DEV COMPANION 0.5.0 — RECORDED LOCAL WORKFLOW Excerpted commands/outputs from a maker evaluation; private paths redacted, status JSON summarized. This is not an independent developer trial or a coding-quality benchmark. Local inference; this run was not network-isolated. Model: qwen3:4b-instruct-2507-q4_K_M Model digest: 0edcdef34593eac1aa2be9c7d06c432dcf81945adca5eca2f27662c18f168ba0 A multi-file Python billing fixture contains catalogue, money, pricing and invoice modules, an app and two unittest modules. Nine tests, three failures. Presence was running separately; no companion chat was open. $ dot check add invoice-tests -- python3 -m unittest discover -s tests -v $ dot verify invoice-tests Recorded run #1, exit 1. Original output: Recorded [run:1] · exit 1 · 108 ms Ask for help with `devcompanion explain`, or continue in `devcompanion chat`. test_discount_reduces_total (test_invoice.InvoiceTests.test_discount_reduces_total) ... FAIL test_empty_basket (test_invoice.InvoiceTests.test_empty_basket) ... ok test_full_discount (test_invoice.InvoiceTests.test_full_discount) ... FAIL test_no_discount (test_invoice.InvoiceTests.test_no_discount) ... ok test_catalogue_quantity (test_pricing.PricingTests.test_catalogue_quantity) ... ok test_discount_rounds_to_cents (test_pricing.PricingTests.test_discount_rounds_to_cents) ... FAIL test_excess_discount_rejected (test_pricing.PricingTests.test_excess_discount_rejected) ... ok test_negative_discount_rejected (test_pricing.PricingTests.test_negative_discount_rejected) ... ok test_unknown_product_rejected (test_pricing.PricingTests.test_unknown_product_rejected) ... ok ====================================================================== FAIL: test_discount_reduces_total (test_invoice.InvoiceTests.test_discount_reduces_total) ---------------------------------------------------------------------- Traceback (most recent call last): File "[evaluation project]/tests/test_invoice.py", line 6, in test_discount_reduces_total self.assertEqual(invoice_total([("keyboard", 2), ("mouse", 1)], 0.2), 100.0) ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ AssertionError: 150.0 != 100.0 ====================================================================== FAIL: test_full_discount (test_invoice.InvoiceTests.test_full_discount) ---------------------------------------------------------------------- Traceback (most recent call last): File "[evaluation project]/tests/test_invoice.py", line 10, in test_full_discount self.assertEqual(invoice_total([("keyboard", 1)], 1), 0.0) ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ AssertionError: 100.0 != 0.0 ====================================================================== FAIL: test_discount_rounds_to_cents (test_pricing.PricingTests.test_discount_rounds_to_cents) ---------------------------------------------------------------------- Traceback (most recent call last): File "[evaluation project]/tests/test_pricing.py", line 7, in test_discount_rounds_to_cents self.assertEqual(discounted_total(19.99, 0.1), 17.99) ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ AssertionError: 21.99 != 17.99 ---------------------------------------------------------------------- Ran 9 tests in 0.002s FAILED (failures=3) $ dot proactive show 1 Insight #1 · incident #1 · ready Observed: tests/test_invoice.py:6:1: AssertionError: 150.0 != 100.0 Hypothesis: The discount calculation in invoice_total is incorrectly applying the discount rate as a multiplier instead of a reducer. Suggested next step: Verify the pricing logic in discounted_total and ensure it correctly applies a discount by subtracting the discount from the subtotal, not adding it. Evidence: [reported:tests/test_invoice.py:6:1] Local model: qwen3:4b-instruct-2507-q4_K_M (0edcdef34593eac1aa2be9c7d06c432dcf81945adca5eca2f27662c18f168ba0) at 1790919080 Rating: unrated The explanation appeared automatically after project opt-in, with chat closed. The observation is recorded evidence; the cause is a tentative model hypothesis. No automatic patch preparation, application or execution took place. $ dot incident fix 1 --foreground $ dot task review 1 diff --git a/billing/pricing.py b/billing/pricing.py --- a/billing/pricing.py +++ b/billing/pricing.py @@ -3,4 +3,4 @@ def discounted_total(subtotal, discount_rate): if not 0 <= discount_rate <= 1: raise ValueError("discount_rate must be between zero and one") - return cents(subtotal * (1 + discount_rate)) + return cents(subtotal * (1 - discount_rate)) Proposed code has not been tested. Approve with `task apply 1` (in chat: /apply 1); reject with `task reject 1`. The maker inspected this diff before explicit application. Only pricing.py changed; every test and other protected file retained its original SHA-256. $ dot task apply 1 Applied task [1]. Run the relevant checks before relying on the change. $ dot task verify 1 invoice-tests Task [1] verified; [run:2] invoice-tests exit 0. test_discount_reduces_total (test_invoice.InvoiceTests.test_discount_reduces_total) ... ok test_empty_basket (test_invoice.InvoiceTests.test_empty_basket) ... ok test_full_discount (test_invoice.InvoiceTests.test_full_discount) ... ok test_no_discount (test_invoice.InvoiceTests.test_no_discount) ... ok test_catalogue_quantity (test_pricing.PricingTests.test_catalogue_quantity) ... ok test_discount_rounds_to_cents (test_pricing.PricingTests.test_discount_rounds_to_cents) ... ok test_excess_discount_rejected (test_pricing.PricingTests.test_excess_discount_rejected) ... ok test_negative_discount_rejected (test_pricing.PricingTests.test_negative_discount_rejected) ... ok test_unknown_product_rejected (test_pricing.PricingTests.test_unknown_product_rejected) ... ok ---------------------------------------------------------------------- Ran 9 tests in 0.001s OK Saved statuses (summarized): task verified, verification_current=true, incident check-passed. Source changes made the earlier insight stale. REMAINING ERRORS — ALSO RECORDED The returning billing chat correctly named run #2 and nine passing tests, but invented a basket example not in the code. That text is not accepted evidence. The recorded check independently establishes the pass. A separate official Flask 3.1.2 checkout (commit 2c1b30d0503cfb064f1cb252e6614a06915a362a, 234 tracked files) passed all 35 helper tests before an injected boolean-condition fault. Four tests then failed. Context followed Python imports through its src/ layout. Model attempts proposed wrong or incomplete patches. The maker rejected one that accepted only 'true'. A narrowed request produced a patch missing empty strings; the unchanged suite caught it (one failed, 34 passed). Revision attempts failed exact-text validation and changed no files. A citation pointed at a real but unrelated source label. Citation existence does not prove a diagnosis is supported. This larger trial exposed a file-priority bug and a revision prompt size bug; both were corrected with regression tests. The model reliability failures remain. Do not treat this preview as a reliable unattended coding agent. Version 0.5 has 95 passing Rust tests. Coverage includes multiple error formats, related source, quiet/cooldown/cancellation, stale evidence, setup, updating and private trial reports. Generated English voice passed in v0.4; voice code did not change for v0.5. A real microphone and speaker still need hands-on testing.