codeprobe-run
Execute eval task suites against an AI coding agent using codeprobe, spawning isolated sessions and scoring each task with automated tests.
Identity proven by GitHub's verified organisation badge (sourcegraph)
- Agents
- Claude Code
- License
- Apache-2.0 (permissive)
- Pin
- 349658apinned Oct 9, 2026
- Context cost
- about 1,500 tokenswhen the skill loads
- GitHub stars
- 2on sourcegraph/codeprobe
- Scripts
- None found
What it does
“Execute eval tasks against an AI coding agent using the codeprobe CLI. Spawns isolated per-task sessions, scores with automated tests, and emits NDJSON events plus a terminal envelope. Triggers on run eval, run tasks, benchmark agent, evaluate agent, score agent, compare agents. Use this when the agent needs to produce scored results on a mined or scaffolded task suite.”
Generated summary from the skill's own text
Execute eval task suites against an AI coding agent using codeprobe, spawning isolated sessions and scoring each task with automated tests. Outputs structured JSON events and a final envelope with per-config scores, task counts, and cost metrics. Handles errors, retries, and supports offline runs with cost budgets.
When to use it
- Benchmark a coding agent against a mined or scaffolded task suite
- Generate scored results for agent comparison or evaluation
- Run isolated per-task eval sessions with cost tracking
- Automate agent performance measurement in CI or experimentation
When not to
- Agent backends or credentials are not pre-configured and available
- Task suite contains no runnable tasks to evaluate
- Human authorization required for budget changes mid-run
Install
These commands install from the publisher's repository and need no account.
skills CLI
npx skills add sourcegraph/codeprobe --skill codeprobe-runInstalls from the repository's default branch, which may have changed since the commit we checked.
GitHub CLI 2.90 or later
gh skill install sourcegraph/codeprobe codeprobe-run@349658a0c14622dccccc9be038cabca63a1e4a0bInstalls the commit we checked (349658a).
Copy by hand
Download the folder at 349658a and put it in .claude/skills/codeprobe-run for Claude Code and in .agents/skills/codeprobe-run for Codex, Cursor, Gemini CLI, GitHub Copilot and OpenCode.
Claude Code plugin
/plugin install sourcegraph@claude-plugins-officialThe publisher's plugin in Claude's official marketplace. It may hold more than this skill, and we have not pinned it.
Copy for your agent
A prompt for any coding agent. It installs this skill from GitHub at the pinned commit, copies the files as they are, and runs nothing unless you ask.
Prompt
Install the agent skill "codeprobe-run" by Sourcegraph. Source, pinned to commit 349658a0c14622dccccc9be038cabca63a1e4a0b: https://github.com/sourcegraph/codeprobe/tree/349658a0c14622dccccc9be038cabca63a1e4a0b/src/codeprobe/skills_data/codeprobe-run 1. Download that folder at exactly this commit. Do not use the default branch. 2. Copy every file unchanged, keeping the folder structure, into .claude/skills/codeprobe-run/ for Claude Code and .agents/skills/codeprobe-run/ for Codex, Cursor, Gemini CLI, GitHub Copilot and OpenCode. Create the folders if they do not exist. 3. Do not edit the files, and do not run any script in the folder unless I ask you to. 4. When you are done, list the files you copied and where you put them. Listing and checks: https://www.agentnative.dev/skills/sourcegraph/codeprobe-run
SKILL.md at the pin
---
name: codeprobe-run
description: Execute eval tasks against an AI coding agent using the codeprobe CLI. Spawns isolated per-task sessions, scores with automated tests, and emits NDJSON events plus a terminal envelope. Triggers on run eval, run tasks, benchmark agent, evaluate agent, score agent, compare agents. Use this when the agent needs to produce scored results on a mined or scaffolded task suite.
user-invocable: false
---
# codeprobe run (autonomous agent contract)
Run an eval suite against an AI coding backend, scoring each task with the
configured scorer. Emits a dual output surface on non-TTY stdout: NDJSON
per-task events followed by a terminal envelope.
## Environment (pre-loaded)
- !`codeprobe doctor --json`
If doctor reports failures (missing backends, bad credentials, permission
issues), resolve them first. Do not retry `codeprobe run` until doctor
reports `ok: true`.
## Bare invocation
Minimum viable call. `--json` collapses the dual-surface output into a single
envelope — recommended for agents that only want the final verdict:
```bash
codeprobe run <path> --json --agent claude
```
Dry-run to estimate resource requirements without spawning any agent:
```bash
codeprobe run <path> --dry-run --json --agent claude
```
With a cost ceiling and suite filter:
```bash
codeprobe run <path> --json --agent claude --max-cost-usd 5.0 --suite <suite.toml>
```
For airgapped runs, pair with the offline preflight:
```bash
codeprobe run <path> --json --agent claude --offline --offline-expected-run-duration 2h
```
## Output surface (NDJSON + envelope, non-TTY default)
On non-TTY stdout, the default output is a **dual surface** per §7.2 + §5.4:
- Per task, one or more JSON records with `"record_type": "event"` are written
as newline-delimited JSON (NDJSON).
- The final record has `"record_type": "envelope"` and contains the aggregate
run summary (per-config score summaries, totals, cost).
Pass `--json` to collapse this into a single terminal envelope (no per-task
events). Pass `--json-lines` to force NDJSON mode even when `--json` would
otherwise be selected from `CODEPROBE_JSON`.
## JSON fields to parse
Terminal envelope shape (emitted with `--json` or as the last NDJSON
record). Top-level keys are exactly:
```json
{
"record_type": "envelope",
"ok": true,
"command": "run",
"version": "<codeprobe version>",
"schema_version": "1",
"exit_code": 0,
"data": {
"command_schema_version": "1",
"experiment": "...",
"configs": [
{
"label": "...",
"tasks": 10,
"scored_count": 9,
"mean_score": 0.72,
"perfect": 4,
"cost_usd": 3.21,
"quota_error_count": 0,
"infra_failure_count": 1,
"errored_count": 1
}
],
"total_tasks": 10,
"total_cost_usd": 3.21,
"tenant": "<tenant or null>",
"tenant_source": "<flag|env or null>"
},
"error": null,
"warnings": [],
"next_steps": []
}
```
Parse `ok`. On `false`, inspect `error.code`; `error` is a single object
with `code` / `message` / `kind` / `terminal` / `next_try_flag` /
`next_try_value` / `diagnose_cmd` / `message_for_agent` / `detail`, not an
array. On `true`, read `data.configs[]`. `mean_score` and `perfect` cover
the scored population only (`scored_count`); check `quota_error_count`,
`infra_failure_count`, and `errored_count` before treating two configs as
comparable.
Per-task event shape (NDJSON stream):
```json
{
"record_type": "event",
"event": "task_done",
"task_id": "...",
"score": 1.0,
"duration_seconds": 12.3,
"cost_usd": 0.42
}
```
## Error handling
Only the codes below may surface. At runtime the envelope's `error` object
carries the authoritative message and remediation for whichever code fired.
| Code | Kind | Retryable? | Action |
|---|---|---|---|
| NO_EXPERIMENT | diagnostic | yes (with fix) | Select an experiment with `--config` or run `codeprobe experiment new`. |
| AMBIGUOUS_EXPERIMENT | prescriptive | yes (with fix) | Pass `--config <experiment-id>` explicitly. |
| NO_TASKS | diagnostic | no | The suite has no runnable tasks; adjust suite or mine more. |
| NO_SUITE_MATCH | diagnostic | yes (with fix) | Re-issue with a valid `--suite` path. |
| INVALID_PERMISSION_MODE | prescriptive | yes (with fix) | Adjust config; permission-mode must be one of the accepted modes. |
| UNKNOWN_BACKEND | prescriptive | yes (with fix) | Re-issue with `--agent <valid>` — see doctor output for installed backends. |
| NO_BACKENDS_CONFIGURED | diagnostic | no | No backend installed; install one before retrying. |
| OFFLINE_PREFLIGHT_FAILED | diagnostic | no | Credential TTL too short; refresh creds or drop `--offline`. |
| OFFLINE_NET_ATTEMPT | diagnostic | no | Component attempted network IO while offline; inspect and fix config. |
| CANARY_PROOF_FAILED | diagnostic | no | Canary indicates unreliable env; run `codeprobe doctor`. |
| CANARY_PROOF_REQUIRED | prescriptive | yes (with fix) | Enable canary proof and retry. |
| CANARY_MISMATCH | diagnostic | no | Env drift; do not auto-retry — surface to caller. |
| CANARY_GATE_FAILED | diagnostic | no | Canary gate blocked; inspect run artifacts. |
| CAPABILITY_DRIFT | diagnostic | no | Capability snapshot mismatch; run `codeprobe check-infra drift`. |
| TRACE_OVERFLOW_FIRED | prescriptive | yes (with fix) | Re-run with `--trace-overflow truncate` or raise the budget. |
| TRACE_BUDGET_EXCEEDED | prescriptive | yes (with fix) | Re-run with a higher `--trace-overflow` mode or disable trace. |
| BUDGET_EXCEEDED | diagnostic | **TERMINAL — do not retry** | Cost/time budget intentionally terminal. A human must explicitly raise the budget. |
| INTERRUPTED | diagnostic | **TERMINAL — do not retry** | SIGINT/SIGTERM halted the run. Partial results preserved under `.codeprobe/runs/<run-id>`. |
| LLM_UNAVAILABLE | diagnostic | yes (bounded) | Transient provider outage; one retry permitted. |
## Retry policy
- Maximum retry depth per error chain: **2**. After two consecutive errors
sharing the same code, stop and surface the envelope to the caller.
- Terminal errors (BUDGET_EXCEEDED, INTERRUPTED) are **never** retried —
treat `terminal: true` in the envelope as a hard stop signal.
- Do not mutate `--max-cost-usd` or `--timeout` on retry; budget changes are
human-authorized only.
- Between retries, re-inspect the pre-loaded doctor envelope. If doctor now
reports failing checks, stop retrying and escalate.
From sourcegraph/codeprobe at 349658a, shown as plain text under its license: Apache-2.0 (permissive). The license text is in LICENSE.
View SKILL.md on GitHub at 349658aFiles
- SKILL.md6.5 KB · de97c3e33402
1 file, 6.5 KB in total. Each file has a sha256 in the pin's manifest; the first 12 characters are shown.
Skill folder on GitHub at 349658aMore from Sourcegraph
searching-sourcegraph
This skill enables agents to search and navigate codebases using Sourcegraph, with guided workflows for finding implementations, understanding systems…
Claude Code, Codex, Cursor +4 · about 2.2K tokens · pinned 6 days ago
codeprobe-calibrate
This skill runs a validation gate for a curator version by comparing it against a holdout dataset and a reference curator.
Claude Code · about 870 tokens · pinned 6 days ago
codeprobe-check-infra
This skill diagnoses whether mined task directories can run safely by checking for capability drift and verifying credential time-to-live against expected…
Claude Code · about 970 tokens · pinned 6 days ago
codeprobe-interpret
This skill analyzes results from codeprobe benchmark runs by comparing configurations, computing statistical rankings, and generating actionable…
Claude Code · about 920 tokens · pinned 6 days ago
codeprobe-mine
Extracts reusable eval tasks from a repository's merge history using the codeprobe CLI.
Claude Code · about 1.6K tokens · pinned 6 days ago
sourcegraph-nav
Navigates code using Sourcegraph to locate where symbols are defined and find all places they are used.
Contains scripts
Claude Code, Codex, Cursor +3 · about 580 tokens · pinned 6 days ago
Related in AI and agents
agent-builder
Agent Builder orchestrates the creation, modification, testing, and publishing of n8n Agents.
by n8nOfficial publisher
Claude Code, Codex, Cursor +3 · about 2.3K tokens · pinned 6 days ago
intent-recognition
Classifies automation requests by determining which primitive owns the top-level control flow (workflow, agent, needs clarification, or out-of-scope) and…
by n8nOfficial publisher
Claude Code, Codex, Cursor +4 · about 6.9K tokens · pinned 6 days ago
planned-task-runtime
This skill manages follow-up interactions with a background task execution system, handling task completion synthesis, failure replanning, workflow building…
by n8nOfficial publisher
Claude Code, Codex, Cursor +3 · about 2.2K tokens · pinned 6 days ago
skill-creator
This skill guides agents through creating new skills that extend Gemini CLI with specialized knowledge, workflows, and tool integrations.
by GoogleOfficial publisher
Contains scripts
Gemini CLI · about 5K tokens · pinned 6 days ago
-21risk-automation
This skill automates 21risk operations by connecting to the 21risk toolkit through Composio's Rube MCP server.
by ComposioOfficial publisher
Claude Code, Codex, Cursor +3 · about 690 tokens · pinned 6 days ago
-2chat-automation
This skill automates 2chat tasks by integrating with Composio's 2chat toolkit through Rube MCP.
by ComposioOfficial publisher
Agents not labelled yet · about 680 tokens · pinned 6 days ago
Open roles at Sourcegraph
2 open roles on the jobs board
- ML & Agentic Systems Engineer [IC4]
Remote (Europe or North America) (Remote)
- Tech Lead - Code Plane [IC5]
Remote (Remote)