Output, environment, and limits¶

Output selection¶

--output human and --output json override automatic selection where the command supports structured output. Interactive TUI JSON is rejected. ANSI and control sequences are emitted only after an interactive-terminal check.

For run, human output deliberately omits run/session identifiers, provider routing, event counts, and timing. Use --output json when those fields are required.

Renderers consume post-policy released contracts. They do not change execution, authorization, persistence, or audit semantics. Untrusted released content is sanitized and never interpreted as terminal control input.

Environment contract¶

• Configuration stores credential references such as env:VARIABLE, not values.
• Sandbox configuration grants environment variable names, not values.
• Permit-bearing adapters resolve credentials only after authorization.
• Process helpers start from a cleared environment and pass only declared names.
• Pack and MCP environment references are bounded by both their declaration and the deployment sandbox.
• Raw credentials, authorization headers, private keys, key material, and hidden reasoning are hard-redacted from policy input, provider output, transcripts, and audit evidence.

Default operational bounds¶

These are baseline configuration values, not permission grants. Adapters may enforce additional hard protocol and input bounds. Exact effective values come from config show, config effective, and the relevant command help. See Runtime limits configuration for field interactions, validation ranges, and tuning examples.

Provider raw-stream bytes, model output tokens, and process-tree memory are separate limits. The provider byte count includes the full SSE framing and metadata before normalization. The memory value is an enforced effect-process-tree ceiling, not memory preallocated by the Colossus process. Policy, adapter, and request-specific obligations may narrow either configured ceiling and never widen it.

Important hard bounds¶

• Provider tool arguments receive at most two correction turns.
• Workflow call depth is at most 16.
• Workflow schedule cadence is 60 seconds through 31 days.
• The TUI queues at most eight future turns.
• The TUI loads at most 1,000 submitted-history entries.
• Pre-effect logical policy input is bounded at 1 MiB. Post-effect policy input is bounded at 8 MiB so the base64 envelope for a permitted 4 MiB result remains inspectable without widening the pre-effect request boundary.
• Workflow conditions are non-executable and bounded by size, token, recursion, and boolean-composition limits.

Completion and uncertainty¶

A successful transport status alone is not application completion. Provider streams need a terminal item. Effects need a durable terminal event. If execution may have escaped but completion evidence is missing, the outcome is unknown. Colossus preserves already released content but never synthesizes completion or blindly retries a non-idempotent external effect.