# Reference validation — October 10, 2026 Execution: existing Docker installation, Linux `aarch64`, OpenCode **v2.0.20**, model **opencode-go/deepseek-v4.1-flash**. The test binaries were compiled from the current checkout and run against a separate SQLite database and isolated fixtures. No definitions were published to the normal server database and no phone alerts were sent. Notification checks used an isolated local HTTP endpoint. | Reference workflow | Live result | Duration | |---|---|---| | Incident extraction → report | Passed: exact incident evidence/counts, handoff, session readback and local success receipt | 6.63s | | Edit → independently verify | Passed: exact file bytes, untouched fixtures, harness-run `cmp`, completed read/edit/shell evidence, retained stage diff | 14.60s | | Missing required input | Passed: deliberate blocked outcome terminates as failed; missing file not invented; read attempt and local failure receipt retained | 4.21s | | Accounting and one-token budget | Initially failed the acceptance test; passed after the accounting fix below | 2.43s on focused rerun | ## Defect found and fixed The first accounting run succeeded despite a one-token limit. Its retained V2 session had **no messages and zero usage counters**, even though `/generate` returned a valid result. Polling that session treated placeholder zeros as known accounting and did not enforce the intended limit. Budgeted V2 text calls now use tracked `/prompt` → bounded active-session polling → message readback, with **all tools denied**. Their messages are retained in the stage audit, usage is read from the tracked session, and a completed token-budgeted response with zero reported tokens fails closed. Native V1 handling and unbudgeted text generation keep their existing paths. A later reliability fix replaced the V2 `.../experimental/session/{id}/wait` request, which timed out at the CLI's own ~six-minute limit (HTTP 499) on long stages, with roughly one-second `GET /api/session/active` polls under the stage context/deadline. That change has fake-CLI regression coverage in `internal/stage`; it has not yet been exercised by a fresh live run. The focused live rerun retained **952 reported tokens and $0.00016440**, failed with `stage budget exceeded`, and stopped its owned session. This is reported usage, not an independently reconciled provider bill. A one-token limit can be overshot by a single in-flight request; this run demonstrates exhaustion handling, not an exact hard cap or proof of immediate mid-stream cancellation. The three previously passing workflows were not charged again. Their copied artifacts were rechecked with current independent assertions; the fixed accounting artifacts were rechecked separately. There has **not** been a fresh full-suite live run of all four cases after the fix. ## Evidence and reproduction Retained container directories for this trial: - `/data/reference-validation/pipelines-reference-284076920`: first complete run. - `/data/reference-validation/pipelines-reference-2247403140`: fixed accounting rerun. Each has isolated SQLite history, pinned definitions, run/session JSON, fixture files and `report.json`. Host copies are in the session's approved temporary directory under `reference-first/` and `reference-accounting-fixed/`. Commands that performed live inference: ```sh docker exec -e PIPELINES_REFERENCE_LIVE=1 \ -e PIPELINES_REFERENCE_MODEL=opencode-go/deepseek-v4.1-flash \ -e PIPELINES_REFERENCE_DIR=/data/reference-validation \ pipelines-pipelines-1 /data/reference-validation/reference-20261010.test \ -test.run '^TestReferenceWorkflowsLive$' -test.v -test.timeout 15m docker exec -e PIPELINES_REFERENCE_LIVE=1 \ -e PIPELINES_REFERENCE_MODEL=opencode-go/deepseek-v4.1-flash \ -e PIPELINES_REFERENCE_DIR=/data/reference-validation \ pipelines-pipelines-1 /data/reference-validation/reference-20261010-fixed.test \ -test.run '^TestReferenceWorkflowsLive/accounting-and-budget$' -test.v -test.timeout 5m ``` `TestReferenceArtifactsReview` rechecks copied run JSON/fixtures without inference; `PIPELINES_REFERENCE_REVIEW_DIR` selects the copy and `-run` selects its cases. See [reference workflows](reference-workflows.md) for a fresh run and its opt-in, cost, retention and safety boundaries. ## Remaining unverified - Other models/providers and native V1 schema/tool compatibility. - Real ntfy authentication/storage and actual phone reception. - Actual provider outages, timeouts, long-stream interruption and retry behavior. - General workflow correctness beyond these bounded fixtures. The fix is in the source/test binaries, **not deployed to the installed Docker server**. Rebuild the installation when ready to use it for ordinary runs.