docs/worker-compatibility.md

# Worker compatibility

The publishing helper runs with `agentenv-framework==0.9.1324`. Its default gateway deployment and
saved Docker image archives produce records that the older `agent-env==0.9.1181` worker can read.
The older worker uses protocol `0.1.196`; it contains every step in the seven-step TradeOps graph.

The supported older-worker path requires:

- Build and publish images with their tar archives. Its image schema rejects context-only artifacts.
- Use gateway mode. Its `MCPServerEnv` reader always constructs a gateway provider and does not honor
  the newer `env_provider_type="server"` setting.
- Configure the gateway, service-database environment, and environment-state provider required by
  that worker's deployment path.
- Use its configured S3 object store for the default prompt trajectory location, or supply an
  explicit `trajectory_output_prefix` understood by its object store.

No change to the TradeOps environment, public tools, A2A adapter, or business verifier was needed.
Gateway tools retain their nine original names. The existing verifier discovers the TradeOps child
card and invokes its advertised grading extension through the gateway.

## Recorded checks

`docs/qa/hub-built-publication-result.json` records a real local publication: two Docker images built
and pushed to a loopback registry, both tar archives uploaded, and all eighteen records created.
A retry reused all eighteen original versions.

`docs/qa/legacy-gateway-result.json` records checks through the exact older gateway: all thirteen
rules scenarios and a scripted three-tool API recovery passed through its MCP and extension routes.

`docs/qa/legacy-worker-result.json` records the older worker reading that built publication's two
image artifacts, verifier, environment, A2A agent, and thirteen tasks. The task project IDs survived
deserialization. Its real `Task.run` executed scenario loading, an A2A prompt, and all three verifier
steps against real loopback services:

| Case | Business score | Agent completion | Combined reward |
|---|---:|---:|---:|
| Scripted recovery of `late_fill_duplicate` | 1.0 | 1.0 | 1.0 |
| Deliberate agent failure in `clean_reconciled` | 1.0 | 0 | 0.0 |

Both trajectories were retrieved from the A2A server, saved by the older worker, and read back.
The failing case included one real workspace read before the local response fixture returned an
error. These are protocol and execution checks using scripted responses, not model-performance data.

The worker and protocol source versions were exact. The worker probe used the installed dependency
versions listed in its receipt, rather than reproducing the entire deployed dependency lock. The
modern A2A service used protocol `0.1.339`. Environment and agent deployment records were injected;
the original deploy steps were skipped. The only prompt override redirected trajectories into the
isolated local object store. The probe did not exercise remote credentials, worker sandbox
provisioning, or a hosted Hub run.

`scripts/check_legacy_worker.py` is the worker-side probe. It requires the selected older source
trees first in `PYTHONPATH`, local publication records, and running loopback gateway/A2A fixtures.
It verifies the imported source versions before running and writes isolated local state beneath a
new `--out` directory. Older dependency sources, runtime stores, raw trajectories, configuration
files, and machine-specific paths are excluded from the public receipt.