cawdex 1.35.74 → 1.35.76

package/resources/hal/cawdex_agent/requirements.txt

@@ -1,2 +1,2 @@
-# Cawdex HAL adapter has no Python dependencies.
-# It shells out to the installed cawdex CLI.
+# Cawdex HAL adapter has no Python dependencies.
+# It shells out to the installed cawdex CLI.

package/resources/kbench/cawdex_agent/README.md

@@ -1,107 +1,107 @@
-# Cawdex KBench Adapter
-
-This directory is a KBench `custom-adapter` for Cawdex.
-
-```bash
-kbench run \
-  --benchmark swe \
-  --harness custom-adapter \
-  --adapter /path/to/resources/kbench/cawdex_agent \
-  --model-name openrouter/free \
-  --instruction "Fix the bug"
-```
-
-The runner reads the KBench JSON payload from `KBENCH_ADAPTER_INPUT` or stdin,
-invokes `cawdex --prompt "/benchmark ..."` in task mode, and emits one
-`AdapterRunnerOutput` JSON object to stdout.
-Known KBench slugs are mapped to benchmark profiles before dispatch:
-`swe`/`swe-bench`, `tb2`/`terminal-bench`, `terminalworld`/`terminal-world`,
-`swe-chain`,
-`swe-cycle`/`fullcycle`/`swe-judge`, `swe-ci`/`swecibench`, `swe-prbench`/`prbench`/`pr-review`, `tml-bench`/`tabular-ml`/`kaggle-ml`, `pi-bench`/`proactive-assistant`, `ci-repair`/`ci-repair-bench`, `roadmapbench`, `saasbench`,
-`swe-bench-mobile`, `webdevbench`/`swe-webdev-bench`, `appworld`, `browsecomp`/`browsecompplus`, and
-`tau2`/`tau-bench` use specialized prompts; unknown slugs use
-`generic`.
-
-The output includes redacted instruction/stdout/stderr artifact refs, native
-Cawdex trace refs, and redacted git patch/status refs when the task
-worktree is a git repo. If a native `summary.json` exists, compact verifier
-evidence, including parsed counts, compact failure signatures, and final-answer
-verification-claim plus incomplete/blocked completion evidence, usage/cost
-telemetry, cost-efficiency risk, invalid tool-action telemetry, task-contract checklist completion/no-edit/test-edit signals,
-task-alignment risk signals, spec-compliance risk signals, reward-hack risk signals, long-horizon coverage risk signals, Pi-Bench proactivity ledger signals, incomplete/inconclusive verifier markers,
-environment setup/reconstruction signals for missing dependencies, toolchains,
-or build artifacts, dependency manifest/lockfile setup-validation signals,
-HarnessAudit-style harness-safety signals for protected-resource access, external information transfer, destructive operations, and oracle access,
-candidate-file dossier signals for broad pre-edit inspection without a compact dossier,
-root-cause hypothesis signals for repair edits after failed verifiers without an explicit diagnosis,
-targeted-fix manifest signals for repair edits after failed verifiers without a fix plan,
-trajectory-cleanup signals for base64/data-URI blobs, high-entropy encoded output, duplicate output, and excessive truncation,
-skill-view fit/timing signals, per-target edit localization signals, large edit-surface
-signals, scratch/probe artifact signals, redundant tool-call signals,
-redundant failing-verifier rerun signals, blind-repair signals, post-edit regression-cycle signals,
-AHE publish-state mutation signals, latest post-edit verifier signals, post-edit and final-state
-diff-review signals, final-edit validation stability/lucky-pass signals, broad-validation signals,
-CI-derived validation signals,
-source-research recency signals,
-process-defect scoring, AHE-style change-evaluation verdicts, submission bundle manifest readiness/hash metadata,
-and trajectory-quality fields are copied to
-`benchmarkResult.traceSummary` for harness-side scoring. `benchmarkResult.usage`
-also aliases the native usage block for cost-aware leaderboards. Native verifier
-trace previews preserve both head and tail output so final test summaries survive
-noisy install/build logs. `benchmarkResult.experienceCard` includes bounded
-task-alignment/spec-compliance/reward-hack/harness-safety/long-horizon/proactivity risk blocks, component-observability edit classification for AHE-style surface attribution, including SWE-WebDevBench canary/frontend-backend/security validation signals, SWE-Cycle lifecycle/setup/test-generation/judge validation signals, SWE-CI evolution/checklist/CI-loop validation signals, and Pi-Bench context-contract/hidden-intent/clarification/privacy/completion evidence, root-cause hypothesis state and targeted-fix counts for failed-verifier repair edits, decision-observability predictions for edits and validation-reliability evidence
-for final verifier stability, broad validation, and CI-derived validation, plus
-context-utilization precision/miss evidence, candidate-dossier status, and
-trajectory-cleanup summaries for retrieval-aware scoring and avoiding noisy prior
-traces, plus run-efficiency action/usage/cost/time evidence for cost-aware scoring. Prior
-experience hints also expose compact source-research coverage, including
-hit/error counts, targeted/fresh coverage, recency windows, top URLs, and
-Kaggle fallback status. When present, `benchmarkResult.traceSummary` also
-includes the redacted ACC-style task/context/answer compilation from the native
-Cawdex trace for retrieval, replay, or training-data curation.
-It also includes `changeEvaluation` and `submissionBundleManifest` when present, so leaderboard
-submission tooling can inspect artifact hashes and missing official score/session
-fields without parsing the full summary.
-
-Inside benchmark mode, the read-only `benchmark_context` preflight also surfaces
-CI workflow run commands plus setup actions, env key names, service containers,
-job containers, and images from GitHub Actions, GitLab CI, CircleCI, Azure
-Pipelines, and Jenkins files. Env values are not printed. Agents can reconstruct
-the relevant CI environment and then reproduce project-native test/build/lint
-steps before finalizing.
-It also separates reusable prior local benchmark experience from similar failed
-or unsafe prior runs, so context reuse stays method-level and current verifier
-evidence remains authoritative. Pi-Bench-like tasks additionally prefer prior
-experience with complete context/hidden-intent/clarification/privacy/completion
-proactivity ledgers and surface incomplete ledgers as warnings. AHE change
-evaluations also participate in reuse: confirmed manifests can rank higher,
-while contradicted, regression-risk, pending-verification, missing-prediction,
-or missing-regression-forecast manifests are warnings rather than replay hints.
-Context-utilization evidence participates in reuse as well: concise runs whose
-inspected context was used by the eventual patch and whose pre-edit search was
-compressed into a candidate-file dossier can rank higher, while low-utilization,
-missing-dossier, or pre-edit context-bloat runs are warnings rather than replay
-hints.
-AHE-style cleanup evidence participates in reuse too: prior runs with encoded blobs,
-duplicate observations, or excessive truncation are surfaced as warnings instead of
-replay hints.
-AHE-style diagnosis evidence participates in reuse too: prior runs that repaired after
-failed verifiers without a root-cause hypothesis are surfaced as warnings instead of
-replay hints.
-AHE-style fix-plan evidence participates in reuse too: prior runs that repaired after
-failed verifiers without a targeted-fix manifest are surfaced as warnings instead of
-replay hints.
-Interactive type-ahead is preserved across active-turn cancellation and
-permission interruptions, so user drafts return to the prompt instead of being
-silently submitted while the harness is still running.
-
-Useful env vars:
-
-- `CAWDEX_KBENCH_COMMAND` or `CAWDEX_KBENCH_COMMAND`: command used to launch Cawdex, default `cawdex`.
-- `CAWDEX_KBENCH_PERMISSION`: permission flag value, default `yolo`.
-- `CAWDEX_KBENCH_EXTRA_ARGS`: extra Cawdex CLI flags.
-- `CAWDEX_KBENCH_ARTIFACT_DIR`: directory for redacted instruction/stdout/stderr and trace files.
-- `CAWDEX_BASH_TIMEOUT_MS`: default Cawdex `bash` tool timeout; the adapter defaults to `300000` when unset.
-
-Provider keys should be passed via normal Cawdex env config or KBench's
-`--api-key-env`, which the runner forwards as Cawdex `--api-key-env`.
+# Cawdex KBench Adapter
+
+This directory is a KBench `custom-adapter` for Cawdex.
+
+```bash
+kbench run \
+  --benchmark swe \
+  --harness custom-adapter \
+  --adapter /path/to/resources/kbench/cawdex_agent \
+  --model-name openrouter/free \
+  --instruction "Fix the bug"
+```
+
+The runner reads the KBench JSON payload from `KBENCH_ADAPTER_INPUT` or stdin,
+invokes `cawdex --prompt "/benchmark ..."` in task mode, and emits one
+`AdapterRunnerOutput` JSON object to stdout.
+Known KBench slugs are mapped to benchmark profiles before dispatch:
+`swe`/`swe-bench`, `tb2`/`terminal-bench`, `terminalworld`/`terminal-world`,
+`swe-chain`,
+`swe-cycle`/`fullcycle`/`swe-judge`, `swe-ci`/`swecibench`, `swe-prbench`/`prbench`/`pr-review`, `tml-bench`/`tabular-ml`/`kaggle-ml`, `pi-bench`/`proactive-assistant`, `ci-repair`/`ci-repair-bench`, `roadmapbench`, `saasbench`,
+`swe-bench-mobile`, `webdevbench`/`swe-webdev-bench`, `appworld`, `browsecomp`/`browsecompplus`, and
+`tau2`/`tau-bench` use specialized prompts; unknown slugs use
+`generic`.
+
+The output includes redacted instruction/stdout/stderr artifact refs, native
+Cawdex trace refs, and redacted git patch/status refs when the task
+worktree is a git repo. If a native `summary.json` exists, compact verifier
+evidence, including parsed counts, compact failure signatures, and final-answer
+verification-claim plus incomplete/blocked completion evidence, usage/cost
+telemetry, cost-efficiency risk, invalid tool-action telemetry, task-contract checklist completion/no-edit/test-edit signals,
+task-alignment risk signals, spec-compliance risk signals, reward-hack risk signals, long-horizon coverage risk signals, Pi-Bench proactivity ledger signals, incomplete/inconclusive verifier markers,
+environment setup/reconstruction signals for missing dependencies, toolchains,
+or build artifacts, dependency manifest/lockfile setup-validation signals,
+HarnessAudit-style harness-safety signals for protected-resource access, external information transfer, destructive operations, and oracle access,
+candidate-file dossier signals for broad pre-edit inspection without a compact dossier,
+root-cause hypothesis signals for repair edits after failed verifiers without an explicit diagnosis,
+targeted-fix manifest signals for repair edits after failed verifiers without a fix plan,
+trajectory-cleanup signals for base64/data-URI blobs, high-entropy encoded output, duplicate output, and excessive truncation,
+skill-view fit/timing signals, per-target edit localization signals, large edit-surface
+signals, scratch/probe artifact signals, redundant tool-call signals,
+redundant failing-verifier rerun signals, blind-repair signals, post-edit regression-cycle signals,
+AHE publish-state mutation signals, latest post-edit verifier signals, post-edit and final-state
+diff-review signals, final-edit validation stability/lucky-pass signals, broad-validation signals,
+CI-derived validation signals,
+source-research recency signals,
+process-defect scoring, AHE-style change-evaluation verdicts, submission bundle manifest readiness/hash metadata,
+and trajectory-quality fields are copied to
+`benchmarkResult.traceSummary` for harness-side scoring. `benchmarkResult.usage`
+also aliases the native usage block for cost-aware leaderboards. Native verifier
+trace previews preserve both head and tail output so final test summaries survive
+noisy install/build logs. `benchmarkResult.experienceCard` includes bounded
+task-alignment/spec-compliance/reward-hack/harness-safety/long-horizon/proactivity risk blocks, component-observability edit classification for AHE-style surface attribution, including SWE-WebDevBench canary/frontend-backend/security validation signals, SWE-Cycle lifecycle/setup/test-generation/judge validation signals, SWE-CI evolution/checklist/CI-loop validation signals, and Pi-Bench context-contract/hidden-intent/clarification/privacy/completion evidence, root-cause hypothesis state and targeted-fix counts for failed-verifier repair edits, decision-observability predictions for edits and validation-reliability evidence
+for final verifier stability, broad validation, and CI-derived validation, plus
+context-utilization precision/miss evidence, candidate-dossier status, and
+trajectory-cleanup summaries for retrieval-aware scoring and avoiding noisy prior
+traces, plus run-efficiency action/usage/cost/time evidence for cost-aware scoring. Prior
+experience hints also expose compact source-research coverage, including
+hit/error counts, targeted/fresh coverage, recency windows, top URLs, and
+Kaggle fallback status. When present, `benchmarkResult.traceSummary` also
+includes the redacted ACC-style task/context/answer compilation from the native
+Cawdex trace for retrieval, replay, or training-data curation.
+It also includes `changeEvaluation` and `submissionBundleManifest` when present, so leaderboard
+submission tooling can inspect artifact hashes and missing official score/session
+fields without parsing the full summary.
+
+Inside benchmark mode, the read-only `benchmark_context` preflight also surfaces
+CI workflow run commands plus setup actions, env key names, service containers,
+job containers, and images from GitHub Actions, GitLab CI, CircleCI, Azure
+Pipelines, and Jenkins files. Env values are not printed. Agents can reconstruct
+the relevant CI environment and then reproduce project-native test/build/lint
+steps before finalizing.
+It also separates reusable prior local benchmark experience from similar failed
+or unsafe prior runs, so context reuse stays method-level and current verifier
+evidence remains authoritative. Pi-Bench-like tasks additionally prefer prior
+experience with complete context/hidden-intent/clarification/privacy/completion
+proactivity ledgers and surface incomplete ledgers as warnings. AHE change
+evaluations also participate in reuse: confirmed manifests can rank higher,
+while contradicted, regression-risk, pending-verification, missing-prediction,
+or missing-regression-forecast manifests are warnings rather than replay hints.
+Context-utilization evidence participates in reuse as well: concise runs whose
+inspected context was used by the eventual patch and whose pre-edit search was
+compressed into a candidate-file dossier can rank higher, while low-utilization,
+missing-dossier, or pre-edit context-bloat runs are warnings rather than replay
+hints.
+AHE-style cleanup evidence participates in reuse too: prior runs with encoded blobs,
+duplicate observations, or excessive truncation are surfaced as warnings instead of
+replay hints.
+AHE-style diagnosis evidence participates in reuse too: prior runs that repaired after
+failed verifiers without a root-cause hypothesis are surfaced as warnings instead of
+replay hints.
+AHE-style fix-plan evidence participates in reuse too: prior runs that repaired after
+failed verifiers without a targeted-fix manifest are surfaced as warnings instead of
+replay hints.
+Interactive type-ahead is preserved across active-turn cancellation and
+permission interruptions, so user drafts return to the prompt instead of being
+silently submitted while the harness is still running.
+
+Useful env vars:
+
+- `CAWDEX_KBENCH_COMMAND` or `CAWDEX_KBENCH_COMMAND`: command used to launch Cawdex, default `cawdex`.
+- `CAWDEX_KBENCH_PERMISSION`: permission flag value, default `yolo`.
+- `CAWDEX_KBENCH_EXTRA_ARGS`: extra Cawdex CLI flags.
+- `CAWDEX_KBENCH_ARTIFACT_DIR`: directory for redacted instruction/stdout/stderr and trace files.
+- `CAWDEX_BASH_TIMEOUT_MS`: default Cawdex `bash` tool timeout; the adapter defaults to `300000` when unset.
+
+Provider keys should be passed via normal Cawdex env config or KBench's
+`--api-key-env`, which the runner forwards as Cawdex `--api-key-env`.

package/resources/kbench/cawdex_agent/adapter.manifest.json

@@ -1,19 +1,19 @@
-{
-  "schemaVersion": "kbench.adapter/v1",
-  "id": "cawdex",
-  "kind": "node",
-  "entry": "./runner.mjs",
-  "version": "0.1.0",
-  "supportedBenchmarks": ["swe", "tb2", "sae"],
-  "capabilities": {
-    "runModes": ["task"],
-    "machineReadableStdout": true,
-    "supportsPatchOutput": false,
-    "supportsTrajectory": true,
-    "supportsToolCallTrace": true,
-    "supportsResume": false,
-    "supportsImages": false,
-    "supportsSandboxBridge": false,
-    "supportsPromptTemplate": false
-  }
-}
+{
+  "schemaVersion": "kbench.adapter/v1",
+  "id": "cawdex",
+  "kind": "node",
+  "entry": "./runner.mjs",
+  "version": "0.1.0",
+  "supportedBenchmarks": ["swe", "tb2", "sae"],
+  "capabilities": {
+    "runModes": ["task"],
+    "machineReadableStdout": true,
+    "supportsPatchOutput": false,
+    "supportsTrajectory": true,
+    "supportsToolCallTrace": true,
+    "supportsResume": false,
+    "supportsImages": false,
+    "supportsSandboxBridge": false,
+    "supportsPromptTemplate": false
+  }
+}