agent-tools

Skills for driving other agent CLIs from inside one, and for keeping a live progress report of the work — without Spec Kit.

Install the Claude Code plugin bundle:

/plugin marketplace add samykabu/sanduq
/plugin install agent-tools@sanduq

Or install a skill on its own:

npx skills add samykabu/sanduq-skills --skill delegate-task
npx skills add samykabu/sanduq-skills --skill progress-report

After installing: delegate-task

The driver is a dependency-free Node script (Node ≥ 18) that lives inside the skill, so where the skill was installed decides where the driver is. Resolve it once:

# plugin install
DELEGATE="$CLAUDE_PLUGIN_ROOT/skills/delegate-task/delegate.mjs"
# npx skills install, or a copy committed into the repository
DELEGATE=".claude/skills/delegate-task/delegate.mjs"

export DELEGATE_RUNS_DIR="$PWD/.delegate/runs"
node "$DELEGATE" doctor

⚠️ Set DELEGATE_RUNS_DIR on a plugin install, and add .delegate/ to the project's .gitignore. Run artifacts hold your task text, the full prompt, and the harness's raw output. Left at the default they are written beside the driver — the shared plugin directory — mixed across every project you delegate from. node "$DELEGATE" prune --keep 20 --yes clears old ones.

doctor probes each CLI's version, so it distinguishes not installed from installed but broken. At least one agent CLI must be on PATH and already logged in; git is what makes the measurement layer work.

See the plugin changelog and the skill's README — also available in Arabic — for worked examples, the result schema, and the three normative contracts. The root README lists every sanduq skill and extension.

After installing: progress-report

A standard-library Python script (3.11 or newer) that lives inside the skill. Resolve it once, from the repository root:

# plugin install
PROGRESS="python $CLAUDE_PLUGIN_ROOT/skills/progress-report/scripts/progress.py"
# npx skills install, or a copy committed into the repository
PROGRESS="python .claude/skills/progress-report/scripts/progress.py"

$PROGRESS init --plan docs/plan.md        # any Markdown checklist; creates .progress/plan/
$PROGRESS hooks install --host claude     # optional: model, effort, context, child agents
$PROGRESS status

Plan mode needs no Spec Kit; Spec Kit projects use init --feature specs/<dir>. See the skill's README for the quick starts, every command and the error codes.

Tests

cd skills/delegate-task && node test/run.mjs
python -m unittest discover -s skills/progress-report/tests -q    # standard library; browser tests skip without Playwright

173 cases. They drive a fake harness through DELEGATE_BIN_<HARNESS>, so the suite spends no tokens and needs no agent CLI installed.

B6 validation (2026-09-27)

Codex exec options, including -s read-only, now precede the resume subcommand. Verified against codex-cli 0.157.1 help and a live sandboxed continuation.

• Driver suite: 173 reported passing cases, 0 failures. The existing Windows-only no-op skip cases remain unexecuted POSIX coverage, not live passes.
• Live parent: codex-20260926232642-f56c99; continuation: codex-20260926232849-8610ae; same session 01a0e00a-de27-71a2-900c-8f02ea434698.
• Continuation: successful (harness_telemetry), exit 0, permission mode codex: -s read-only, no measured Git file delta. This proves CLI continuation compatibility; the driver's Git measurement is not proof against every possible write.
• The delegate-task.driver.v1 and delegate-task.result.v2 schemas are unchanged. Plugin version 1.1.1 is packaged as a standalone tagged release. Validation is local plus the live Codex continuation above; no GitHub CI pass or consumer adoption is claimed.