Repository navigation
Conversation
…tric A seeded referral stream with an exception mix and its ground truth, a scripted operator charged from a versioned cost model, an independent SQL oracle, the arm interface (lifecycle-arm/1), and the metric (verified completed items per operator-hour with x0.5/x2 bands). The OpenEMR adapter injects faults and server-side UI changes; an in-memory app and toy arms check the bookkeeping in tests. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The OpenAdapt arm runs the certified referral program on the governed Standard path (or one CLI run per item with mode=cli) and records what the product refused. Trigger SQL uses a delimiter; ports and project are overridable; the fixture is in Env.extra; arms may report maintenance phases inside an item. Workload 2026-10-20.2 widens the surname pool. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… soak harness) Agent alone (screenshots), agent with structured tools and REST read-back, agent-written Playwright+REST script it maintains after breakage, hand-written script baseline; dated price table, shared spend cap, one-command resumable runner, cost per 1,000 verified items, offline OpenEMR mock and tests. Records that current main refuses the OpenEMR referral's API step. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…nners that have them Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ases, safer dry runs - local_model.py: an Ollama /api/chat client in the SDK's messages shape; the computer toolset maps to one 'computer' function. Local calls cost $0. - compare --models local: cua_local (qwen3-vl:8b, first 10 items), agent_tools_local and agent_script_local (llama3.1:8b). - agent_script returns maintenance as ItemResult.phases; economics no longer re-reads detail (the runner writes them as maintenance rows). - dry runs keep the save blocked 5 s after the script returns, then reload. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
abrichr
force-pushed
the
claude/lifecycle-arms
branch
from
October 20, 2026 17:51
510618e to
7f52886
Compare
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Draft. Stacked on the lifecycle harness commit 4ec5247 (branch claude/objective-sanderson-e24b08); merge that first.
Adds the competing arms to benchmark/lifecycle/arms/: a screenshot-only agent (cua_alone), an agent with DOM/accessibility tools and a read-only REST read-back (agent_tools), an agent-written Playwright + REST script that the agent fixes when it breaks (agent_script), and a hand-written script (human_script). Also adds a dated model price table, a $30 spend cap across all arms, a one-command resumable runner (compare.py), cost per 1,000 verified items (economics.py, report.py), the scoreboard listing, and protocol amendment 10.
Records that current main refuses the OpenEMR referral's API step (lifecycle/results/api_tier/), so OpenAdapt's gated numbers are the GUI path's.
Not run on the real fixture yet: no API key, and no fixture lease. Tests: tests/test_lifecycle_arms.py and tests/test_lifecycle_arms_mock.py (an offline OpenEMR mock with scripted model stand-ins), 33 passing locally together with the harness tests.
🤖 Generated with Claude Code