AI DevKit · Track dev-lifecycle / structured-debug progress on a durable task with the ai-devkit task CLI. Use to record phase, progress, next step, blockers, and validation evidence.
SkillScore64Maintained
Automatic
◆Project Management pack
Last verified 8/12/2026
What it changes
Passed 9 of 14 test cases with the skill, and 7 of 14 without it.
Without skill50%
With skill64%
Checks it passed that the model alone missed
Invokes task assign with task name order-export.
Includes kind flag in the artifact command invocation.
Invokes task next search-api with the next step description.
and 3 more in the full report
6 fixed8 passed either way5 still failing3 broke
Pass-rate lift
+14 pp
Tokens
+104%
Agent turns
0%
Response time
+4%
Model gemini-3.6-flashMeasured 8/12/2026v114 of 22 cases passed with skill · 14 comparable
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 6 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 6
Prompt
An existing task named 'order-export' was created earlier by another worker. We just identified our agent as `--agent worker-a --agent-type subagent --pid 1020 --session s-123`. Show the CLI command to claim ownership of 'order-export'.
What was checked, without and with the skill
Checked
Without skill
With skill
Invokes task assign with task name order-export.
Missed
Passed
case-08 · gemini-3.6-flash · 8/12/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md126 linessingle file
Task Progress Tracking
Record development progress on a durable task: phase, progress, next step,
blockers, and validation evidence.
Requires the optional task command. Use npx ai-devkit@latest for task and
agent commands. Before recording task events, run a real read probe:
npx ai-devkit@latest task list --json
# or, when a task name is known:
npx ai-devkit@latest task list --name <task-name> --json
Only treat task tracing as available when the read probe exits 0. If it fails,
continue without task logging and include the failed command plus stderr/stdout
summary in the final report. Do not block the user's work just because optional
task tracing is unavailable or unusable.
Core idea
One task per work item. Create it once; advance its phase field as work
moves through the lifecycle or debug workflow.
<id> can be a task name. Every command below accepts the task name in
place of a task id, resolving to the latest non-terminal task. Prefer
<task-name> so agents do not track task ids.
Choose stable names. For lifecycle work, use the feature key as the task
name. For debugging or review work, choose a short kebab-case task name.
Emit at checkpoints, not streaming. Phase transitions, task toggles,
immediate next-step changes, fresh evidence, blockers discovered/resolved. A
handful of calls per session.
Sequence mutations. Never run task mutation commands in parallel for the
same task. Each mutation reads the current task snapshot and writes it back;
parallel writes can clobber snapshot fields even though events append. Run
create/assign/phase/next/progress/evidence/blocker/artifact/close commands
one at a time, then read back with show --events --json when the final state
matters.
Attribution is explicit. Identify self once, then pass actor flags on
mutation commands.