Documentation
¶
Overview ¶
Package store persists the store-owned sidecar files of a completed eval run into the canonical run directory <root>/.agents/eval/runs/<run-id>/. It is one participant in an INCREMENTALLY ASSEMBLED directory: the sandbox creates the run dir, the score stage writes iteration-log/{iter-1.yaml,iter-1.score.yaml} into it, and this package writes the two files it owns (taskspec.yaml, eval-run.yaml). The store therefore does NOT own the whole directory — it never publishes it atomically as a unit, never refuses an existing dir, and never deletes it. Crash-safety is per-file: each file it writes appears atomically (temp-then-rename via internal/fsops) or not at all.
Index ¶
Constants ¶
This section is empty.
Variables ¶
var ( // ErrEmptyRunID is returned when the run carries no run ID. ErrEmptyRunID = errors.New("store: run id is required") // ErrUnsafeRunID is returned when the run ID contains a path separator, // a ".." segment, or otherwise does not name a single path element — any // of which could escape the <root>/.agents/eval/runs/ directory. ErrUnsafeRunID = errors.New("store: run id is not a safe path element") // ErrNilSpec is returned when the run carries no task spec. ErrNilSpec = errors.New("store: task spec is nil") // ErrEmptyRoot is returned when the root dir is empty. ErrEmptyRoot = errors.New("store: root dir is required") // ErrEmptyRecordPath is returned when the run score has no iteration-record // path. The score stage writes it (scoringbridge.Result.RecordPath); its // absence means the scoring stage did not complete. ErrEmptyRecordPath = errors.New("store: score record path is required") // ErrEmptyScorePath is returned when the run score has no score-sidecar // path (scoringbridge.Result.ScorePath) — the same "scoring did not // complete" signal for the score artifact. ErrEmptyScorePath = errors.New("store: score sidecar path is required") )
Typed errors callers can match with errors.Is.
Functions ¶
Types ¶
type PersistedEvalRun ¶
type PersistedEvalRun struct {
RunID string `yaml:"run_id"`
BaseCommit string `yaml:"base_commit"`
TaskID string `yaml:"task_id"`
Language string `yaml:"language"`
Difficulty string `yaml:"difficulty"`
Agent agentIdentity `yaml:"agent"`
Verify verifyOutcome `yaml:"verify"`
Score scoreOutcome `yaml:"score"`
}
PersistedEvalRun is the on-disk YAML shape of eval-run.yaml. It aggregates run identity, the agent-runner identity + reproducibility metadata (spec R9 / R10, plan D6), the verification result, and the score summary so the R2 dashboard can render a run without reading all four sidecar files.
type Result ¶
type Result struct {
// RunDir is the run's sidecar directory (<root>/.agents/eval/runs/<run-id>).
RunDir string
// TaskspecPath is the persisted task spec (<RunDir>/taskspec.yaml).
TaskspecPath string
// EvalRunPath is the persisted run aggregate (<RunDir>/eval-run.yaml).
EvalRunPath string
// IterLogDir is the iteration-log subdirectory (<RunDir>/iteration-log).
IterLogDir string
// RecordPath is the iteration record (<IterLogDir>/iter-1.yaml).
RecordPath string
// ScorePath is the score sidecar (<IterLogDir>/iter-1.score.yaml).
ScorePath string
}
Result reports the persisted artifact paths produced by WriteEvalRun. Every path names the canonical location under the run dir.
func WriteEvalRun ¶
WriteEvalRun persists the store-owned files of a completed eval run into the (incrementally assembled) canonical run directory:
<root>/.agents/eval/runs/<run.RunID>/ taskspec.yaml — the eval.TaskSpec (store-owned) eval-run.yaml — run aggregate + R9/R10 (store-owned) iteration-log/iter-1.yaml — R1-shaped iteration record (score stage) iteration-log/iter-1.score.yaml — score sidecar (score stage)
It ensures the run dir (and iteration-log) exist, then writes taskspec.yaml and eval-run.yaml with per-file atomic writes (temp-then-rename via fsops.WriteFileAtomic) — the crash-safety unit is the individual file, so a mid-write failure of one leaves no torn file and never touches the others.
The iteration-log artifacts belong to the score stage. WriteEvalRun ADAPTS to both wirings: when run.Score.RecordPath / ScorePath already resolve to their canonical location inside this run dir (the merged harness, where the score stage wrote them in place), it adopts them — validating they exist, never copying a file onto itself. When they resolve to a scratch location outside the run dir, it copies them in atomically. It never deletes the run dir and never refuses an existing one.
Source Files