A Connect conversation turn fails
Inmeerkat-connect, a confirmed failed turn can
return to the You: prompt. Enter a new question there to continue the same
conversation. Connect does not automatically retry the failed request.
If Connect reports an uncertain effect, follow its restriction to status checks
for that exact request. A new question does not authorize repeating the effect.
If the CLI closes instead, starting it again creates a new conversation; it does
not resume or replay the old request.
Bootstrap failed or was interrupted
Do not rerun bootstrap into the same client, output or workspace. It preserves partial directories; a timeout or lost response can occur after an owner has written valid receipts. A receipt’s presence alone does not establish that the whole bootstrap completed.- Preserve the selected archive, command, fixed error code and partial outputs. Check that the bootstrap/installer process has settled before further work.
- Inspect the exact client
installed.json, hostprepared.jsonand available step logs. Compare them with the original inputs. Bootstrap normally verifies the client through the installer’s--verify-existing; do not replace that check with a directory-exists test. - If the host validates a complete existing client, use
prepare_host.pywith its exact interpreter/hash and fresh output/workspace paths. Otherwise use a new bootstrap client, configuration and workspace. Preserve the previous attempt for inspection.
Preparation refuses paths or runtimes
The setup MCP worked once and refuses to restart
Its generated command creates one fresh workspace. It is not a restartable server command for that same directory. Keep one setup process alive for the original task, or use saved-work readback with new configuration/workspace paths after the old workers settle. Do not delete the old workspace to make the original command run again.A saved setup file does not resume installation
That file selects retained requests. Use--recover-manifest for readback or
--prior-manifest plus an exact request ID for a new configuration proposal.
These are different workflows. Recovery restores
no install or retry tools. matched is a current file observation; the old
outcome can remain unconfirmed. Missing acquired files remain unresolved.
If preparation rejects a saved file, verify its exact hash, original workspace,
request anchors and original host bindings. Keep configured-install and
acquisition-only requirements distinct. Do not rewrite hashes, discard a
selected request or substitute a newer library to bypass the refusal.
The connector installed but no tools appear
Installation emits a launch; the host still needs to approve and register it. For Meerkat, follow the native MCP mapping and attachment guide. Astaged result can precede a missing-executable or connection failure.
Check native catalog search/load and an actual operation, keeping the setup
server’s name and process stable. Two accounts can expose colliding tool names;
use the host’s explicit naming map instead of assuming server-name prefixes.
Report a setup problem
Send the relevant host or release maintainer a bounded report containing:- Archive filename/SHA-256 and platform, architecture, libc and runtime versions.
- The selected family/profile and exact command flags, with private values removed.
- Which step failed, its fixed error code, job/request identifiers and whether the process settled or the result is unconfirmed.
- The relevant receipt/check result and the smallest useful log excerpt.
find_and_install returns unavailable
Check that the application enrolled its native setup owner, that the owner is active, and that the invocation carries the required authenticated context. Constructing a dispatcher alone does not bind that owner. See host composition.A recipe exists but discovery cannot find it
Check the compiled include list incrates/toolkit-package/src/library.rs. Drive 0.1.0 and Slack 0.3.0 files currently exist outside that registered inventory. Also confirm whether the host chose the metadata-only catalog or the versioned library. A stale paging snapshot requires a fresh listing.
A read fails with a digest or unknown-field error
Use the exact identity, package digest and listed library digest. Do not substitute a newer version, normalize a publisher/name or omit required joined-recipe identity. The tools reject unknown fields and a guidance-only strategy cannot be promoted by adding made-up execution data.Python refuses an unprepared client
Callprepare_httpx_client once at the host construction site before sharing the client. Keep its response hook first throughout borrowed use. require_prepared_httpx_client checks without mutating. Do not reorder hooks during active operations or patch around the refusal in each call.
An SDK call succeeded but the agent cannot use the tool
The SDK operation and the agent’s runtime binding are separate. Observeidentity_resolved_tools for the actual member and perform a real dispatch through that surface. Check profile selection, package/grant revisions and the current target incarnation.
Check the application’s restoration, helper and spawn configuration when a tool is available in one member but missing in another.
Startup shows progress but never reports ready
Inspect the native startup result and the configured startup budgets. A timeout or lost reply calls for reconciliation through the startup owner before another attempt. See native application lifecycle.Authentication completed but access is denied
Check actual account identity, requested operations/resources, host grant and provider scope separately. Developer registration or administrator installation may still be missing. A saved connection is not a grant, and a broad provider credential does not establish the intended local restriction.A write timed out or returned an HTTP error
Inspect the typed effect outcome. Once entered, a write can remain unknown even on an error response. Preserve the original business invocation and reconcile according to that provider/domain contract. Retry-After is not permission to replay; a Gmail Message-ID search is not proof that an earlier send did not happen.Slack retrieval finished but the summary is incomplete
Inspect intervalcoverage_gaps, chain states and budgets. Older roots, hidden/deleted content, retention, permission and concurrent changes can limit results. finished only means the collector exhausted queued work. One page, a short page or an exhausted visible chain is not a global completeness proof.