reasoning tokens. Cost comes from declared pricing; otherwise it is shown as unavailable. See Cost and token control.

Agent activity

The run page’s Agent activity tab keeps what each agent did while it ran: system prompt, prompts, replies, reasoning, and every tool call with its input and result (or error), plus each model call’s token counts.

Executor	How entries are captured
opencode	The worker follows OpenCode’s event stream for the step’s session
claude-agent-sdk	The SDK’s assistant messages, tool uses and results
codex	codex exec --json items
model agent	The server records each turn

Entries are redacted twice (on the worker and on the server) using the step’s secrets and token-shaped patterns such as provider keys, GitHub, Slack and AWS tokens, bearer headers and private keys. Each text field is kept up to 32 KB and a step attempt keeps up to 2,000 entries. Set AZHI_AGENT_TRANSCRIPTS=0 to stop storing transcripts.