Skip to content

Fix task/tasks skill: fast single-call lookup, run, and create-offer#33

Merged
skapoor8 merged 1 commit into
mainfrom
fix/task-speedup
Jul 19, 2026
Merged

Fix task/tasks skill: fast single-call lookup, run, and create-offer#33
skapoor8 merged 1 commit into
mainfrom
fix/task-speedup

Conversation

@skapoor8

@skapoor8 skapoor8 commented Jul 19, 2026 •
edited
Loading

Copy link
Copy Markdown
Contributor

Intent

Make /task (and its /tasks alias) fast. This skill sits on a hot path — the spec, em, dev, and gcp skills all call /task detect / /task list before running any build/test/lint. Each call was taking ~2 minutes for a job that is really ~1ms of shell. This branch collapses that into a single deterministic script call, and makes no-arg invocations list tasks, query invocations run the matching task, and unmatched queries offer to create a runner-native task.

Bug

Detecting the project's task runner took ~2 minutes per /task call. The detection work itself is trivial (a few test -f marker checks plus one jq on package.json, ~1ms). The slowness was structural: the model did the work by transcribing-and-executing bash across four markdown files. SKILL.md dispatched to references/workflows/detect.md; list.md re-inlined the same detection and its list command; run.md did it a third time. Every "read this file, transcribe its bash, run it, read the output, move to the next file" hop is a slow LLM round-trip. Four files × round-trips = minutes for a millisecond job.

Fix

Move the deterministic part into one shipped script and reduce the model's role to "run one command, read its output."

  • New skills/task/scripts/task.sh — single POSIX-sh source of truth for detect/list/run and the no-match offer, one process per call:
    • no-arg / list → detect runner, print RUNNER: <name>, then the task listing (or a no-runner message with a mise.toml suggestion).
    • detect → runner name plus run/list commands in stable KEY=value form.
    • run <query> [args…] → detect, resolve query (exact → unique prefix/substring), exec the runner on match; on no match, exit distinctly and print the task list plus a NO_MATCH:<query> marker.
  • skills/task/SKILL.md rewritten to a thin dispatcher: every verb maps to one bash scripts/task.sh … call. No-arg → list; an unrecognized first token → run <query> (so /task test, /task build --release just work); explicit detect/list/run still take precedence. On NO_MATCH, it offers a runner-native entry (mise [tasks.*], just recipe, Taskfile.yml task, package.json script — or a scaffolded mise.toml if no runner), gated behind user confirmation.
  • detect.md / list.md / run.md thinned to short reference docs describing the script contract — no executable bash on the hot path.
  • /tasks alias reconciled and kept in sync: .mise-tasks/gen-task-alias regenerates skills/tasks/ from skills/task/ (identical except the name/trigger token), wired via mise run skills:sync-task-alias with a drift --check folded into skills:validate. docs/skills/task.md updated for the new behaviour.

Detection order (mise → just → Taskfile → package.json → none) and cross-skill references (/task detect, /task list, /tasks list) are unchanged — no other skill needs to change.

Fixes:

Validation

  • mise run skills:validate passes — 44/44 SKILL.md checks plus the alias drift --check
  • /task with no args lists tasks in a single Bash call (no multi-file dispatch)
  • /task <known-task> runs the matching runner command; extra args pass through
  • /task <unknown> offers a runner-native create snippet for the detected runner
  • Existing /task detect|list|run verbs and aliases produce equivalent output

@skapoor8 skapoor8 self-assigned this Jul 19, 2026
skapoor8
skapoor8 commented Jul 19, 2026
View reviewed changes

@skapoor8 skapoor8 left a comment

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Automated review: 7 inline comments (3 BLOCKING run-path bugs). See details inline.

Comment thread skills/task/scripts/task.sh Outdated
Comment thread skills/task/scripts/task.sh Outdated
Comment thread skills/task/scripts/task.sh Outdated
Comment thread skills/task/scripts/task.sh
Comment thread skills/task/scripts/task.sh Outdated
Comment thread .mise-tasks/gen-task-alias Outdated
Comment thread skills/task/scripts/task.sh
skapoor8 added a commit that referenced this pull request Jul 19, 2026
@skapoor8
perf(task): deterministic single-script runner dispatch + /tasks alias
3d1459e 
Model-driven dispatch (SKILL.md -> detect.md -> list.md -> run.md)
made `/task detect` and `/task list` take ~2 minutes: every verb
walked several reference docs and ran ad-hoc bash the model
assembled per call. Replace that hot path with one shipped POSIX-sh
script so detection, listing, running, and the no-match offer happen
in a single deterministic process.

**Single shipped dispatcher**
- Add scripts/task.sh: one source of truth for runner detection
  (mise -> just -> Taskfile -> package.json -> none), listing, and
  running.
- No argument lists tasks; an unrecognised first token is treated as
  a run query and matched against the runner's task names; a query
  with no match returns NO_MATCH so the skill can offer to create a
  runner-native task entry instead of failing.
- SKILL.md maps every verb to a single Bash call; the three workflow
  docs shrink to thin references pointing at the script, so no
  executable bash remains on the hot path.

**Correct lookup signals**
- List names via `mise tasks ls --no-header` so the first task is no
  longer dropped by a header-skipping awk (mise 2026.3.x emits no
  header); dropping it made `/task <first>` wrongly offer to create
  an existing task.
- Capture enumeration stderr: a runner present but unable to
  enumerate (e.g. untrusted mise on a fresh clone) is reported as
  ENUM_FAILED, surfacing the trust error rather than offering to
  create the task.
- Two or more prefix/substring candidates return AMBIGUOUS with the
  candidate list so the query is disambiguated, not treated as
  no-match.
- Match literally via `grep -F` and a lowercased case-glob prefix;
  the query is never interpreted as a regex.

**/tasks alias with a drift guard**
- Ship skills/tasks/ as a generated mirror of canonical skills/task/
  (identical but for the name, `/tasks` trigger token, and SKILL.md
  H1), so both triggers share the same fast script.
- Add .mise-tasks/gen-task-alias to regenerate it and a
  skills:sync-task-alias task; fold a --check drift guard (content
  and exec-bit modes) into skills:validate so CI fails if the two
  diverge.
- Update docs/skills/task.md for no-arg list, bare-query run, and the
  create-offer.

Closes #33
@skapoor8
perf(task): deterministic single-script runner dispatch + /tasks alias
d56c89f 
Model-driven dispatch (SKILL.md -> detect.md -> list.md -> run.md)
made `/task detect` and `/task list` take ~2 minutes: every verb
walked several reference docs and ran ad-hoc bash the model
assembled per call. Replace that hot path with one shipped POSIX-sh
script so detection, listing, running, and the no-match offer happen
in a single deterministic process.

**Single shipped dispatcher**
- Add scripts/task.sh: one source of truth for runner detection
  (mise -> just -> Taskfile -> package.json -> none), listing, and
  running.
- No argument lists tasks; an unrecognised first token is treated as
  a run query and matched against the runner's task names; a query
  with no match returns NO_MATCH so the skill can offer to create a
  runner-native task entry instead of failing.
- SKILL.md maps every verb to a single Bash call; the three workflow
  docs shrink to thin references pointing at the script, so no
  executable bash remains on the hot path.

**Correct lookup signals**
- List names via `mise tasks ls --no-header` so the first task is no
  longer dropped by a header-skipping awk (mise 2026.3.x emits no
  header); dropping it made `/task <first>` wrongly offer to create
  an existing task.
- Capture enumeration stderr: a runner present but unable to
  enumerate (e.g. untrusted mise on a fresh clone) is reported as
  ENUM_FAILED, surfacing the trust error rather than offering to
  create the task.
- Two or more prefix/substring candidates return AMBIGUOUS with the
  candidate list so the query is disambiguated, not treated as
  no-match.
- Match literally via `grep -F` and a lowercased case-glob prefix;
  the query is never interpreted as a regex.

**/tasks alias with a drift guard**
- Ship skills/tasks/ as a generated mirror of canonical skills/task/
  (identical but for the name, `/tasks` trigger token, and SKILL.md
  H1), so both triggers share the same fast script.
- Add .mise-tasks/gen-task-alias to regenerate it and a
  skills:sync-task-alias task; fold a --check drift guard (content
  and exec-bit modes) into skills:validate so CI fails if the two
  diverge.
- Update docs/skills/task.md for no-arg list, bare-query run, and the
  create-offer.

Closes #33
@skapoor8
skapoor8 marked this pull request as ready for review July 19, 2026 19:41
@skapoor8
skapoor8 merged commit c95361b into main Jul 19, 2026
2 checks passed
@skapoor8
skapoor8 deleted the fix/task-speedup branch July 19, 2026 19:41
github-actions Bot pushed a commit that referenced this pull request Jul 19, 2026
@semantic-release-bot
chore(release): 1.66.1 [skip ci]
391958f 
## [1.66.1](v1.66.0...v1.66.1) (2026-07-19)

### Performance Improvements

* **task:** deterministic single-script runner dispatch + /tasks alias ([#33](#33))

Model-driven dispatch (SKILL.md -> detect.md -> list.md -> run.md)
made `/task detect` and `/task list` take ~2 minutes: every verb
walked several reference docs and ran ad-hoc bash the model
assembled per call. Replace that hot path with one shipped POSIX-sh
script so detection, listing, running, and the no-match offer happen
in a single deterministic process.

**Single shipped dispatcher**
- Add scripts/task.sh: one source of truth for runner detection
  (mise -> just -> Taskfile -> package.json -> none), listing, and
  running.
- No argument lists tasks; an unrecognised first token is treated as
  a run query and matched against the runner's task names; a query
  with no match returns NO_MATCH so the skill can offer to create a
  runner-native task entry instead of failing.
- SKILL.md maps every verb to a single Bash call; the three workflow
  docs shrink to thin references pointing at the script, so no
  executable bash remains on the hot path.

**Correct lookup signals**
- List names via `mise tasks ls --no-header` so the first task is no
  longer dropped by a header-skipping awk (mise 2026.3.x emits no
  header); dropping it made `/task <first>` wrongly offer to create
  an existing task.
- Capture enumeration stderr: a runner present but unable to
  enumerate (e.g. untrusted mise on a fresh clone) is reported as
  ENUM_FAILED, surfacing the trust error rather than offering to
  create the task.
- Two or more prefix/substring candidates return AMBIGUOUS with the
  candidate list so the query is disambiguated, not treated as
  no-match.
- Match literally via `grep -F` and a lowercased case-glob prefix;
  the query is never interpreted as a regex.

**/tasks alias with a drift guard**
- Ship skills/tasks/ as a generated mirror of canonical skills/task/
  (identical but for the name, `/tasks` trigger token, and SKILL.md
  H1), so both triggers share the same fast script.
- Add .mise-tasks/gen-task-alias to regenerate it and a
  skills:sync-task-alias task; fold a --check drift guard (content
  and exec-bit modes) into skills:validate so CI fails if the two
  diverge.
- Update docs/skills/task.md for no-arg list, bare-query run, and the
  create-offer.

Closes #33
@github-actions

github-actions Bot commented Jul 19, 2026

Copy link
Copy Markdown

🎉 This PR is included in version 1.66.1 🎉

The release is available on GitHub release

Your semantic-release bot 📦🚀

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

@skapoor8 skapoor8

Labels

released

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@skapoor8