Each model_request/model_payload repeated the full context, so the trace
grew with the square of the request count: batch 6 Qwen runs hit 64 MiB
at ~75 requests and 14 attempts (13 graded PASS) were discarded.
Requests now record only items after the common prefix with the previous
one (schema embedded-rollout.pi/2), and pi_run.expand rebuilds full
payloads before the unchanged broker/feedback checks. Needs a new image
(trace6) and recalibration before use.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The other users of /home left 57 GiB free; the user chose to keep running
and stop only at 30 GiB. Batch start and runtime disk_low now share that floor.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Agent containers wrote to bind-mounted /workspace and /trace with no
size limit, so a runaway Agent could fill the shared /home. Containers
no longer start below 50 GiB free; a running Agent is stopped when the
disk drops below that floor or its attempt directory passes 2 GiB.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Output limits were baked into pi.mjs inside the runtime image, so every
change (GLM 32000 -> 131072, Qwen 8192 -> 65536) needed a new image and a
full 785-trial recalibration, and the Qwen baseline ran on an image the
repo could not rebuild.
- MODELS holds max_tokens and its wire field per model; pi_run passes
--model and --max-tokens to the adapter, which has no defaults left.
- Receipts record model_max_tokens and model_sampling=service-default;
evidence requires every payload to carry exactly that limit, which
would have caught the SDK's silent 32000 cap.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Bos tasks with 131072-token reasoning produce ~110 bytes per streamed
token plus a full context per request; the trace3 batch hit 16 MiB
before finishing. Adapter and host share the new limit, and the size
tests now derive their data from MAX_TRACE_BYTES. Runtime trace4 carries
this adapter; all 21 tasks recalibrated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The first Bos pilot batch exposed two adapter faults on long responses:
- fs.writeSync(1, ...) ignored partial writes to a nonblocking stdout pipe,
so host stdout lost the tail of 100 KB lines and could no longer match
the native trace. writeAll now loops until every byte is written and
waits briefly on EAGAIN.
- Pi SDK caps max_tokens at 32000 unless told otherwise; GLM-5.3 spent
all of it reasoning and stopped with "length" before any edit. The
adapter now passes the registered 131072 (accepted upstream).
Runtime trace3 is the trace2 digest plus this pi.mjs; 12/12 SDK checks
pass inside it and all 21 tasks were recalibrated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Tasks may declare up to 16 candidate files; the whole template becomes
the Agent workspace while only candidate files are frozen and graded.
- Optional fixtures add more cheats or legal alternatives (five trials
each); optional grader_files ship private tests next to verify.py.
- Tasks without an argv oracle export call_type=verifier ground truth.
- Upstream model routes live in one MODELS table.
- One bounded no-follow reader, one exclusive writer and one SIGTERM
helper replace per-module copies; drop unused validate_task.
frame-decoder keeps its task digest, calibration receipts and exported
query. Verified on host 36: internal network probe, 20-trial calibration
and a real GLM rollout (PASS, evidence VERIFIED, cleanup confirmed).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Add the current and target deployment diagrams, the task author standard
and copyable spec template, the five-field data contract, and record the
Qwen Responses trial and pending-create cleanup fix in progress and
acceptance.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The dedicated Pi network was a plain bridge, so the Agent could reach the
internet and fetch upstream fixes. Create it with --internal, make run
accept only internal networks, and default to --network none.
Verifier limits of 1s per case and 15s per grade turned host load into
false FAIL verdicts. Use 5s per case, 30s compile, 10s isolation probe and
timeout_seconds=60; frame-decoder task version becomes 2.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>