Repository navigation
Ichandra2 - #624
Ichandra2#624ichandrasharma wants to merge 33 commits into
Conversation
# Conflicts: # AGENT_COORDINATION.md
…nd types as agent requests fix(next-env): correct import path for route types in TypeScript environment
… prevent overwriting successful responses
… a delivered agent answer is never overwritten Value from PR #624 by @ichandrasharma, adapted: the PR shipped the behavior change without test updates (6 pinned assertions would have failed as-is) and without a regression test for the bug it fixed. Three changes, one contract: 1. Legacy-lane ComputerUse fallback fires ONLY when nothing handled the turn. The old condition (not handled OR AGENT_REQUEST) fired after the Atom meta-agent had already answered, overwriting its final_output with a generic task-ID bubble and spawning a duplicate background agent — every tool-plan "agent_request" turn lost its answer this way. Atom failing leaves handled=False, so the safety net survives (test_explicit_agent_request still pins it, plus a new regression test: successful answer preserved, ComputerUse not called). 2. Unclassifiable input routes to the general agent instead of a fake search. _classify_intent maps CommandType.UNKNOWN (and unmapped types) plus _fallback_intent_analysis's keyword-free else-branch to AGENT_REQUEST. Rationale: the legacy lane's SEARCH handler returns success over an empty store, so the old SEARCH_REQUEST default dead-ended in "I found 0 results" for precisely the turns no classifier understood. One deadline-bounded Atom run (the legacy lane is already sliced by _FEATURE_ROUTING_MAX_SECONDS) instead of a template non-answer. R82's scoping stands: genuinely search- classified turns still never reach ComputerUse. command_type in the fallback dict stays "search" for rule-classified intents and is "agent" only when that was actually decided (no consumer reads it; kept honest for the _TOOL_PLAN_INTENT_MAP contract). 3. next-env.d.ts: restore the canonical .next/dev/types/routes.d.ts reference (established in 2ae66a0, clobbered by snapshot commit 5449299 which leaked a preview instance's custom distDir .next-preview-crosssrc_fix_0929 into the tracked file). Verified (TESTING=1, backend venv): 31 covpush intent/deadline tests + 39 w115 classify/fallback/route tests green; w89 intent tests identical to pristine; test_chat_orchestrator.py shows the SAME 11 pre-existing env-dependent evidence-test failures on pristine origin/main and on this tree, with one additional pass (the new regression test). Nothing new fails anywhere.
|
Closing — the value from this PR has been harvested and landed on main in aaa883d (credit @ichandrasharma for all three findings). The PR itself is not being merged as-is because it shipped the behavior change without test updates (6 pinned characterization assertions across 4 covpush suites would have failed) and without a regression test for the bug it fixed. What landed (adapted + fixed + tested):
Verified: 70+ intent/routing/deadline tests green; failure sets identical to pristine main everywhere else (11 pre-existing env-dependent evidence tests, unrelated). The earlier commits on this branch (owner-dashboard live payload, pdf-worker path, next-env) had already landed on main via other paths — nothing else in the branch carries unique changes. |
1 similar comment
|
Closing — the value from this PR has been harvested and landed on main in aaa883d (credit @ichandrasharma for all three findings). The PR itself is not being merged as-is because it shipped the behavior change without test updates (6 pinned characterization assertions across 4 covpush suites would have failed) and without a regression test for the bug it fixed. What landed (adapted + fixed + tested):
Verified: 70+ intent/routing/deadline tests green; failure sets identical to pristine main everywhere else (11 pre-existing env-dependent evidence tests, unrelated). The earlier commits on this branch (owner-dashboard live payload, pdf-worker path, next-env) had already landed on main via other paths — nothing else in the branch carries unique changes. |
3 participants
Description
Type of Change
Testing
pytest)tsc --noEmit)Checklist