How to use
- Copy the installation command and review the source package before installing it in your own agent environment.
- Check the source document for required tools, credentials and supported hosts.
- Describe your input and desired deliverable; review the output against the source material.
npx skills add https://github.com/flyingnobita/flyingnobita-deliver --skill deliver-implementBefore you start
- Installation does not include a model subscription, API allowance or permission to access your files.
- Repository stars and installation counts describe activity, not an independent quality or security test.
Rendered from the source repository. Headings, examples, code, tables, links, and referenced images are preserved.
Implement tickets
You are the orchestrator. The tickets are the agreed spec: there is no brief, grilling, plan, reviewer role or QA. Each ticket gets its own worktree, branch and implementer worker running implement (or mattpocock-implement); you verify each result yourself and fast-forward the target branch to it, one ticket at a time.
Run to completion without asking, except where a step below names a stop. When a ticket blocks, notify the user at once and keep landing independent tickets. The final report is the user's notification that the run is done.
Use the shared delivery files, resolved from this skill's directory: host detection, the selected adapter in adapters, configuration, run records and scheduling. Run the runtime helper as python3 <its path>.
1. Resolve the run
- Host: follow host detection and select one adapter. A host named in the
request is used only when that session is confirmed; otherwise stop with the failed check.
- Tickets: resolve each reference, a GitHub issue (
26,#26) or Kaneo
key (DLV-26), to both identities (kaneo task list -p <PROJECT> --all) and read it fully, with its "Blocked by" edges and GitHub issue dependencies. Name every ticket as DLV-26 (#26) when a Kaneo board exists. An ambiguous reference is a stop.
- Graph: reject cycles. A blocker that is closed or merged is satisfied; a
blocker outside the requested set and still open is a stop.
- Target branch: a branch named in the request, else the current branch of
your checkout. Record its name and SHA and state both in your first status line. Detached HEAD or a remote-tracking ref is a stop.
- Settings: resolve per configuration, using only the implementer
role and these keys: max_workers (default 3, capped by any host limit) and human_schedule (default normal). A worker count or schedule in the request overrides them.
- Model: the request's override (harness, model, effort, and mode on
Herdr; a profile ID on Paseo), else the project implementer role, else the adapter default. Validate it the adapter's way, passing only the implementer table. An override that is unclear, unknown or fails validation is a stop before anything is created: show what you parsed, why it failed, and the presets (project role, adapter default, Paseo profile IDs) to choose from.
- Human steps: a human ticket, or a human step inside an agent ticket, is
scheduled by human_schedule as in scheduling. Load wizard (or mattpocock-wizard) for each one and notify the user with the script path when it becomes ready. Its dependents wait until the user reports it done.
- Run record: on Herdr or Orca,
createwithpurpose: "implement", the
resolved concurrency, the implementer profile and the target branch as base, then approve with approved.tickets holding the ticket list and the user's request as evidence. On Paseo, open the adapter's run directory instead and record the host, target name and SHA, implementer profile, max_workers, tickets and request there; Paseo keeps native IDs and never creates a runtime record.
- Baseline: run the full suite and linters on the target tip as the repo
defines them. A failing baseline is a stop.
Done when the run record is approved and the baseline is green.
2. Dispatch
The frontier is every ticket whose blockers have landed or were already satisfied. While a worker slot is free, take the next frontier ticket in the request's order:
- Claim it: set the repo's in-progress
status:label when the repo uses
status labels (only you change statuses).
- Create its workspace through the adapter, branched from the target's current
tip, on a branch named <kaneo-key>-<slug> (else <issue>-<slug>). On Paseo pass that tip's SHA as baseBranch, not origin/<default>. On Orca pass the name with --branch: --name alone yields <git-user>/<name>. Use the branch the host reports in every later step.
- Start one implementer worker in it with: the ticket reference, the command
to run implement (or mattpocock-implement) for that ticket alone, the repo's commit rules (or the configured commit_skill), a Fixes #N footer for a GitHub issue, the target branch name, and a last step before reporting: git rebase <target> onto its current tip, resolve any conflict, and re-run the tests. Its report: commits, the target SHA it rebased onto, and test and lint commands with one status line each.
Refill the frontier each time a ticket lands.
3. Land each finished ticket
Land one ticket at a time, keeping the target's history linear: no merge commits, so a later publish accepts the range.
- Record the worker's report:
reporton Herdr,orca-reporton Orca, and on
Paseo a run-directory note of its commit, target SHA and check lines. In the ticket's worktree, with the tree clean, run git rebase <target-sha>; when that moved the branch, record it with rebased on Herdr or orca-rebased on Orca (onto, commit, evidence). The runtime accepts only a rebase whose patches equal the reported ones, so the work it verifies is the work the worker reported. On Paseo, write that receipt in the run directory instead: onto, before/after SHAs and the patch-id comparison. On a conflict run git rebase --abort.
- Run the full suite and linters there yourself, then
verifythe branch tip
(orca-verify on Orca; on Paseo, save the commands, output and result in the run directory). Make any red-proof probe in a temporary copy of the worktree, never in the worktree itself. A conflict, a refused rebase claim or a failure goes back to the ticket's worker once with the receipt (followup on Herdr, send_agent_prompt on Paseo): it rebases its branch onto the new target tip, resolves any conflict and re-runs its tests. Repeat steps 1-2 on its next report. A second failure blocks the ticket. A target that moved before verification is not a failure: rebase again and record the new receipt.
- Fast-forward the target to the rebased branch tip. When the target is
checked out, run git merge --ff-only <branch> in that checkout; git refuses rather than overwrite local changes, and that refusal blocks the ticket. Otherwise run git update-ref refs/heads/<target> <branch-sha> <target-sha>.
- Set the repo's done or review
status:label. Leave the issue open: its
Fixes #N footer closes it when the user publishes.
- Clean up through the adapter: close the worker, then remove its workspace
and worktree with the adapter's single removal call (herdr worktree remove --workspace on Herdr, worktree rm on Orca, archive_workspace on Paseo); never close a Herdr workspace first, or the removal loses its ID. Then run git branch -d <branch>, which succeeds because the target now contains the branch.
When a ticket blocks, notify the user through the adapter's notify operation with the ticket, the failed command and receipt, and its worktree and workspace. Keep both, hold its dependents, and continue independent tickets. Check the ticket again at each later wait; once the user reports a fix, land it from step 1. A ticket the user drops is recorded as abandoned with their words as evidence; its dependents stay blocked.
Done when every ticket has landed, is blocked, or is abandoned.
4. Report
Record the run's end with the runtime's checkpoint: stage complete when every ticket landed, otherwise blocked, with the landed range as evidence. On Paseo, append the same final state to the run directory's run note instead.
Report: the target branch and its final SHA; each ticket as landed (its commit range), blocked (reason, kept worktree and workspace) or abandoned; the requested and effective implementer profile; the verification commands you ran; and the issues that close on publish. Nothing is pushed: publishing is a separate request. Suggest deliver-review (or flyingnobita-deliver-review) on the landed range when the combined change warrants it, such as tickets that touched the same module or a ticket the user fixed by hand.