Two needs_opencode live tests are stale: the agent set predates the subagents, and the shape floors have drifted #97
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
With the pinned opencode 1.18.21 on PATH,
pytest -m needs_opencode tests/integration/has two failures on main. Both fail the same way before and after #89, so they are stale tests, not regressions.test_only_the_builtin_agents_exist_in_an_isolated_homeexpects agent names to be a subset of opencode's own built-in agents. Since412cea4(2026-09-13), the run config also declares Braid'sreaderandtestsubagents, so the set is larger. The allowed set should include the roster the config declares (SUBAGENT_NAMES), and the test should still catch an operator agent leaking in.test_each_shape_puts_its_measured_tools_on_the_wiremeasures the implement floor at 33,195 characters againstMEASURED_FLOOR_CHARS[IMPL]= 32,341 (tolerance 200). The lean floor sits at 22,736 against 22,545, just inside the tolerance. Re-measured per tool:task: 3,899 → 4,562. Its description now lists theexplore,readerandtestroster (412cea4).bash: 5,312 → 5,425. Its description now names the run's ownTMPDIR(#36), so the floor moves with the length of the run's path.editstill measures 3,049, which iseditpluswrite(1,996 + 1,053) because turningeditoff removeswritetoo. It is unchanged.The fix updates
MEASURED_FLOOR_CHARS,TOOL_COST_CHARSand theosf.engine.shapedocstring table, and the unit tests derived from them. The lean saving is now 10,459 characters, or 3,486 tokens a turn.shape.floor_tokensis used only by tests, so budgeting doesn't change.config.FLOOR_LEAN/FLOOR_IMPL(7,515 / 11,009), the budget estimator's inputs, come from a separate osfd measurement and are left as they are.Fixed and pushed to main in
2ee5069.Rebasing onto main brought in #17, which made a third test stale. The numbers here supersede the ones in the description:
test_only_the_builtin_agents_exist_in_an_isolated_homenow allows opencode's own agents plusSUBAGENT_NAMES(reader,explore,test,browser).test_the_run_is_isolated_from_the_operators_opencodeandtest_operator_config_leaks_in_without_home_isolationnow expect the run's own Playwright MCP server (RUN_MCP = {"playwright"}, from #17) and nothing else.test_each_shape_puts_its_measured_tools_on_the_wire: re-measured on 1.18.21 under the test's own path.taskis 4,973, because its description lists the four-subagent roster.bashis 5,445, because its description names the run'sTMPDIR.config.FLOOR_LEAN/FLOOR_IMPL, the budget estimator's inputs, are unchanged.Results:
-m needs_opencode tests/integration/passes 23 of 23 with 1.18.21, and./tools/test.sh fullpasses (backend 1,858, UI 444, browser 108; no osfd, so no@livespecs).