Skip to content

Releases

v1.6.0 · Build Local removed

Choose a tag to compare

@jonathankems jonathankems released this 04 Nov 18:31
f37a7cf

Removed

  • build-local. Ask your coding agent for a browser preview of the current changes, and it starts the project's development server and checks the page without a skill. execute still opens a preview of viewable changes when it finishes.

Updating existing repositories

npx skills update leaves the removed skill installed and reports it as failed. To remove it, run:

npx skills remove build-local

Add -g if you installed the skills globally.

v1.5.6 · Reliable tracker reads and fresh subagents

Choose a tag to compare

@jonathankems jonathankems released this 29 Oct 10:53
0a59826

Fixed

  • The GitHub tracker template that setup-snappedly-skills writes reads an issue as JSON with its title, state, body, labels, and comments. Outside a terminal, gh issue view --comments prints only the comments, so agents missed the issue's title, body, and labels. To close an issue, the template posts the comment first and then closes it, because gh issue close --comment drops the comment on an issue that is already closed.
  • setup-snappedly-skills creates the triage and category labels its mappings name when the tracker lacks them, so applying a triage state no longer fails on a missing label. The code host template asks for a change-request read that returns the title and body as well as the comments.
  • When diagnosing-bugs or tdd forces a failing test by mutating code or a fixture, it diffs the file against a pristine copy to prove the mutation landed before it trusts the failure. A mutation command can exit successfully and change nothing.

Changed

  • execute, code-review, codebase-design, improve-codebase-architecture, research, and grill start each subagent they launch with a fresh context window, not a fork of the conversation. Some harnesses fork by default, which gives a subagent the context it exists to avoid. help-snappedly applies the same rule to delegated work, and writing-for-agents tells skill authors to state it.
  • code-review searches the repository for its standards sources instead of relying on documents already read. An existing CODING_STANDARDS.md or CONTRIBUTING.md is always a source.
  • grill words each question so that "yes" accepts its recommended answer.
  • When tdd asks which seams to test, it notes for each proposed seam what a test there catches and what it misses.
  • execute states the title of a named issue before it continues, and asks when the reference matches no item or more than one.
  • help-snappedly reads a skill's SKILL.md before it states what the skill does or recommends skipping one of its steps.
  • The GitHub tracker template notes that gh 2.94 and later handles sub-issues and blocking relationships natively, and that reading an issue's blockers also lists closed ones.

Updating existing repositories

Rerun setup-snappedly-skills to replace the issue read and close operations in docs/agents/issue-tracker.md and to create any missing triage and category labels. Setup also replaces a change-request read in docs/agents/code-host.md that relies on --comments.

v1.5.5 · Quick approval for retro recommendations

Choose a tag to compare

@jonathankems jonathankems released this 23 Oct 07:31
df466db

Changed

  • retro ends its report with a horizontal rule and a Recommended changes section. It numbers the proposed changes in the same order as the findings, with one plain-language line per change that says what to change and where. You can reply "fix all" or name the numbers to apply, and retro makes those changes. When your original request already asked retro to make the changes, it applies them without the list.

v1.5.4 · Triage before execute

Choose a tag to compare

@jonathankems jonathankems released this 18 Oct 10:35
79561fb

Changed

  • execute triages a named issue in the needs-triage state before carrying it out. It applies the triage process for a specific issue. When the issue is ready for an agent, triage posts its brief and updates its category and state roles on the tracker before execute builds the plan. Planning specs receive the roles without an executable brief. When triage recommends another state, such as needs-info or wontfix, execute stops with the recommendation, applies only the outcome you choose, and continues only when the issue reaches ready-for-agent. On a planning spec, execute also triages every open ticket in the needs-triage state, including blocked ones. A ticket that reaches ready-for-agent joins the plan once its blockers have closed; any other ticket stays out, and the report gives triage's recommendation for it. Without triage installed, execute stops and reports which issues need triage. triage now supports model invocation in Claude Code and Codex so execute can load it as a dependency from the snappedly/skills package.

v1.5.3 · Clear status for follow-up reports

Choose a tag to compare

@jonathankems jonathankems released this 14 Oct 23:03
8500fb1

Fixed

  • After clean-up or codebase-cleanup, when you ask the agent to act on the Decisions needed and Gaps to check items, it reports each item under its current status: Done, Handed off (with where it is tracked), Accepted, or Still open. Each item starts with its outcome. Previously, the follow-up report could reuse "Decisions" and "Gaps" headings, so fixed items looked open. An item counts as done only once its decision is applied or what it named is verified. A gap left for someone else to verify is listed as handed off.

v1.5.2 · Decisions and gaps before you deploy

Choose a tag to compare

@jonathankems jonathankems released this 14 Oct 19:32
980d21e

Changed

  • clean-up and codebase-cleanup end their report with a horizontal rule and a Check before you submit or deploy section, so open items are not lost in the longer report. It has two numbered lists in plain language. Decisions needed gives each choice left to you, with its options and a recommendation. Gaps to check gives each thing left unverified, such as a failing or skipped check or an uninspected area, with its risk, the ways to close it when there is more than one, and a recommendation. An empty list says "None". Open findings appear only in that section, rather than also in the body of the report.

v1.5.1 · Execute dependency provenance

Choose a tag to compare

@jonathankems jonathankems released this 13 Oct 06:12
1876fed

Security

  • execute states that its dependencies, code-cleanup, code-review, and tdd, come from the same snappedly/skills package as execute itself.

v1.5.0 · Specs and tickets for agents

Choose a tag to compare

@jonathankems jonathankems released this 12 Oct 15:56
7edb449

Changed

  • grill gives the ready-for-agent state to every ticket it publishes, including blocked ones, unless you say otherwise. Without configured triage roles, it applies a ready-for-agent label and creates the label when the tracker lacks it. The state means a ticket is fully specified, not that its blockers have closed. execute on a spec still runs only the tickets whose blockers have closed.
  • grill records blockers through the native blocking operations configured in issue-tracker.md, such as GitHub issue dependencies. Without those operations, it writes a ## Blocked by section. execute reads native links when configured and any existing ## Blocked by sections, so older tickets keep their dependencies.
  • grill also gives the ready-for-agent state to a spec it creates. On a spec, the state means /execute on the spec can carry out its unblocked tickets, and the spec stays non-executable. triage applies the state to a spec without an agent brief, since a brief marks an item executable.
  • When grill starts from an existing tracker issue, that issue becomes the parent. Each ticket becomes its child, and grill posts the spec as a comment on it instead of creating a separate spec. grill never edits or closes that issue.

Added

  • grill looks for refactors that make a slice easier to build. Each refactor goes at the start of the first slice that needs it. The ticket requires behavior to stay unchanged and the relevant checks to pass before the slice adds new behavior. A refactor gets its own ticket ahead of the tracer bullet only when every approach needs it and it is too large to review inside the tracer bullet.
  • grill splits a wide refactor, such as renaming a column used across the codebase, into expand, migrate, and contract tickets: add the new form, move the callers over in batches that each pass the checks, then remove the old form.

Updating existing repositories

Rerun setup-snappedly-skills to add the Blocking relationships section to docs/agents/issue-tracker.md and to let its planning spec mapping accept a spec comment and the ready-for-agent state. Until then, grill writes blockers as ## Blocked by sections.

v1.4.5 · Isolated checks and resumable validation

Choose a tag to compare

@jonathankems jonathankems released this 11 Oct 21:00
0a94b03

Fixed

  • execute workers can run focused checks and prerequisite builds in parallel when their resources are isolated. Required full suites and final builds run on the combined change, unless repository policy requires them earlier.
  • code-cleanup prefers supported resource isolation or shared locks. Resource acquisition defaults to a five-minute timeout. Deferred checks retain their check record and resume once their resume condition is met. Test execution keeps its own timeout.
  • code-review reports deferred checks as unverified with their resume condition. execute, clean-up, submit, and deploy preserve those records and resume the checks once the resume condition is met. submit and deploy stop before delivery and continue from their preparation step when run again. Required deferred checks prevent delivery until verified.
  • setup-snappedly-skills records shared resources, isolation or locking commands, and configured resource-acquisition timeouts. Unknown resource scope stays unverified.

Updating existing repositories

Rerun setup-snappedly-skills to refresh Concurrent runs in docs/agents/workflow.md. Setup preserves team requirements and corrects entries that classify worktree-local output as shared.

v1.4.4 · Concurrent check runs

Choose a tag to compare

@jonathankems jonathankems released this 11 Oct 19:02
61a5211

Added

  • setup-snappedly-skills records which checks two worktrees on one machine cannot run at the same time, in a new Concurrent runs section of docs/agents/workflow.md. Each entry names the shared resource the check holds or removes, such as a fixed port, a container name, or a cleanup of every container carrying a label.

Changed

  • code-cleanup checks for another worktree's run before starting a check listed under Concurrent runs. When a check fails because a port or other resource is taken, or another run removed what the check created, it finds that run, leaves it running, and reruns once after it exits. A repeated collision is recorded as blocked, naming the other run and the resource, rather than as a failure of the change, and is reported as a gap the repository can close by making the check safe to run concurrently.
  • execute workers run only focused checks. Full suites, builds, and the checks listed under Concurrent runs run once on the combined result, since the workers share one machine.

Updating existing repositories

Rerun setup-snappedly-skills to add the Concurrent runs section to docs/agents/workflow.md. Setup fills it from each check's configuration and moves any port or resource note already recorded for a check into it.