Skip to content

Adonis0123/adonis-skills / skills/agent-dispatch

agent-dispatch

Flagged

Choose the lightest agent setup for a coding, review, drafting or design-research task: do it yourself, fan out to 2-3 workers with independent parts across the Claude Code, Codex, Grok Build, Cursor or other CLIs and accounts on this machine, run a scoped architecture loop, or send one brief to independent researchers; pick a cross-model reviewer, test budget and cleanup; track dispatched work, external waits and user decisions on a Run Board. Use when work may need delegation, or the user asks to dispatch, fan out, parallelize, run several agents or accounts, get another AI's review or independent design opinions, or asks how dispatched work is going. Chinese triggers: 派活、并行、要不要多开几个 agent、分给多个 Claude 账号 / codex / grok、找另一个 AI 审、对比优秀设计、多方意见、进度看板、派出的任务进行到哪了。Not for unrelated read-only questions, trivial edits, research with no delegation, a single review of an existing diff (agentic-review-handoff), or the machine-wide session list (sideby sessions). Inside another named workflow, only size its workers.

FileKindSize
GLOSSARY.mdmarkdown2.1 KB
SKILL.mdskill md27.8 KB
evals/README.mdmarkdown7.6 KB
evals/README.zh-CN.mdmarkdown7.5 KB
evals/bench/compare.pyscript5.3 KB
evals/bench/run_bench.pyscript12.3 KB
evals/bench/tasks.jsonconfig6.6 KB
evals/bench/test_bench.pyscript4.4 KB
evals/build_plugin_eval.pyscript11.3 KB
evals/evals.jsonconfig31.4 KB
evals/results.mdmarkdown15.8 KB
evals/test_eval_harness.pyscript3.5 KB
evals/trigger-eval-set.jsonconfig1.6 KB
fixtures/probe-full.jsonconfig1.6 KB
fixtures/probe-single.jsonconfig690 B
fixtures/run-midrun/B9.jsonlother191 B
fixtures/run-midrun/deploy-log.txtmarkdown42 B
fixtures/run-midrun/run.jsonconfig1.8 KB
fixtures/run-sessions/B9.jsonlother191 B
fixtures/run-sessions/deploy-log.txtmarkdown42 B
fixtures/run-sessions/ended.jsonlother95 B
fixtures/run-sessions/run.jsonconfig2.1 KB
fixtures/run-sessions/sessions.jsonconfig804 B
fixtures/session-index.jsonconfig401 B
fixtures/sideby-stub/sessions.jsonconfig1.5 KB
fixtures/sideby-stub/sidebyexecutablescript4.1 KB
references/accounts.mdmarkdown10.4 KB
references/context-budget.mdmarkdown2.9 KB
references/fallbacks.mdmarkdown4.7 KB
references/launch.mdmarkdown16.1 KB
references/measure.mdmarkdown3.3 KB
references/research-brief.mdmarkdown1.3 KB
references/run-board.mdmarkdown48.7 KB
references/task-spec.mdmarkdown5.0 KB
scripts/board.cssother12.6 KB
scripts/board_interaction.jsscript3.7 KB
scripts/board_markdown.jsscript2.9 KB
scripts/board_render.pyscript17.2 KB
scripts/dispatch_board.pyscript95.3 KB
scripts/isolated_case.pyscript1.6 KB
scripts/probe_agents.pyscript12.9 KB
scripts/sideby_cli.pyscript3.2 KB
scripts/sideby_sessions.pyscript3.3 KB
scripts/test_billing_and_quota.pyscript10.9 KB
scripts/test_board_interaction.jsscript2.1 KB
scripts/test_board_markdown.jsscript2.9 KB
scripts/test_board_reload_state.jsscript2.7 KB
scripts/test_bug_regressions.pyscript4.7 KB
scripts/test_dispatch_board.pyscript68.9 KB
scripts/test_harness_limits.pyscript11.7 KB
scripts/test_lane_states.pyscript7.5 KB
scripts/test_probe_agents.pyscript19.2 KB
scripts/test_review_fixes.pyscript6.7 KB
scripts/test_rounds_and_watch.pyscript7.3 KB

evals/results.md

View on GitHub
# agent-dispatch eval results

Newest run first. Add a section per measured change; keep old sections for comparison. How to run each check: [README.md](README.md).

## 2026-10-26 · skill 1.10.1 · claude-opus-5-5 (judge: sonnet) · Claude Code 2.1.309

Follow-up on the UNVERIFIED items of 2026-10-25.

### A. Cases 10 and 15 with a sideby stand-in

The sandbox has no `sideby`, so both cases could only describe what they would do. `fixtures/sideby-stub/sideby` now answers `sessions`, `send` and a few other commands from fixture data in the real `--json` shapes, and sideby cases are judged on the whole trace (case 10's with-skill answer was right but did not name its source, and the judge saw only the last message).

| Case                                     | before with | before without | after with | after without | Run                                                                           |
| ---------------------------------------- | ----------- | -------------- | ---------- | ------------- | ----------------------------------------------------------------------------- |
| 10-global-session-list-is-not-a-dispatch | 0.56        | 0.00           | 1.00       | 0.89          | run6, 3 per arm, $2.53 (run5 with last-message judging: 0.00 vs 0.56, $11.70) |
| 15-reply-local-worker-question           | 0.58        | 0.17           | 1.00       | 0.92          | run5, 3 per arm                                                               |

In case 15 every with-skill run sent the answer through the stand-in and reported its `delivered` receipt. The small delta on both cases is expected: the stand-in makes the right command easy to find for the baseline too.

### Trigger eval (`build_plugin_eval.py --trigger`, `--ablation none`)

| Description                                             | should load (6 prompts × 3)                                                  | should not load (6 × 3) | Cost                                                      |
| ------------------------------------------------------- | ---------------------------------------------------------------------------- | ----------------------- | --------------------------------------------------------- |
| 1.10.0                                                  | 16/18: prompt 06 "这个需求要不要多开几个 agent?还是你自己改就行" loaded 1/3 | 18/18                   | $12.32 (ran as two arms by mistake; with-skill arm shown) |
| 1.10.1: Chinese trigger "多开" → "要不要多开几个 agent" | 18/18 (prompt 06 alone: 6/6, $1.19)                                          | 18/18                   | $7.54                                                     |

Prompt 11 (Orca sidebar icons) hits the 15-turn cap in every with-skill run while searching for Orca, but never loads the skill.

Spend for this follow-up: $35.28.

## 2026-10-25 · skill 1.10.0 · claude-opus-5-5 (judge: sonnet) · Claude Code 2.1.309

### A. With vs without the skill (`claude plugin eval --ablation with-without`)

Scores are the mean share of a case's assertions that held, per arm. Run 1 graded each case with one all-or-nothing grader; runs 2-4 grade every assertion separately, so run 1 is not directly comparable with the later runs.

| Run  | Cases         | Runs per arm | Grading                             | Cost   | Notes                                                                                                                    |
| ---- | ------------- | ------------ | ----------------------------------- | ------ | ------------------------------------------------------------------------------------------------------------------------ |
| run1 | 22            | 3            | one grader per case, all or nothing | $48.91 | with 0.41, meanDelta +0.36; 3 runs ended on API errors                                                                   |
| run2 | 22            | 3            | one grader per assertion            | $58.63 | incomplete: 68 runs ended on API errors (mostly "You've hit your session limit" for cases 12-22) and scored 0 ungraded   |
| run3 | 07, 08, 11-15 | 3            | one grader per assertion            | $90.88 | after the ledger-note and grant fixes below; stopped at `--max-cost-usd 90` (`partial: cost_ceiling`) before cases 16-22 |
| run4 | 16-22         | 2            | one grader per assertion            | $26.48 | completed, no API errors                                                                                                 |

An interrupted rerun of cases 12-13 between run2 and run3 ($5.64) is left out of the table.

Per case (`-` = not run or every run ended on an API error). "Latest" is the newest per-assertion run for that case, which is what the summary line uses.

| Case                                                        | run1 with | run1 without | run2 with | run2 without | run3 with | run3 without | run4 with | run4 without | latest per-assertion run | failing assertions (with, latest) |
| ----------------------------------------------------------- | --------- | ------------ | --------- | ------------ | --------- | ------------ | --------- | ------------ | ------------------------ | --------------------------------- |
| 01-simple-copy-change-stays-direct                          | 0.67      | 0.00         | 1.00      | 0.87         | -         | -            | -         | -            | run2                     |                                   |
| 02-single-product-machine-degrades                          | 0.67      | 0.00         | 0.86      | 0.67         | -         | -            | -         | -            | run2                     | a05×2, a07×1                      |
| 03-cross-package-fanout-plus-article                        | 0.33      | 0.00         | 1.00      | 0.33         | -         | -            | -         | -            | run2                     |                                   |
| 04-composed-pipeline-defers-reviewer-and-pipeline-ownership | 0.00      | 0.00         | 1.00      | 0.75         | -         | -            | -         | -            | run2                     |                                   |
| 05-allow-list-multi-session-headless-permissions            | 0.00      | 0.00         | 0.90      | 0.48         | -         | -            | -         | -            | run2                     | a02×2                             |
| 06-design-research-fans-out-across-products                 | 0.00      | 0.00         | 0.93      | 0.47         | -         | -            | -         | -            | run2                     | a05×1                             |
| 07-progress-question-mid-run-uses-board                     | 0.00      | -            | 0.62      | 0.62         | 0.96      | 0.62         | -         | -            | run3                     | a08×1                             |
| 08-standing-grant-for-deploy-is-scoped                      | 0.00      | 0.00         | 0.25      | 0.58         | 0.92      | 0.83         | -         | -            | run3                     | a04×1                             |
| 09-outward-ask-carries-recipient-and-draft                  | 0.00      | 0.00         | 1.00      | 0.60         | -         | -            | -         | -            | run2                     |                                   |
| 10-global-session-list-is-not-a-dispatch                    | 0.00      | 0.00         | 0.56      | 0.00         | -         | -            | -         | -            | run2                     | a02×1, a03×3                      |
| 11-plain-answer-updates-the-ledger                          | 0.33      | 0.00         | 0.67      | 0.33         | 1.00      | 0.75         | -         | -            | run3                     |                                   |
| 12-board-that-cannot-open-falls-back-to-chat                | 0.33      | 0.00         | -         | -            | 0.92      | 0.75         | -         | -            | run3                     | a04×1                             |
| 13-authorized-local-commits-stay-on-lane-branch             | 1.00      | 0.00         | -         | -            | 1.00      | 0.67         | -         | -            | run3                     |                                   |
| 14-stop-mid-run-stops-what-it-can                           | 0.67      | 0.00         | -         | -            | 1.00      | 0.80         | -         | -            | run3                     |                                   |
| 15-reply-local-worker-question                              | 0.00      | 0.00         | -         | -            | 0.58      | 0.17         | -         | -            | run3                     | a01×2, a02×1, a03×1, a04×1        |
| 16-fix-round-files-old-result-and-keeps-watching            | 0.67      | 0.00         | -         | -            | -         | -            | 1.00      | 0.60         | run4                     |                                   |
| 17-shared-contract-lands-before-fanout                      | 1.00      | 0.33         | -         | -            | -         | -            | 1.00      | 1.00         | run4                     |                                   |
| 18-l1-coordinator-writes-itself                             | 0.67      | 0.00         | -         | -            | -         | -            | 1.00      | 0.50         | run4                     |                                   |
| 19-small-ux-tweak-is-not-research                           | 1.00      | 0.67         | -         | -            | -         | -            | 1.00      | 1.00         | run4                     |                                   |
| 20-worker-summary-carries-decisions                         | 0.67      | 0.00         | -         | -            | -         | -            | 0.83      | 0.00         | run4                     | a01×1                             |
| 21-coordinator-writes-the-article                           | 1.00      | 0.00         | -         | -            | -         | -            | 1.00      | 1.00         | run4                     |                                   |
| 22-close-reports-whether-dispatch-paid-off                  | 0.00      | 0.00         | -         | -            | -         | -            | 1.00      | 0.30         | run4                     |                                   |

**Latest per-assertion run per case, 22 cases: with skill 0.93, without 0.60, delta +0.33.** These mix runs 2-4 (runs 2 and 3 used 3 runs per arm, run 4 used 2).

Remaining with-skill failures:

- 15 (0.58) and 10 (0.56): the sandbox had no `sideby`; resolved on 2026-10-26 (above).
- 02 a05, 05 a02, 06 a05, 07 a08, 08 a04, 12 a04, 20 a01: one or two runs out of 2-3 missed one assertion each; no repeated pattern across cases, not acted on.

### B. Solo vs dispatch outcome benchmark (`bench/run_bench.py`)

3 tasks from this repository's history, hidden tests from the original commit, 3 runs per arm, coordinators rotated across claude:001/002/008. Cost $32.50.

| Task | solo pass | dispatch pass | dispatch tier | solo median s | dispatch median s | solo median $ | dispatch median $ | Flip |
| ---- | --------- | ------------- | ------------- | ------------- | ----------------- | ------------- | ----------------- | ---- |
| T1   | 3/3       | 3/3           | L1            | 100.7         | 173.5             | 1.22          | 1.66              | -    |
| T2   | 3/3       | 3/3           | L1            | 322.3         | 478.2             | 1.83          | 2.95              | -    |
| T3   | 3/3       | 3/3           | L1            | 83.1          | 147.3             | 0.86          | 1.45              | -    |

- solo: 9/9 passed, median 100.7s, median cost $1.22.
- dispatch: 9/9 passed, median 211.2s, median cost $2.11.
- Verdict: dispatch does not pay off on these tasks: not faster and not more correct. Every dispatch run picked L1 (coordinator does the work itself), which is the right tier for tasks this size; the extra time and cost is the skill's own reading and Run Board overhead. All three tasks are solved by both arms, so the set has a ceiling effect: it cannot show a quality gain from dispatching. A future task set needs a task large enough that L2 should win.

### Diagnosis and changes made from these runs

| Finding                                                    | Cause                                                                                                                                                                                                 | Change                                                                                                                                                                                                                               |
| ---------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| Run 1 scored 0/3 on cases that per-assertion review passed | all-or-nothing grader, turn caps                                                                                                                                                                      | `build_plugin_eval.py` emits one LLM grader per assertion and raises limits (timeout 900s, 40 turns)                                                                                                                                 |
| Low scores on 07/08/11 in run 2                            | harness: the agent copies the static ledger fixture, file times reset, after 5 minutes `watch` wakes correctly and the last message reports that false silence; the judge sees middle messages elided | `build_plugin_eval.py` tells the agent `fixtures/run-*` ledgers are static snapshots and a background `watch` only needs to be described                                                                                             |
| 08 missed the grant read-back                              | skill: the grant rule did not say to answer the grant first                                                                                                                                           | `references/run-board.md` rule 7: answer the grant before status news, quote scope and words, deploy/send/publish grants last this Run only and never widen to another environment, ask once as a numbered Ask when scope is unclear |
| Article gap: budgets and gates lived only in prose         | skill                                                                                                                                                                                                 | `dispatch_board.py` enforces max 3 rounds, `--budget-usd`, overlapping `--owns` write Lanes and reviewer-in-writer-session refusal (`scripts/test_harness_limits.py`)                                                                |
| run2 incomplete                                            | account hit its session limit                                                                                                                                                                         | README: API-error runs score 0; rerun them on another account, subsets via a copied `--eval-dir`                                                                                                                                     |

The trigger eval (skill-creator `run_eval.py`) could not run on this machine because the installed agent-dispatch shadows the candidate description; replaced on 2026-10-26 by `build_plugin_eval.py --trigger` (above).

Total spend for this round: A $230.54 (runs 1-4 plus the interrupted rerun), B $32.50, smoke $0.47: $263.51.
finishing-a-development-branchobra/superpowersUse when implementation is complete, all tests pass, and you need to decide how to integrate the workStars: 322kUpdated Claude CodeCodexVerifiedInstall
systematic-debuggingobra/superpowersUse when encountering any bug, test failure, or unexpected behavior, before proposing fixesStars: 322kUpdated Claude CodeCodexVerifiedInstall
test-driven-developmentobra/superpowersUse when implementing any feature or bugfix, before writing implementation codeStars: 322kUpdated Claude CodeCodexVerifiedInstall
skill-creatoranthropics/skillsCreate new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy.Stars: 198kUpdated Claude CodeCodexVerifiedInstall