Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 2 additions & 2 deletions skills/slopaudit/references/skills.md
Original file line number Diff line number Diff line change
@@ -1,8 +1,8 @@
# Skill quality

What lint and evals cannot judge: whether a skill is worth routing to and
cheap to load. Run `skillcheck lint` first; evidence beats stylistic
preference.
cheap to load. Run `npx -y @uinaf/skillcheck lint --root <dir>` first on each
dir holding `skills/`; evidence beats stylistic preference.

## Metadata and discovery

Expand Down
6 changes: 4 additions & 2 deletions skills/slopaudit/references/sources.md
Original file line number Diff line number Diff line change
Expand Up @@ -8,14 +8,16 @@ describe confidential contexts generically.

| Harness | User prompts | Full sessions |
|---|---|---|
| Claude Code | `~/.claude/history.jsonl` (`display`, `project`, `timestamp`) | `~/.claude/projects/<cwd-slug>/<session>.jsonl`; subagents under `<session>/subagents/` |
| Claude Code | user messages in full sessions; `~/.claude/history.jsonl` misses desktop-app prompts | `~/.claude/projects/<cwd-slug>/<session>.jsonl`; subagents under `<session>/subagents/` |
| Codex | `~/.codex/history.jsonl` (`text`, `ts`) | `~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl` (`response_item` messages) |
| T3 Code | — | `~/.t3/userdata/state.sqlite` (`projection_thread_messages`, `projection_threads`) |
| T3 Code | — | `~/.t3/userdata/state.sqlite` (`projection_thread_messages`, `projection_threads`); if absent, Claude Code sessions whose slug names a T3 worktree |

Before counting:

- Keep only what the user typed: drop tool results, reminders, hook output,
and skill bodies.
- Count button clicks that arrive as prompts, such as next-step buttons,
apart from typed prompts: they are stamps, not repeated instructions.
- Separate subagent sessions from top-level ones. Count skill loads for both,
corrections for top-level only.
- Exclude eval and sandbox sessions (usually a working directory under a temp
Expand Down
15 changes: 10 additions & 5 deletions skills/slopcourier/SKILL.md
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,7 @@ Three lanes. Start where the request starts, stop where it stops:

| Request | Lanes |
|---|---|
| "open the PR" | Open, then stop |
| "open the PR" | Open, then Babysit until `ready`; merge only when asked |
| "get it merged", "babysit" | Open (if needed), Babysit, After merge |
| review or CI feedback on an open PR | Babysit, After merge |
| direct push or dispatched workflow | After merge |
Expand Down Expand Up @@ -41,7 +41,7 @@ Three lanes. Start where the request starts, stop where it stops:
came from the owner's `<owner>/.github` template, and `-footer <line>` when
an attribution line is required. Fix what it flags or say why it's wrong.

Done when the PR is open: report its URL and head commit.
Report the PR's URL and head commit, then Babysit.

## Babysit

Expand Down Expand Up @@ -90,16 +90,21 @@ feedback.
- **A human's design question or ambiguous point:** stop rework on that PR,
post nothing, and hand it back to the user with the evidence.
- **Concrete fixes** (from humans or bots): validate each against the code.
Fix the real ones in one pass, run the gates, push, reply on each fixed
Fix each real one's whole defect class across the diff, not only the
flagged line, in one pass. Run the gates, push, reply on each fixed
thread with the commit link alone, and re-request review from anyone who
requested changes or approved an older head. A human's finding you judge
wrong goes to the user with the evidence; post nothing. Out-of-scope
requested changes or approved an older head. Reject a wrong bot finding
with a one-line reason, then resolve its thread; resolve fixed bot threads
too. A human's finding you judge wrong goes to the user with the evidence;
post nothing. Out-of-scope
findings become tracker items.
After the third fix round on one PR, stop patching, reconsider the
approach, and tell the user if it changes.
- **Failed check:** read the log. Fix what this change broke. If the base has
since fixed it, bring in the base. A flake or infra failure gets one rerun;
the same failure twice is real.
- **Behind or conflicted:** bring in the base, resolve conflicts, rerun the
gates, push, and keep watching.
- **`slopguard`:** only when the user asks or the repo has no review bot or
required review; once, on the final head.

Expand Down