[hessabi]

Re-file of the former #6, as requested in #4, rebased onto current main.

The default model is pinned to swe-2-max, which Devin Free accounts can't use. Every run fails with only Error: Upgrade to Pro to access this model (https://devin.ai/pricing) and no guidance.

Changes

• DEFAULT_MODEL honors DEVIN_COMPANION_DEFAULT_MODEL. The pinned default is unchanged, and the override flows into the job's recorded model, so status shows it.
• When a run fails with "Upgrade to Pro" and --model wasn't passed, the output gets one hint line: pass --model <id> (see devin models list) or set the env var. It renders next to the sandbox hint through renderDevinOutput, so sandbox output is unchanged.
• No retry, and exit codes are untouched, so this stays out of the exit-code discussion.
• One README line documents the env var.

Tests: 3 new tests, reusing the fixture's stderrText option:

• the hint shows without --model;
• it's absent with an explicit --model;
• the env var changes the --model passed and the model recorded in status.

Full suite: 112/112.

Note: this PR and the welcome-screen re-file both append tests to the end of tests/companion.test.mjs. Whichever merges second needs a trivial rebase that keeps both blocks; I'm happy to do it.

[TheOnlyFusionCube]

Reviewed against main and issue #4. The override flows into buildDevinPrintArgs and persisted job.model; explicit --model takes precedence, and the three tests exercise these paths. Existing ordinary-failure exit policy is unchanged. One error-reporting issue to address before approval: renderDevinOutput uses stdout as the base when nonempty, and only appends devinFailureMessage(result) when a hint exists. With an explicit --model, modelHint is suppressed, so a failed run with stdout and the Pro-access error only on stderr hides the actual error. Please ensure failed runs with nonempty stdout still show stderr in the rendered result even when no hint is emitted; add a regression test for this case. Also consider matching a more specific model-entitlement error rather than any upgrade to pro text. HOL logs show the same one HIGH hardcoded-secret finding present before this PR; #17 fixes it and has now been merged. Please refresh this branch and rerun both required workflows before merge.