Reward Schema Reference

Every task emitted by Repo2RLEnv writes its reward to /logs/verifier/ inside the container after the agent's patch is applied and the verifier runs. Harbor reads reward.txt as the primary training signal. Native pipelines with a graded verifier also write reward-details.json, the full breakdown for analysis, filtering, and debugging.

Files written per task

File	Always present?	Content
`/logs/verifier/reward.txt`	✅	Single float on one line, e.g. `0.842731`
`/logs/verifier/reward-details.json`	Pipeline-dependent (see below)	Full breakdown as JSON (`pr_diff`, `pr_runtime`, `commit_runtime`, `cve_patches`)
`/logs/verifier/result.json`	Research recipes and Tasksmith	The grader's outcome, e.g. `{"passed": true, "returncode": 0, "statuses": {…}}`; SCALER writes `{"score": …, "accuracy": …}`. CLI-Gym writes `reason.txt` instead when protected source was changed.

`pr_diff`

Reward kind: diff_similarity

reward.txt: weighted combination of 6 components, clamped to [0, 1].

reward-details.json:

{
  "final_reward": 0.74,
  "components": {
    "format_valid":   1.0,
    "size_sanity":    0.85,
    "file_targeting": 0.67,
    "region_overlap": 0.60,
    "similarity":     0.55,
    "llm_judge":      0.80
  },
  "weights": {
    "format_valid":   0.00,
    "size_sanity":    0.08,
    "file_targeting": 0.12,
    "region_overlap": 0.20,
    "similarity":     0.10,
    "llm_judge":      0.50
  },
  "judge_model":  "claude-haiku-4-5-20251001",
  "judge_endpoint": null,
  "judge_status": "ok"
}

Field reference:

Field	Type	Description
`final_reward`	float [0, 1]	Final score, the same value as `reward.txt`. Hard-capped to ≤ 0.40 when `size_sanity < 0.10` (catastrophic-size guard); the breakdown has no separate flag for the cap, so check `components.size_sanity`.
`components.format_valid`	0 or 1	Predicted output parses as a valid unified diff. Weight 0.00: kept as a guard, not a scoring factor.
`components.size_sanity`	[0, 1]	`min(oracle_loc, pred_loc) / max(oracle_loc, pred_loc)`. Catches severe over- or under-generation.
`components.file_targeting`	[0, 1]	F1 over the sets of changed files (not Jaccard; F1 gives partial credit for TP).
`components.region_overlap`	[0, 1]	Predicted hunks overlap oracle hunks within a 5-line slack.
`components.similarity`	[0, 1]	`difflib.SequenceMatcher` ratio over `+`/`-` lines only (no credit for context lines).
`components.llm_judge`	[0, 1] or `null`	LLM semantic judge ("does this address the issue?"). `null` when disabled or the API key is absent, in which case its weight is redistributed to the remaining components.
`weights`	object	Effective per-component weights (overridable via `R2E_W_*` env vars).
`judge_model`	string or `null`	Model used for `llm_judge` as the serving API names it (e.g. `claude-haiku-4-5-20251001`, `Qwen/Qwen3.5-4B`), or `null` if the judge did not score.
`judge_endpoint`	string or `null`	The OpenAI-compatible server the judge was routed to (`R2E_JUDGE_ENDPOINT`), or `null` on the default Anthropic route.
`judge_status`	string	`"ok"` \| `"no_api_key"` \| `"no_judge_model"` \| `"empty_predicted"` \| `"network"` \| `"parse"` \| `"missing_score"`

Weight override env vars (set inside the verifier container via --ve): R2E_W_FORMAT, R2E_W_SIZE, R2E_W_FILE, R2E_W_REGION, R2E_W_SIM, R2E_W_JUDGE

Judge routing env vars (also via --ve): R2E_JUDGE_MODEL, R2E_JUDGE_ENDPOINT, R2E_JUDGE_API_KEY. See ENV.md.

`pr_runtime` · `commit_runtime` · `cve_patches`

Reward kind: test_execution (primary) + diff_similarity (fallback)

reward.txt: f2p_rate × p2p_rate, rounded to 6 decimal places.

reward-details.json (normal path: test output parsed successfully):

{
  "reward":                  0.833333,
  "resolved":                false,
  "command_resolved":        false,
  "f2p_total":               3,
  "f2p_passed":              2,
  "f2p_rate":                0.666667,
  "p2p_total":               5,
  "p2p_passed":              5,
  "p2p_rate":                1.0,
  "regressions":             [],
  "untracked_failed_count":  1,
  "untracked_failed":        ["tests/test_other.py::test_legacy"],
  "parse_status":            "ok",
  "runner":                  "pytest",
  "tests_parsed":            12,
  "exit_code":               1
}

reward-details.json (fallback path: test output unrecognised, no per-test status):

{
  "reward":            1.0,
  "resolved":          false,
  "command_resolved":  true,
  "parse_status":      "fallback_exitcode",
  "eval_trustworthy":  false,
  "runner":            "",
  "f2p_total":         3,
  "p2p_total":         5,
  "exit_code":         0
}

Field reference:

Field	Type	Description
`reward`	float [0, 1]	Dense training signal: `f2p_rate × p2p_rate`. Gold patch → 1.0.
`resolved`	bool	SWE-bench eval signal. At least one F2P test is declared, all F2P tests pass AND all P2P tests pass. Gold patch → `true`. Use this for benchmark-style scoring.
`command_resolved`	bool	Stricter: `resolved` AND no untracked failures AND `exit_code == 0`. Filters tasks with pre-existing flaky tests outside F2P/P2P.
`f2p_total`	int	Number of declared FAIL_TO_PASS tests.
`f2p_passed`	int	F2P tests that passed after the agent's patch.
`f2p_rate`	float [0, 1]	`f2p_passed / f2p_total`. 0.0 when `f2p_total == 0`.
`p2p_total`	int	Number of declared PASS_TO_PASS tests.
`p2p_passed`	int	P2P tests still passing after the agent's patch.
`p2p_rate`	float [0, 1]	`p2p_passed / p2p_total`. 1.0 when `p2p_total == 0` (no regression guard).
`regressions`	list[str]	P2P tests that broke under the agent's patch.
`untracked_failed_count`	int	Tests that failed but were not in F2P or P2P, often from pre-existing flakiness.
`untracked_failed`	list[str]	Names of untracked failures (capped at 20).
`parse_status`	string	`"ok"`: per-test status parsed successfully. `"fallback_exitcode"`: runner output unrecognised, so `reward` is binary and based on the exit code.
`eval_trustworthy`	bool	Only present in `fallback_exitcode` path. `false` when an F2P oracle exists but couldn't be verified.
`runner`	string	Detected test runner: `"pytest"` \| `"go"` \| `"cargo"` \| `"jest"` \| `""`
`tests_parsed`	int	Total tests detected in log output (only in `parse_status == "ok"`).
`exit_code`	int	Raw exit code of the test command.

Choosing between reward, resolved, and command_resolved:

Training: use reward (dense, graded signal from reward.txt).
Benchmark / leaderboard: use resolved (strict SWE-bench tracked resolution, unaffected by untracked flakiness).
Strict eval (clean command required): use command_resolved.

`code_instruct` · `equivalence_tests`

Reward kind: test_execution

reward.txt: 1.0 (tests pass) or 0.0 (tests fail). Binary.

reward-details.json: not written by these pipelines.

The verifier runs the repo's test suite (or the generated pytest equivalence tests) and maps the exit code directly to reward. No per-component breakdown.

Reading the reward for training:

reward = float(open("/logs/verifier/reward.txt").read().strip())  # 1.0 or 0.0

Per-pipeline summary

Pipeline	`reward.txt`	`reward-details.json`	Training signal	Eval signal
`pr_diff`	weighted [0,1]	✅	`reward`	`components` breakdown
`pr_runtime`	f2p × p2p	✅	`reward`	`resolved` / `command_resolved`
`commit_runtime`	f2p × p2p	✅	`reward`	`resolved` / `command_resolved`
`cve_patches`	f2p × p2p	✅	`reward`	`resolved` / `command_resolved`
`code_instruct`	1.0 / 0.0	❌	`reward`	`reward`
`equivalence_tests`	1.0 / 0.0	❌	`reward`	`reward`

Reward Schema Reference

Files written per taskCopy Anchor Link

pr_diffCopy Anchor Link

pr_runtime · commit_runtime · cve_patchesCopy Anchor Link

code_instruct · equivalence_testsCopy Anchor Link

Per-pipeline summaryCopy Anchor Link

On this page

Files written per task

`pr_diff`

`pr_runtime` · `commit_runtime` · `cve_patches`

`code_instruct` · `equivalence_tests`

Per-pipeline summary