fix(headless): run the DeepSWE Maka arm inside the Pier task container #1473

Problem

The DeepSWE A/B in #1343 runs the two arms across different execution boundaries.

Codex CLI and Kimi Code run inside the Pier task container and execute their tools locally. Maka runs the full task-run controller on the host, then sends every Bash, Read, Write, Edit, Glob, and Grep command through _ToolExecutorServer and environment.exec.

That difference exposes only the Maka arm to Pier's command-output transport on every tool call. Pier 0.3.0 starts docker compose exec with stderr=asyncio.subprocess.STDOUT, then returns the merged bytes as stdout and stderr=None. The Maka bridge converts that into merged stdout plus empty stderr even though IsolatedCommandResult requires separate streams.

The recorded DeepSWE canaries contain real consequences:

• GNU Awk warnings from Glob/Grep enter the Base64 stdout frame and produce payload is not canonical Base64.
• Reads of missing paths lose the intended diagnostic and become transport-integrity failures.
• Edit reads can fail closed for the same reason.
• Write and Edit write failures lose their stderr diagnostics.
• Bash stderr is reported as stdout.

Historical artifacts include 11 Read integrity failures and repeated Glob/Grep failures in both serial and parallel runs. The paired Codex arm has none because individual Codex tool results never cross environment.exec; Pier launches only the outer CLI process, and Codex writes its rollout and output as container artifacts.

#1343 describes a system-level comparison, but it also requires infrastructure failures to be separated from model outcomes. These failures are specific to the Maka-side bridge and cannot be accepted as model results.

Reproduction

Use current main, official datacurve-pier==0.3.0, and a DeepSWE 1.1 Linux/amd64 task image. Run the Maka arm in host task-run mode with --isolation harbor-http.

Two reliable triggers are:

1. Run a command-backed Glob or Grep fallback whose generated Awk -v value contains \.. GNU Awk writes a warning to stderr; Pier merges it into the Base64 stdout frame; parseFileToolOutputFrame() rejects the payload.
2. Read a path that does not exist. The file tool writes the path error to stderr. Pier merges it into stdout, so the model receives an integrity failure instead of the path diagnostic.

The same executor behavior can be observed directly with:

printf OUT
printf ERR >&2
exit 7

Pier returns both strings through stdout and leaves stderr empty.

Proposed change

Run the complete Maka Headless task-run controller inside the Pier task container, matching the container-CLI arms. Reuse the existing Headless path:

--mode task-run
--isolation harbor-local

Use a pinned Linux/x64 Maka and Node toolchain mounted read-only. Route provider calls through the existing host credential proxy, with only an attempt-scoped proxy token inside the container. Write task-run, session, runtime, usage, and terminal artifacts under /logs/agent, then hydrate them after the run as the Codex adapter does.

Pier should start and supervise the agent process, not carry structured results for every Maka tool call.

Acceptance criteria

• The DeepSWE Maka arm runs the full task-run controller inside the official Pier task container.
• DeepSWE tool calls use the existing harbor-local executor, not _ToolExecutorServer or harbor-http.
• Maka and its competitor use the same task image, container user, workdir, network policy, deadline authority, and verifier.
• Both arms reach the provider through the host credential proxy; real provider credentials stay on the host.
• A pinned, fingerprinted Linux/x64 Maka and Node toolchain is mounted read-only and needs no container-side network installation.
• Task-run results, runtime events, sessions, usage, and terminal state are recovered from container artifacts.
• A missing-path Read returns the path diagnostic.
• Glob/Grep remain correct when GNU Awk writes a warning to stderr.
• Bash preserves distinct stdout and stderr.
• A real-model two-task canary passes the artifact and execution-identity audit required by eval(headless): DeepSWE 1.1 discriminative-30 harness A/B — Maka vs Kimi Code and Codex CLI #1343 using unmodified official Pier.
• Historical cells containing tool transport-integrity failures are excluded and rerun.
• The run manifest and report record that both agents executed inside the Pier task container.

Scope

This change should not patch or fork Pier, loosen the existing integrity checks, filter known warning strings, or make the two agents use identical internal tools. Their tool implementations are part of the systems being compared; the task-container boundary is not.

The current host bridge may remain for a demonstrated non-benchmark caller. If no supported caller still needs it, remove it instead of maintaining two task-run execution paths.

Related

• eval(headless): DeepSWE 1.1 discriminative-30 harness A/B — Maka vs Kimi Code and Codex CLI #1343 tracks the DeepSWE comparison and is blocked on a valid canary.
• feat(headless): run DeepSWE via Pier as a benchmark axis of the harness A/B #1398 introduced the current DeepSWE/Pier benchmark path.
• fix(headless): prevent isolated Edit from decoding executor stdout noise into files #1111 / fix(headless): frame isolated Edit file reads #1147 made Edit fail closed instead of corrupting files.
• fix(headless): remove Harbor job notifications from tool output #1270 / fix(headless): suppress Harbor job-control notifications #1271 removed known Harbor job-control output.
• fix(headless): frame structured tool output #1273 framed command-backed Read, Glob, and Grep output.
• eval(headless): Pier container-CLI arm serialized on host proxy port 443 caps A/B concurrency #1424 tracks the fixed-port proxy concurrency ceiling. Moving Maka into the container makes its shared-proxy direction more important for parallel runs.