Skip to main content
WORK INWorkstream MCP
BRING

Brand access · Saved baseline · Representative cases

FINISH WITH

A verified saved version and evaluation evidence

Outcome: one justified configuration change with a saved version, evaluation evidence, and a read-back of the applied result. This walkthrough updates a Workstream workflow brand, not a Response support-agent record or an active-horizon recipe release. Choose a small observed problem, such as a response instruction that contradicts your approved policy. Avoid changing routing, review thresholds, model settings, and policy text together: you need to understand what caused the result.

Prerequisites

Connect to Workstream MCP with access to the intended brand. Have an authorized reviewer, representative cases, and the policy or evidence supporting the change. Evaluation may invoke models and consume resources even though it does not apply the candidate to the brand. If you have no existing workflow or real case yet, start with Your First Workflow Trial.

1. Read the baseline

Discover the exposed tool schemas, then read get_brand_config and the relevant scorecard or run evidence. These configuration tools use the argument brand; other operational tools may use brand_id. Inspect the schema rather than assuming one name everywhere. Retain the full deterministic_config, its config_version, the brand identity, and the observed problem. Keep the baseline with the proposed change so a reviewer can compare them. Use check_config_drift where available to identify divergence before editing.
Success: your proposal will be based on a known saved version and a specific failure case.

2. Propose a patch without saving

propose_config_patch accepts the brand and a list of operations. Its supported operations are set, remove, append, and replace_text; it is not a generic implementation of every JSON Patch operation. Use a path observed in the actual configuration and the operation’s current schema. Ask your assistant:
The tool returns proposed_config, changes, validation, and based_on_config_version. It does not save the candidate. Preserve that complete candidate; do not reduce it to the changed fragment for the later apply call.

3. Validate the candidate you intend to apply

Call validate_config with the brand and the returned proposed_config as config. Inspect valid, errors, and warnings. Omitting config validates the brand’s currently saved configuration. That can succeed while your proposed change remains invalid, so confirm which object was checked. Validation checks configuration structure and binding rules. It does not prove answer quality, provider readiness, or that the desired business outcome will occur.
Success: the exact candidate passes structural validation, and any warnings have been reviewed.

4. Compare behavior before applying

Use get_eval_cases or approved hand-written cases, then call evaluate_config with: With simulate_tools: true, recorded case context supplies reads and mutations are recorded as proposed actions rather than executed. This is not a live connector test. Inspect missing context and proposed actions as well as generated text before trusting the result. Use evaluate_golden to check the candidate against the brand’s standing human-corrected cases. An evaluation result informs the decision; do not mistake an earlier advisory evaluation for the apply-time gate. If there are no usable cases, build and review the relevant evidence before claiming that the change passed regression checks. Keep Response’s evaluation records separate from Workstream golden cases. A Response run_eval result does not automatically satisfy this brand’s workflow gate.

5. Apply the authorized candidate

After approval, call apply_config with these fields populated from the reviewed work: A brand’s required golden gate still applies when the caller omits require_golden_pass. Omitting the flag is not a way to bypass policy. Resolve the reported prerequisite or failed result rather than removing controls. If the version conflicts, another change may have landed. Fetch the latest configuration, review the difference, and rebuild/re-evaluate the proposal. Do not simply replace the expected version number and send the old candidate, which could erase someone else’s work. An authorized apply task can be explicit:

6. Verify persistence and subsequent behavior

Inspect the apply result, including applied, config_version, and config_fingerprint when returned. Then read get_brand_config independently and compare the saved values. Use list_config_versions and get_config_version for the version history and snapshots exposed by your connection. Watch relevant subsequent runs. Keep the new config version, run IDs, observed replies/actions, and unresolved regressions with the change record. A successful apply proves persistence; it does not itself prove better live outcomes or change the brand’s routing mode. If a response was lost, read the saved state and version history before deciding whether to repeat the write. Do not infer success solely from an HTTP connection or a tool invocation.

If the change needs to be reversed

Read the current state and the earlier snapshot. Treat restoring prior values as another reviewed, validated, evaluated change against the current version. An older snapshot is not automatically suitable if connectors or other policies have changed since it was captured. This process changes future configuration. It does not unsend replies, undo provider actions, or establish that in-flight runs now use the restored values. Inspect affected executions and handle their outcomes separately.

Troubleshooting

Next steps

Measure the result

Compare subsequent outcomes with the baseline you saved.

Resolve a blocker

Diagnose access problems or a stalled workflow.
Related: Workstream MCP reference.
Last modified on September 20, 2026