intelligent-delegate

Plugin intelligent-delegate from marketplace intelligent-delegate in joshuawowk/intelligent_delegate. Commands read from the repo's manifest — 8 days ago.

Skills (2)

• audit

  Audits where this session's spend and time went — the orchestrator's own context against each subagent's tier, model, effort and cost — and flags the patterns behind an expensive session: an orchestrator-heavy split, a large re-read context, background-completion wake-ups, cache rebuilds, untiered o

• delegate

  Routes work to the right execution shape — inline, one subagent, or a Workflow fan-out — and picks the model tier and reasoning effort for each, for speed, result quality and cost together, including what the orchestrator's own context costs. Use whenever delegation, parallelism, subagents, backgrou

Subagents (8)

• builder

  ESCALATION implementation tier — inherits the session model at high effort WITH Edit and Write, so in an Opus session this IS an Opus agent: measured at 2.4x the time and 3.4x the cost of `drafter` on a spec-driven implementation, for the same result. Reach for `drafter` (sonnet/high, writes) FIRST

• drafter

  Default generation tier (Sonnet, high effort, writes). Use for producing a substantial artifact from source material you supply — writing documentation from a verified evidence file, drafting a skill or runbook, implementing a well-scoped feature whose shape the brief settles, applying a pattern acr

• inspector

  Default inspection tier (Sonnet, medium effort, read-only). Use for ordinary review, verification and audit work — checking a generated artifact against its source, verifying a finding, per-item correctness passes in a fan-out, hunting edge cases in a bounded diff — and for a read whose answer you w

• judge

  Adjudication tier (Opus, xhigh effort). Use only for contested calls where being wrong is expensive and cheaper tiers have already disagreed or already failed — refuting a claimed finding, choosing between competing designs, deciding whether a risky change is safe, settling a split verdict from a ve

• mechanical

  Cheapest writing tier (Haiku, low effort). Use for deterministic, fully specified work whose result is verifiable without judgment — mechanical refactors, renames applied across known files, format/lint fixes, boilerplate generation, converting records by rule and example. Cheap, not fast: on a 40-r

• reviewer

  ESCALATION inspection tier — inherits the session model at high effort, so in an Opus session this IS an Opus agent: measured at 3x the time and 4x the cost of `inspector` on a bounded review, for the same result on a saturated task. Reach for `inspector` (sonnet/medium, read-only) FIRST for ordinar

• scout

  Cheapest reading tier (Haiku, low effort, read-only). Use to pull named facts out of a large input — counting or locating entries in a long log, extracting fields from a changelog, finding which of many files mention something — when the answer is one you can check, or a wrong one is cheap. Cheap, n

• worker

  Default workhorse tier (Sonnet, medium effort, writes). Use for ordinary multi-step implementation and research that needs real reading comprehension but has a knowable right answer — implementing a well-scoped feature, tracing how a subsystem works, writing tests against a clear spec, migrating a k

Hooks (1)

• PreToolUse

  python3 "${CLAUDE_PLUGIN_ROOT}/hooks/tier_check.py"

Similar plugins