Validation and rollout#

Independent backends — 5 November 2026#

Current default onboarding selects hosted ClickUp + GitHub CLI. ClickUp stays on official OAuth MCP with no ClickUp CLI/app/helper token. Reuse gh sign-in across authorized repositories; verify each target without creating a new token per project. --github hosted selects an existing hosted connection instead. Strict --mode remote-only / legacy --mode oauth never invokes gh; local-helper / legacy cli retains its separate runtime. No automatic fallback.

Native ClickUp–GitHub activity linking/automations and the coding agent's direct GitHub actions are separate connections. No native integration or automation is enabled by this installer. The 5 November Kilo report records a bounded pass for hosted ClickUp + gh on Kilo 7.8.18 with implementation commit 1abd0a79439f4b1b980fb74234416202c87766cf. After reload, all six skill bodies matched the pinned source. The same GitHub account read two repositories. The agent created one child, added one issue backlink while preserving the original body, and found both on a second pass with zero writes. Hosted GitHub acceptance remains pending after the separate all-hosted test lacked a credential. Earlier dated sections describe their original revisions, not the new default.

Historical remote-only mode alignment — 27 October 2026#

The mode guide describes current generated manifests, profiles and onboarding. Tests use isolated homes/projects and mocked service fixtures. They check that remote-only/the-then-default/legacy-oauth onboarding never authenticates local ClickUp or GitHub CLIs, while local-helper/legacy-cli retains its runtime. Generated configuration and JSON/TOML contracts are not installed-client or OAuth acceptance. No real client installation, new OAuth authorization, live ClickUp mutation or organizational policy change was exercised in this implementation. The historical Kilo pilot used gh and does not prove end-to-end remote-only GitHub support. #9–#12 retain live acceptance/safety work; #13 retains journal parity feasibility. PR #6's separate worker write-budget defect remains tracked under #12/#23 and is not repaired here.

Automated coverage#

python3 scripts/check.py runs offline behavioral tests, bytecode compilation, manifest/config validation, source-attribution checks, and skill-frontmatter checks. The tests use fresh temporary Git repositories and mocked service data. Real subprocess tests cover both the workflow MCP server and the REST MCP client against an isolated protocol fixture. They do not measure real-world token cost or server latency, and do not verify credentials or third-party service uptime.

Three independent skill scenarios were exercised with minimal task-local facts:

1. MCP-only execution with a merged PR and pending integration CI: the agent chose host tools, preserved reviewer text, and withheld completion.
2. A partial PR-link failure plus a proposed parent from a different workspace: the agent preserved the successful GitHub side, required complete remote reconciliation evidence, and rejected the out-of-scope subtask parent.
3. MCP-only issue-before-child creation, a timed-out child write, compact task identity, 40 comments/replies, and pending PR CI: the agent preserved the successful issue, required complete remote recovery evidence, skipped mirror echoes, and withheld completion. Ambiguities around host identity fallback, reply export policy, host receipts, and shared provenance were clarified.

New behavior tests cover GitHub-first creation, lost issue responses, resumed child/backlink writes, explicit repository targets, changed-input rejection, issue body preservation, two-way comment snapshots, edits, host provenance, GitHub pagination/review types, and incomplete ClickUp-history rejection.

Packaged installer validation — 20 October 2026#

The suite now includes isolated installation and update checks with a fixture npm CLI. These verify preservation of JSONC comments, unrelated settings and servers, managed TOML updates, legacy default-config migration, skill conflicts, failed dependency installs, repeated setup, and subprocess arguments containing spaces. An actual npm archive is unpacked and installed, then its downloaded source is deleted before the persistent MCP initializes and lists all nine workflow tools. No API accounts are used by these offline tests.

A separate Linux smoke install fetched @krodak/clickup-cli@latest from npm, verified its installed version (1.46.2 at test time), created the Python environment, copied five skills, and wrote project-scoped Kilo configuration in an isolated directory. This version is observed evidence, not an installation pin. No real ClickUp/GitHub account authentication or task writes were performed in this test. Installed editor discovery and authenticated client behavior remain live team acceptance steps. The CI matrix runs Python 3.12 and Node 22 on Linux/macOS/Windows.

Worker and rollout preparation — 21 October 2026#

The offline suite now contains 96 passing tests. The optional polling worker has checks for explicit task scope, a read-only preview, preflight across every task/target before writing, per-pass write limits, binding changes, echo suppression and repeat runs. Tests lose responses after successful writes on each service and verify marker-based recovery without duplicates; absent remote evidence blocks replay. Old sync receipts remain recoverable. Two ClickUp tasks sharing a PR have separate new receipt keys while preserving the canonical source provenance.

OS-lock tests use an actual competing process, kill it, and verify the lock is released. Linked worktrees share the same writer lock; separate clones and machines still need one agreed owner per task. The packaged installer test deletes the downloaded source, initializes the persistent MCP, and loads the worker's command help. No editor or live account is involved in these checks.

Additional fixtures cover exact plain issue-comment endpoints and literal multiline file inputs, malformed paginated records, inline review replies, and excluding pending/unsubmitted reviews. These establish offline behavior only.

The reviewed upstream sources were rechecked: cup comments and REST MCP list_comments still expose one page. The worker refuses a pass that would reach 25 ClickUp comments with those default readers. The explicit clickup-api comment-reader override adds GET-only cursor pagination, with fixtures for multiple pages, inclusive overlap, equal timestamps, malformed or stalled cursors, redirect rejection, and credential/error isolation. Task reads and all writes still use the configured upstream runtime. No live worker service, credentials, native connection, or automation was installed. See the worker guide for its current boundaries and recovery.

Kilo official OAuth pilot — 22 October 2026#

A user-run Kilo session authenticated clickup_official via its existing browser OAuth session and read the approved Workspace, test task, statuses and official tool catalog. The token-backed cup backend was not used. The pilot then performed two reported remote writes on the approved PIP-Technical-Team/wbpip fixture:

1. Created one child 868mmcp85 below 868mjzgzh, after verifying the existing GitHub issue #269. The child contained the issue URL, [cu-work-github:...] association and unique [pilot-op:...] marker.
2. Updated only the body of issue #269 to append the child backlink, preserving the original content.

The Kilo agent reported a guarded second pass with zero additional writes, correct parent/list/status, and unchanged draft PR #270. Independent read-back later confirmed the new child and the single issue-backlink block. ClickUp's markdown_description escaped/reformatted the markers, while text_content preserved them literally. This distinction is now mandatory in the host workflow guidance.

Kilo 23 October acceptance: all six skills, ClickUp OAuth, real task access and agent-enforced checks passed. Two installer issues were found: unvalidated alternate-name ClickUp MCP endpoints, and a GitHub remote MCP that Kilo could not authorize because it lacks dynamic client registration. The installer now rejects unsafe/duplicate ClickUp aliases and uses authenticated gh for Kilo, disabling only an unchanged previously installer-managed GitHub remote entry. A subsequent Kilo live test is still necessary to confirm the corrected migration on the user's machine.\n\nLimits: This proves a successful agent-controlled existing-issue reuse sequence, not a local-helper transaction or guaranteed exact-once write. It does not validate new-issue creation, induced network timeouts, simultaneous writers, production comment threading, MCP quota accounting or host OAuth behavior in every supported client. No permanent worker, native integration, or OAuth-backed bridge to clickup_work was installed.

Live rollout acceptance#

Choose a disposable or team-approved test task before broad rollout. Each teammate authenticates their own tools. Configure the real workspace/list/root and exact statuses; do not put tokens into the binding.

Run one task read, one small subtask, one progress comment, and a link to a draft test PR. Repeat with identical operation keys and verify no duplicates. Verify the PR's original body stays intact. Confirm the helper refuses an unrelated task and refuses done while the PR is draft/open or CI is pending.

Run the same read/update on the free REST MCP runtime. For host official/context and GitHub MCP, discover current schemas and verify the skills' manual scope and recovery sequence. Verify every teammate gets only their own account's access.

Record elapsed time, calls, and returned bytes for equivalent tasks if tuning routing priorities. Change project overrides based on those measurements. Do not claim a faster/cheaper backend merely from the default ordering.

Live ClickUp access was verified through the official host MCP on 2026-10-19: one scoped repository-associated child, one linked progress comment, repeat-read reuse, and an in-progress/backlog status round trip. Initial GitHub issue creation returned HTTP 403. On 2026-10-20 the user installed the GitHub App for the actual repository owner organization, resolving the missing installation. The follow-up created issue #269 and draft PR #270 in wbpip, verified reciprocal task links and PR body preservation, transferred labeled conversation fixtures in both directions, and reused exactly one mirror per canonical source-event key on a second pass. The child and parent remain open; the original parent is unchanged.

See the pilot record for source/destination IDs and exact scope. These are live host-MCP/skill checks. Local CLI/REST runtime credentials, plain issue-comment writes, live inline review/thread transfers, native integration/automation behavior, and real installed-client plugin loading remain separate acceptance steps from the passing offline suite. No always-on comment worker, merge, or deployment was installed.