benchmark-runner

name	benchmark-runner
description	Auto-discover all skills with evals in RConsortium/pharma-skills, benchmark each with vs. without skill using matched isolated sessions, and post scored results to the linked GitHub issue. Use whenever someone says "run benchmarks", "compare skill performance", "eval the skills", or wants to measure whether a skill improves output quality.

Skill Benchmark Runner

Benchmark every evaluation case in the _automation/evals/ directory of the RConsortium/pharma-skills repository. Each routine invocation is one of two short phases (~20 min each). The routine inspects GitHub issue comments on startup to decide which phase to execute — no configuration needed, no commits required, no repo write access required from the human user.

Repository: RConsortium/pharma-skills (https://github.com/RConsortium/pharma-skills)

---

Model Selection — Series Label and Sub-Agent Model

{CURRENT_MODEL_NAME} throughout this skill is the benchmark series label, fixed to the literal string Claude Routine. Use it verbatim everywhere it appears: get_next_eval.py --model, record_run_result.py --model, the marker JSON ("model"), and the report metadata. The scripts normalise this label (lowercase, strip spaces/punctuation) for deduplication, so all runs group into one consistent series.

Sub-agent model:

• Agent A and Agent B are launched without an explicit --model flag (claude -p ...), so each sub-agent inherits this session's default model — the same model orchestrating the routine. This keeps the measurement clean (bare model ± skill, identical model on both sides) without ever writing a concrete model ID into a repository artifact. Never hardcode a concrete model ID anywhere in this skill or its outputs.
• The prompt files constructed in Step 2 / Step 6 and the CLAUDE_CODE_MAX_OUTPUT_TOKENS setting are the only information passed from the parent session to a sub-agent. Do not forward any other parent context: no conversation history, no additional environment variables, no eval assertions, no scoring prompt, no blinded map. This prevents the orchestrator's context from leaking into either candidate.

---

Routine Setup (one-time)

Create a single routine at claude.ai/code/routines:

Field	Value
Prompt	Read _automation/benchmark-runner/SKILL.md and execute.
Repository	RConsortium/pharma-skills
Schedule	0 1,6 * * * (1 AM and 6 AM UTC — 5 h gap, matches rolling usage window)

That is all. The skill determines its own phase on every invocation.

---

GitHub Access — Use Whichever Method Works

Throughout this skill you will read issue comments, post issue comments, and create release assets. Use whatever method is available in your environment — pick the one that works without prompting:

Method	Best when	Notes
mcp__github__* MCP tools	Running inside Claude Code with the GitHub MCP server	No token required; preferred when available
gh CLI (gh issue view, gh release upload, etc.)	Running locally with gh authenticated	Concise, supports all operations
REST API via curl	Anywhere with GH_TOKEN / GITHUB_TOKEN set	Universal fallback; use for release-asset upload (no MCP equivalent)
Provider-specific GitHub tools (Codex, Gemini, etc.)	Running under another agent CLI	Use whatever the host provides

Reason about which method to use; do not enforce a rigid order. If one fails, try another. Always confirm the operation succeeded (e.g., the comment URL came back, the asset was uploaded) before continuing.

For release-asset upload there is currently no MCP tool — use gh release upload or curl POST to the upload URL.

---

Phase Detection — Run Before Any Other Step

Scan all benchmark eval issues to find any that are waiting for Phase 2 (Agent B + scoring) for the model you are running:

1. List all eval files: ls _automation/evals/*.json — extract each id field (e.g. github-issue-27 → issue #27).

2. For each issue number, fetch comments using whichever GitHub access method is available (see above). Scan each comment body for a <!-- BENCHMARK_PARTIAL: marker.

3. Filter and evaluate each BENCHMARK_PARTIAL marker found:

   • Skip if state.model does not match {CURRENT_MODEL_NAME}. This is critical: a partial run by another user on a different model belongs to that user. Only pick up partials matching your current model.
   • Skip if a later comment on the same issue contains a matching <!-- BENCHMARK_COMPLETE: {"eval_id":"{same}","model":"{same}" — Phase 2 already finished for that combination.
   • Otherwise → Phase 2 candidate. Extract the JSON from the marker (see format below) and note the partial comment id.

4. Decision:

   • One or more Phase 2 candidates found → pick the oldest (earliest created_at) → enter Phase 2 with that state.
   • No candidates for your model → enter Phase 1.

BENCHMARK_PARTIAL marker format (hidden HTML comment embedded in the issue comment body):

<!-- BENCHMARK_PARTIAL: {"eval_id":"github-issue-27","model":"Claude Routine","skill_sha":"b5ede6a...","issue_number":27,"blinded_map":{"candidate_1":"output_B","candidate_2":"output_A"},"agent_a_asset_url":"https://github.com/RConsortium/pharma-skills/releases/download/benchmark-results/benchmark_agent_a_github-issue-27.zip","run_date":"2026-05-03T06:00Z","tokens_a":199382,"partial_comment_id":4367060533} -->

---

Phase 1 — Agent A Run (With Skill)

Runs when no Phase 2 candidate is found. Executes Agent A, archives its output, and posts a partial comment that holds state for Phase 2.

Step 0 — R Environment Pre-flight

Always run first. Idempotent — safe to re-run.

bash _automation/benchmark-runner/scripts/setup_r_env.sh

Exits non-zero on failure — stop and report the error. Do not proceed.

R packages installed: jsonlite, digest, gsDesign, gsDesign2, lrstat, graphicalMCP, eventPred, ggplot2

Step 1 — Discover Next Eval

python3 _automation/benchmark-runner/scripts/get_next_eval.py --model {CURRENT_MODEL_NAME}

• STATUS: UP_TO_DATE → all evals complete for this model+SHA. Exit.
• JSON output → parse to a temp file; extract _skill_name, _skill_sha, _skill_content, _bundled_resources, _prompt_a, _blinded_scoring_map, and the issue number from id.

Optional flags:

--runner-id {YOUR_NAME}           # stable per-person ordering
--priority-issue github-issue-{N} # force a specific eval

Step 2 — Run Agent A (With Skill)

Create the working directory:

mkdir -p /tmp/benchmark_{id}/agent_A/output_A

Stage bundled resource files to disk (progressive disclosure — files read on demand, not embedded in the prompt):

import os, json
agent_a_dir = "/tmp/benchmark_{id}/agent_A"
for rel_path, content in eval_case["_bundled_resources"].items():
    if rel_path == "SKILL.md":
        continue
    dest = os.path.join(agent_a_dir, rel_path)
    os.makedirs(os.path.dirname(dest), exist_ok=True)
    with open(dest, "w", encoding="utf-8") as f:
        f.write(content)

Write prompt_A.txt — _skill_content (SKILL.md) followed by _prompt_a only. No bundled resource content in the prompt:

prompt_a = eval_case["_skill_content"] + "\n\n" + eval_case["_prompt_a"]
with open(os.path.join(agent_a_dir, "prompt_A.txt"), "w", encoding="utf-8") as f:
    f.write(prompt_a)

Launch Agent A:

export CLAUDE_CODE_MAX_OUTPUT_TOKENS=64000
cd /tmp/benchmark_{id}/agent_A && \
  cat prompt_A.txt | claude -p \
  --allowedTools "Bash,Read,Write,Edit,Glob" \
  --output-format json > agent_A_run.json 2>&1

Note: export is required — a prefix (VAR=val cat ... | claude) only sets the variable for cat, not for the claude process receiving the pipe.

--output-format json emits a single JSON object when the agent finishes — resilient to long-running agents and session timeouts.

When Agent A returns, extract token count:

import json
d = json.load(open("/tmp/benchmark_{id}/agent_A/agent_A_run.json"))
u = d.get("usage", {})
tokens_a = u.get("input_tokens", 0) + u.get("cache_creation_input_tokens", 0) + u.get("output_tokens", 0)
is_error_a = d.get("is_error", False)

Record in runs.json:

python3 _automation/benchmark-runner/scripts/record_run_result.py \
  --eval-id {id} --model {CURRENT_MODEL_NAME} \
  --status partial_a --tokens-a {tokens_a}

Step 3 — Archive Agent A Output

Create the zip:

cd /tmp/benchmark_{id} && zip -r benchmark_agent_a_{eval_id}.zip \
  agent_A/output_A/ agent_A/agent_A_run.json

Upload to the benchmark-results GitHub release as a named asset. The release must already exist (create it once if needed). Use whichever method works in your environment — examples below; pick what works:

• gh CLI (simplest if available):

  gh release view benchmark-results --repo RConsortium/pharma-skills \
    || gh release create benchmark-results --repo RConsortium/pharma-skills \
       --prerelease --title "Automated Benchmark Results" --notes "Rolling release."
  gh release upload benchmark-results /tmp/benchmark_{id}/benchmark_agent_a_{eval_id}.zip \
    --repo RConsortium/pharma-skills --clobber
  

• REST API via curl (when only GH_TOKEN is available):

  # Get-or-create release, then POST to its upload_url with the zip as data-binary.
  # See https://docs.github.com/en/rest/releases for the exact endpoints.
  

• MCP does not currently expose a release-asset upload tool — use gh or curl for the upload step. Comment posting and reading can still use MCP.

Construct the asset download URL (used in the partial comment state):

https://github.com/RConsortium/pharma-skills/releases/download/benchmark-results/benchmark_agent_a_{eval_id}.zip

If no upload method works (no gh, no token), skip the upload and set agent_a_asset_url: null in the partial state. Phase 2 will detect the null URL and re-run Agent A for that eval — wasteful but correct.

Step 4 — Post Partial Comment

Write the partial comment body to /tmp/partial_comment_{eval_id}.md:

## Automated Benchmark Results — `{_skill_name}` 🟡 In Progress

### Run Metadata

| Field | Value |
|---|---|
| **Eval ID** | `{id}` |
| **Run date** | {YYYY-MM-DD HH:MM UTC} |
| **Model** | `{CURRENT_MODEL_NAME}` |
| **Skill version** | `{_skill_sha[:7]}` |
| **Phase** | 1 of 2 complete — Agent A (with skill) finished |

Agent A has completed. Agent B (without skill) will run in the next scheduled window (~5 h).
Results will be updated here automatically.

<!-- BENCHMARK_PARTIAL: {"eval_id":"{id}","model":"{CURRENT_MODEL_NAME}","skill_sha":"{_skill_sha}","issue_number":{N},"blinded_map":{_blinded_scoring_map},"agent_a_asset_url":"{asset_url}","run_date":"{ISO8601}","tokens_a":{tokens_a}} -->

Post it using whichever GitHub access method is available (see "GitHub Access" above). The partial comment id returned by the API is not needed for Phase 2 (Phase 2 discovers it by scanning), but log it for debugging.

Phase 1 is complete. Print this summary to the user before exiting:

✓ Phase 1 complete — Agent A finished for {eval_id} ({model})
  • Output archived: {asset_url}
  • Partial comment: {comment_url}
  • Tokens used: {tokens_a:,}

NEXT STEP — Phase 2 (Agent B + scoring):
  • If running as a scheduled routine: nothing to do. The next scheduled
    invocation (≥5 h from now, after the rolling usage window resets) will
    detect this partial state automatically and run Phase 2.
  • If running manually: re-invoke this skill any time. It will
    detect the BENCHMARK_PARTIAL marker on issue #{N} and run Phase 2 to
    completion.

Then exit cleanly.

---

Phase 2 — Agent B Run + Scoring

Runs when a BENCHMARK_PARTIAL state is found in a GitHub issue comment. Loads Agent A's output, runs Agent B, scores both, posts the full result.

Step 5 — Load Partial State

Parse the BENCHMARK_PARTIAL JSON from the comment body found during Phase Detection:

import re, json
marker_re = re.compile(r'<!-- BENCHMARK_PARTIAL: ({.*?}) -->', re.DOTALL)
m = marker_re.search(comment_body)
state = json.loads(m.group(1))
# state keys: eval_id, model, skill_sha, issue_number, blinded_map,
#             agent_a_asset_url, run_date, tokens_a

Also reload the full eval case (for assertions, scoring prompt, prompt_b):

# Send stdout (the eval-case JSON) to the file and stderr (warnings such as
# the >100 KB bundle notice) to a separate log. Do NOT use `2>&1` here — it
# would merge warning lines into the JSON file and break the json.load below.
python3 _automation/benchmark-runner/scripts/get_next_eval.py \
  --model {state["model"]} \
  --priority-issue {state["eval_id"]} \
  > /tmp/eval_case_{id}.json 2>/tmp/eval_case_{id}.log

Restore Agent A's output. If agent_a_asset_url is set, download and unzip it. Use whichever method works:

mkdir -p /tmp/benchmark_{id}/agent_A/output_A

# Option A — gh CLI:
gh release download benchmark-results --repo RConsortium/pharma-skills \
  --pattern "benchmark_agent_a_{eval_id}.zip" --dir /tmp/benchmark_{id}/

# Option B — curl (release assets are public for public repos; token only needed for private):
curl -L "{agent_a_asset_url}" -o /tmp/benchmark_{id}/benchmark_agent_a_{eval_id}.zip

# Then unzip:
cd /tmp/benchmark_{id} && unzip -q benchmark_agent_a_{eval_id}.zip

If agent_a_asset_url is null (Phase 1 could not upload), re-run Agent A from scratch using the same procedure as Phase 1 Step 2 before continuing.

Step 6 — Run Agent B (Without Skill)

mkdir -p /tmp/benchmark_{id}/agent_B/output_B

Write prompt_B.txt — contains only _prompt_b. No skill content, no resource files:

with open("/tmp/benchmark_{id}/agent_B/prompt_B.txt", "w") as f:
    f.write(eval_case["_prompt_b"])

Launch Agent B:

export CLAUDE_CODE_MAX_OUTPUT_TOKENS=64000
cd /tmp/benchmark_{id}/agent_B && \
  cat prompt_B.txt | claude -p \
  --allowedTools "Bash,Read,Write,Edit,Glob" \
  --output-format json > agent_B_run.json 2>&1

Extract token count and record:

d = json.load(open("/tmp/benchmark_{id}/agent_B/agent_B_run.json"))
u = d.get("usage", {})
tokens_b = u.get("input_tokens", 0) + u.get("cache_creation_input_tokens", 0) + u.get("output_tokens", 0)
is_error_b = d.get("is_error", False)

python3 _automation/benchmark-runner/scripts/record_run_result.py \
  --eval-id {state["eval_id"]} --model {state["model"]} \
  --status completed --tokens-b {tokens_b}

Step 7 — Score Blinded Outputs

Copy outputs per state["blinded_map"] to /tmp/benchmark_{id}/scoring/:

mkdir -p /tmp/benchmark_{id}/scoring/candidate_1 /tmp/benchmark_{id}/scoring/candidate_2
# blinded_map: {"candidate_1": "output_B", "candidate_2": "output_A"} (or reversed)
cp -r /tmp/benchmark_{id}/agent_{X}/output_{X}/. /tmp/benchmark_{id}/scoring/candidate_1/
cp -r /tmp/benchmark_{id}/agent_{Y}/output_{Y}/. /tmp/benchmark_{id}/scoring/candidate_2/

For each candidate, evaluate every assertion in the eval case:

• Pass — clearly met
• Partial — partially met
• Fail — not met

Score = (passes + 0.5 × partials) / total_assertions

Then unblind using state["blinded_map"] to map candidate scores back to "With Skill" and "Without Skill".

Step 8 — Format Full Report

Write /tmp/benchmark_comment_{skill}_{eval_id}.md:

## Automated Benchmark Results — `{_skill_name}`

### Run Metadata

| Field | Value |
|---|---|
| **Eval ID** | `{id}` |
| **Run date** | {YYYY-MM-DD HH:MM UTC} |
| **Model** | `{model}` |
| **Skill version** | `{skill_sha[:7]}` |
| **Triggered by** | Scheduled |

### Scorecard

| Metric | With Skill | Without Skill |
|---|---|---|
| **Score** | {score_A} ({pct_A}%) | {score_B} ({pct_B}%) |
| **Assertions** | {pass_A} Pass · {partial_A} Partial · {fail_A} Fail | {pass_B} Pass · {partial_B} Partial · {fail_B} Fail |
| **Skills loaded** | 1 | 0 |
| **Execution time** | {time_A} min | {time_B} min |
| **Token usage** | {tokens_a} | {tokens_b} |
| **{Key Metric 1}** | {value_A1} | {value_B1} |
| **{Key Metric 2}** | {value_A2} | {value_B2} |

### Key Observations

- {2-4 bullet points comparing both agents}

### Verdict

{1-2 sentence overall verdict}

---

## Technical Details & Artifacts

<details>
<summary>View Assertion Breakdown, Code Artifacts, and Logs</summary>

### Assertion Breakdown

| Assertion | With Skill | Without Skill |
|---|---|---|
| {assertion_text_1} | {Pass/Partial/Fail} | {Pass/Partial/Fail} |

### Debugging Information


  {numAerror value, or "None"}


  {numBerror value, or "None"}



 []()


{Key output files — .R, .json, text summaries}


{Key output files}



---
<!-- BENCHMARKid":"{id}","model":"{model}","skillsha}"} -->

Note the <!-- BENCHMARK_COMPLETE: --> marker at the bottom — this tells future Phase Detection scans that Phase 2 is done for this eval+model+sha.

Step 9 — Post Full Results Comment

Post as a new comment using whichever GitHub access method is available (see "GitHub Access" above). The new comment carries the BENCHMARK_COMPLETE marker; the partial comment can stay in place — future Phase Detection scans will see the COMPLETE marker on a later comment and skip the partial.

If you prefer to also edit the partial comment to mark it as superseded (cleaner timeline), use whatever update method works in your environment (gh api PATCH, REST PATCH with GH_TOKEN, etc.). Optional — not required for correctness.

Phase 2 is complete. Print this summary to the user before exiting:

✓ Phase 2 complete — full benchmark posted for {eval_id} ({model})
  • Score: With Skill {pct_A}% · Without Skill {pct_B}%
  • Comment: {comment_url}
  • Tokens — A: {tokens_a:,} · B: {tokens_b:,}

Then exit cleanly.

---

Execution Flow

EVERY ROUTINE INVOCATION:

  Phase Detection
    │
    ├─ BENCHMARK_PARTIAL found (no BENCHMARK_COMPLETE for same eval+model) ──► Phase 2
    │     Step 5: load state from comment + restore Agent A output
    │     Step 6: run Agent B (without skill)
    │     Step 7: score blinded
    │     Step 8: format full report
    │     Step 9: post full results comment (with BENCHMARK_COMPLETE marker)
    │     EXIT
    │
    └─ No partial found ──► Phase 1
          Step 0: R pre-flight
          Step 1: get_next_eval.py → if UP_TO_DATE, EXIT
          Step 2: run Agent A (with skill)
          Step 3: archive + upload Agent A output
          Step 4: post partial comment (with BENCHMARK_PARTIAL marker + state JSON)
          EXIT

---

Notes on Model Name

{CURRENT_MODEL_NAME} is the fixed series label Claude Routine — see "Model Selection — Series Label and Sub-Agent Model" at the top of this skill. Sub-agents inherit the host's default model rather than being pinned with --model, so no concrete model ID is recorded in issue comments or release assets. The dedup logic normalises the label, so every run posted under Claude Routine groups into one series.

Notes on Distributed Selection

When several people run the same model, set distinct --runner-id values. The dispatcher hashes runner-id + model + UTC minute + eval-id + skill-SHA to spread different runners across different pending evals. Runners starting in the same minute may collide; the GitHub issue-comment deduplication (checking for BENCHMARK_COMPLETE markers) prevents redundant Phase 1 runs.

Notes on Rate Limits

If Agent A or Agent B hits a usage rate limit mid-run (is_error: true, result contains "You've hit your limit"):

• Agent A rate-limited in Phase 1: record status: error_a_rate_limited in runs.json, do NOT post a partial comment, exit. The next Phase 1 invocation will retry.
• Agent B rate-limited in Phase 2: record status: error_b_rate_limited, do NOT post a full results comment. Leave the BENCHMARK_PARTIAL comment in place so the next Phase 2 invocation retries Agent B. Include a note in the partial comment body edit if possible.

Success Criteria

• One phase executed per invocation (~20 min each)
• State persists entirely in GitHub issue comments — no commits, no repo write access needed from the human user
• Blinded scoring: _blinded_scoring_map is never visible to the scorer
• Deduplication: BENCHMARK_COMPLETE marker prevents re-running finished evals
• Results posted on the correct GitHub issue with full assertion breakdown