Macro is a unified workspace for teams: email, chat, docs, tasks, agents, calls, and CRM — @-linked together with shared AI memory.
Agent effort capabilities
Effort belongs to a selected model in a particular harness. Native API support,
Claude Code support, and Cursor variants are different contracts. The UI reads
ACP configOptions, selecting category thought_level (legacy id fallback:
reasoning_effort). External ids and values are opaque; ultra, for example,
is not a native Macro effort level.
Cursor
The authenticated Cursor catalog supplies concrete model variants. Discovery uses the selected model's default variant; an active session uses its effective variant. Only effort values with identical remaining parameters are shown. Changing effort preserves those parameters and selects an actual catalog variant. Automatic selection, missing reasoning parameters, or fewer than two compatible values produce no effort control. No static provider effort list is used.
In-memory
The engine's model catalog controls model availability. Native effort profiles
are defined in crates/agent/src/model/reasoning_effort.rs; adding a model to the
catalog does not guess its effort support from its name. Update the profile and
provider serialization tests when adding support.
| Routed model | Explicit effort values |
|---|---|
| anthropic/claude-sonnet-5-5, anthropic/claude-opus-5-5 | low, medium, high, xhigh, max |
| openai/gpt-5.5 | none, low, medium, high, xhigh |
| openai/gpt-5-mini | minimal, low, medium, high |
| Haiku 4.5, retired Claude ids (Sonnet 5, Opus 5), and unknown models | No effort control |
Every supported profile also offers Default, meaning no session override.
It preserves Macro's existing adapter defaults (including low for GPT-5 mini).
It is distinct from OpenAI's explicit none and is never sent as a provider
value. Anthropic receives output_config.effort; OpenAI Responses receives
reasoning.effort. A compatible Chat Completions endpoint does not inherit
native OpenAI capabilities merely because its model name resembles GPT.
Thinking token budgets are not exposed as effort levels.
Sources checked September 21, 2026: Anthropic effort, GPT-5.5, and OpenAI reasoning.
Session lifecycle
POST /agent-capabilities/discoveraccepts a selected model for Cursor and in-memory. The existing/agent-models/loadAPI remains available.- Preflight capability cache identity includes the harness and model. A selection from another target is discarded. External harnesses expose controls after opening a session; model-specific preflight is not guessed.
- A model chosen at creation is the session's model from its row. In-memory and Cursor runtimes start on it; the harness selects it over ACP for the others and holds the first prompt until the runtime accepts it.
- An effort chosen at creation waits for the runtime's reported configuration, checks the effort choices, then confirms effort before sending the prompt. HTTP queue acceptance alone is insufficient. Rejection or timeout stops startup.
- New/resumed sessions and setting changes return the complete configuration. Clients replace their previous options, including removing unsupported controls.
- In-memory rejects unsupported models and effort values. A model change resets an incompatible effort to Default. Replay restores confirmed snapshots only; failed or unanswered requests do not become session settings.
The ACP configuration contract documents full snapshots, semantic categories, and opaque values.
Probe evidence and limits
Prompt-free local ACP probes initialized a session, switched advertised models, set advertised effort values, and tried an invalid value. No inference request was sent. Observed capabilities are version- and account-specific:
| Harness tested | Observations |
|---|---|
| Claude Code 2.1.278, claude-agent-acp 0.73.0 | Sonnet/Opus: default, low, medium, high, xhigh, max; switching to Haiku removes effort. Invalid value rejected. |
| Codex CLI 0.154, codex-acp 1.8.0 | GPT-5.5: low/medium/high/xhigh; GPT-5.6 Sol adds max/ultra. Initial Astra session omitted effort. Invalid value rejected. |
| OpenCode 1.18.23 | Tested Haiku 4.5 and Sonnet 4.5 advertised high/max; Opus 4.5 advertised low/medium/high. Initial model omitted effort. Subsequent probe startups failed. |
| Hermes 0.20.4 | Legacy model/mode metadata, no configOptions observed. |
Sixteen advertised effort changes were accepted with complete configuration responses. These observations verify ACP advertisement and acceptance, not the provider request made during inference. OpenClaw was unavailable. Cursor and in-memory are covered by repository ACP adapter tests and provider serialization tests; these are deterministic tests, not claims of live provider execution.
To reproduce the protocol check with an authenticated harness, launch its ACP
stdio command in an empty temporary directory; send initialize, session/new,
and session/set_config_option using the returned session id and advertised
config ids/values. Re-read the full configOptions after every model change.
Record harness and adapter versions. Never infer a native provider profile from
an external harness result.
Model menu interaction
The selected model also has a speedometer-style effort button beside its model selector in new-chat and running-session composers. It opens a compact desktop popover or a dedicated bottom sheet on mobile. Its needle reflects the selected level. A simple list shows the supported effort levels and checks the current choice. Selecting a level applies it and closes the menu; models with fewer than two choices hide the button.
On desktop, effort remains available in each model’s submenu. Discovery runs when the submenu opens, scoped to that model and harness; the current session's advertised options take precedence for its selected model. The trigger displays the model and selected effort together. Keyboard Right Arrow opens the submenu. On mobile, tapping a model selects it directly; use the adjacent effort button to adjust its supported effort levels. Loading, discovery errors, and models with no effort options keep ordinary model selection available.
A combined running-session selection waits for the model's correlated ACP confirmation, validates effort against the new snapshot, then confirms effort. A model rejection prevents the effort request. An effort rejection leaves the accepted model in place and reports the failure. New-chat choices remain bound to their model while discovery refreshes, so immediately sending cannot silently drop a selected effort.