Repository navigation
Supersede ACE - #346
Draft
vemvcs wants to merge 40 commits into
Draft
Supersede ACE#346vemvcs wants to merge 40 commits into
vemvcs wants to merge 40 commits into
Conversation
Replace the old ACE backend with the local cs-agent binary for refactoring operations. This eliminates dependency on the ACE backend service while keeping the existing CWF webview UI. - Add AgentRefactoringService to invoke cs-agent binary - Add agent types and configuration modules - Add VS Code settings for agent model and AWS config - Add platform-specific package targets in Makefile - Update .vscodeignore to include bin/ directory - Add unit tests for mapping logic and config building - Add integration test for real binary invocation
Aligns with the agentic-ace branch reference implementation: - Remove agentModel, awsProfile, awsRegion settings (agent reads its own config) - Update user-facing strings from "CodeScene ACE" to "CodeScene Agent" or "CodeScene Refactoring Agent" - Simplify agent config to only pass token and plugins
Clicking "Extension Settings" in the ACE acknowledgement panel was logging an error because the panel didn't have a handler for the open-settings message type. Reuse the existing handleOpenSettings function from cwf-message-handlers.
The agent was being spawned with a temp directory as cwd, causing "not inside a git repository" errors. Use the git repo root so the agent can access the entire project and run git commands.
The agent was failing with "UnknownError" because the opencode_config with Amazon Bedrock provider settings was missing from the agent config. Also adds the --model argument to explicitly request the Bedrock model.
The agent outputs whole_file_content instead of replacements when change_type is "whole_file". Also fixes the integration test to initialize a git repo and adds assertions verifying the agent writes the expected output file.
Stream progress messages from the refactoring agent binary to the loading view instead of showing hardcoded rotating messages. The most recent stdout line is displayed, with newer lines replacing older ones.
Automatically build and copy CWF from ../cs-webview when running as vemv, otherwise download from GitHub releases. Make build depend on updatecwf to ensure CWF is always up to date.
Update Makefile package targets to reference ~/cs-agent instead of ~/refactoring-agent for darwin arm64 and x64 builds.
Extract spawnAgentProcess, setupStderrHandler, setupStdoutHandler from invokeAgent to reduce cyclomatic complexity. Refactor applyChanges to use filter/reduce pattern to eliminate nested conditionals. Fix post-refactoring tests by initializing git repo and mocking extension path.
Add global mocha afterAll hook to clean up test directories after all tests complete. This catches async operations that recreate directories after per-suite teardowns run.
Match the new cs-agent schema that uses fixed enum values for summary and reasoning instead of freeform strings.
Move render-code-fix input/output JSON files to a temp directory instead of the repository root to avoid cluttering user projects. The agent's io_json_dir config option specifies where the reporter plugin reads and writes these files.
Add assertions that prove the agent actually reads from the temp directory: task_id must match between input and output, and when fix_proposed, changes must reference the input file.
Add `codescene.agentModel` dropdown setting with predefined models and `codescene.agentModelOverride` text field for custom model identifiers. The override takes precedence when non-blank.
Replace hardcoded opencode_config with user-configurable key-value pairs in VSCode settings. Keys come from a fixed list (amazon-bedrock:profile, amazon-bedrock:region) and render as an editable table in the settings UI.
Add a clickable link in the providerOptions setting description that runs cs-agent's check-connectivity skill to verify provider credentials are correctly configured.
The connectivity test was failing with "not inside a git repository" because `spawn()` wasn't given a `cwd`, inheriting whatever directory VS Code was running from.
The spawned cs-agent process was blocking because stdin remained open. Adding `stdio: ['ignore', 'pipe', 'pipe']` closes stdin and allows the process to complete. Also adds timeout (60s default), abort signal support, and makes the progress notification cancellable. Integration test requires `CS_AGENT_AWS_PROFILE` and `CS_AGENT_AWS_REGION` environment variables.
Extend `providerOptions` to support API key authentication for direct provider access alongside the existing Amazon Bedrock profile/region options. The `agentModel` dropdown now includes models from all four providers. Provider options map to cs-agent's expected JSON structure: - `anthropic:api-key` -> `anthropic_api_key` - `openai:api-key` -> `openai_api_key` - `google:api-key` -> `google_api_key`
Include `tracking.environment: "codescene-vscode"` in the agent config so analytics and MCP events identify this extension as the source.
Split the configuration building logic to distinguish between API key credentials (Anthropic, OpenAI, Google) and Bedrock-specific options (profile, region). The `buildProviderCredentials` function now handles API keys and populates the top-level `provider` field in `AgentConfig`, while `buildOpencodeConfig` remains focused solely on Bedrock configuration within `opencode_config.provider`.
Extract the complex conditional into a `hasAnyApiKey` helper function to reduce cyclomatic complexity and improve readability.
When using OpenAI or Anthropic API keys without an explicit model selection, the connectivity test was failing because it used the default Bedrock model. Now getAgentModel() checks providerOptions and returns a matching model (openai/gpt-4o or anthropic/claude-sonnet-4-0) when API keys are configured.
When agentModel had its package.json default value, the model inference from providerOptions was never reached. Now getAgentModel() checks if agentModel equals the default before returning it, allowing OpenAI or Anthropic keys to select the appropriate model automatically.
When the agent returns a proposed refactoring but the resulting code is identical to the original, override the response to `unable_to_fix` with low confidence. This prevents no-op refactorings from being presented as successful suggestions.
![codescene-access[bot]](https://avatars.githubusercontent.com/in/3755875?s=60&v=4){kind=small}
This comment was marked as outdated.
This comment was marked as outdated.
Sorry, something went wrong.
There was a problem hiding this comment.
Code Coverage Gates Failed
(required = 95%)
Review details by gate
- New & Changed Code Coverage
-
Pass/Fail Reason:
New or changed code lacks code coverage: 48.3% covered < threshold = 95% -
Action:
Add tests to:src/refactoring/agent-service.ts— uncovered line(s) = [16-19,28-34,37-74,77-89,92-110,113-122,125-133,136-164,198-199,268-275]src/refactoring/connectivity.ts— uncovered line(s) = [20-27,32-104]src/extension-impl.ts— uncovered line(s) = [335-336,338-364]src/refactoring/agent-config.ts— uncovered line(s) = [73-79,81-83]src/codescene-tab/webview/ace/cwf-webview-ace-panel.ts— uncovered line(s) = [57,101,235-241]
-
* chore: add updatecwf target for local CWF development Automatically build and copy CWF from ../cs-webview when running as vemv, otherwise download from GitHub releases. Make build depend on updatecwf to ensure CWF is always up to date. * chore: fix local builds with local CWF * fix: make updatecwf work on Windows Move the local-vs-release CWF logic from a POSIX shell recipe into scripts/update-cwf.js, since make on Windows runs recipes via cmd.exe. * fix: run fetch-github-release CLI directly The package's bin launcher passes a raw Windows path to import(), which Node rejects. Invoking dist/src/cli.js directly works on all platforms. --------- Co-authored-by: vemv <vemv@users.noreply.github.com>
This comment was marked as outdated.
This comment was marked as outdated.
Sorry, something went wrong.
# Conflicts: # .gitignore # CLAUDE.md # Makefile # package-lock.json # package.json # scripts/package-platform.js # scripts/update-cwf.js # src/code-health-monitor/addon.ts # src/codescene-tab/webview-panel.ts # src/configuration.ts # src/control-center/view-provider.ts # src/devtools-api/index.ts # src/extension-impl.ts # src/simple-executor.ts # src/test/setup.ts # src/test/suite/git-change-lister.test.ts
This comment was marked as outdated.
This comment was marked as outdated.
Sorry, something went wrong.
Preflight and fns-to-refactor were ACE-specific. With cs-agent every code smell is refactorable, so the function to refactor is now resolved locally from the review (or the whole file for file-level smells). Test Connectivity no longer requires the "Auth token" setting, and the Refactor action, CodeLens, status bar item and docs panel are shown only when Enable Auto Refactor is on and the user is authenticated by any route. cs-agent hands CS_ACCESS_TOKEN to the CodeScene MCP server and silently produces no output without it. When no token setting or legacy session exists, fetch one from `cs-ide auth token --client ide` as an interim until cs-ide exposes it over JSON-RPC. Missing output now surfaces agent stderr.
This comment was marked as outdated.
This comment was marked as outdated.
Sorry, something went wrong.
Replace the interim `cs-ide auth token` spawn with the cs-ide/auth-token request, so the server URL, active account and token refresh are handled by the running cs-ide instead of a separate process. Pin the CLI build that ships the request. The delta presentation tests now expect a single event, since the ACE enrichment pass that produced the second one is gone.
This comment was marked as outdated.
This comment was marked as outdated.
Sorry, something went wrong.
…ailing-newline drift The agent returns search/replace edits written with LF line endings and, for whole-method edits, a trailing newline. The function body taken from the editor keeps CRLF on Windows files and ends right after the closing brace. Neither form matched, so String.replace silently did nothing, the result equalled the original and a proposed fix was downgraded to "unable to fix" with an empty diff. Apply replacements on LF-normalised text, restore the original line endings afterwards, and retry without trailing newlines when the exact search is not found. Replacement text is now inserted literally so "$&" sequences are not interpreted. A search that still does not match logs a warning with the search and the body.
cs-agent reads the target file from disk, but the refactoring is applied to the function body taken from the editor buffer. With unsaved edits the two disagree, the agent's search strings no longer match and the user ends up on a "Refactoring Failed" panel that only points at the log. Prompt with a modal before the panel opens: "Save and refactor" saves the document and continues, dismissing it cancels without opening the panel. The Retry button goes through the same command. runRefactoring still rejects a dirty document as a backstop for other callers.
This comment was marked as outdated.
This comment was marked as outdated.
Sorry, something went wrong.
cs-agent performs refactorings now, so nothing calls the ACE refactor request. Remove the client method, its cancellation and HTTP-failure mapping, the response normalizer, log summaries and test fixtures. Reword the "Auth token" setting, README entry and the missing-sign-in error so signing in is presented as the normal route for refactoring.
There was a problem hiding this comment.
Gates Passed
See analysis details in CodeScene
Quality Gate Profile: Clean Code Collective
Install CodeScene MCP: safeguard and uplift AI-generated code. Catch issues early with our IDE extension and CLI tool.
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
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.
Commits:
feat: integrate cs-agent for refactoring
Replace the old ACE backend with the local cs-agent binary for
refactoring operations. This eliminates dependency on the ACE
backend service while keeping the existing CWF webview UI.
refactor: rename ACE to CodeScene Refactoring Agent
Aligns with the agentic-ace branch reference implementation:
(agent reads its own config)
to "CodeScene Agent" or "CodeScene Refactoring Agent"
fix: handle open-settings message in ACE acknowledgement panel
Clicking "Extension Settings" in the ACE acknowledgement panel was
logging an error because the panel didn't have a handler for the
open-settings message type. Reuse the existing handleOpenSettings
function from cwf-message-handlers.
fix: spawn refactoring agent with git repository root as cwd
The agent was being spawned with a temp directory as cwd, causing
"not inside a git repository" errors. Use the git repo root so the
agent can access the entire project and run git commands.
fix: add Amazon Bedrock provider config for refactoring agent
The agent was failing with "UnknownError" because the opencode_config
with Amazon Bedrock provider settings was missing from the agent config.
Also adds the --model argument to explicitly request the Bedrock model.
fix: handle whole_file change type in agent output
The agent outputs whole_file_content instead of replacements when
change_type is "whole_file". Also fixes the integration test to
initialize a git repo and adds assertions verifying the agent
writes the expected output file.
chore: remove debug log statements
feat: forward agent stdout to CWF progress widget
Stream progress messages from the refactoring agent binary to the
loading view instead of showing hardcoded rotating messages. The
most recent stdout line is displayed, with newer lines replacing
older ones.
chore: gitignore bin/cs-agent
chore: add updatecwf target for local CWF development
Automatically build and copy CWF from ../cs-webview when running
as vemv, otherwise download from GitHub releases. Make build
depend on updatecwf to ensure CWF is always up to date.
chore: rename refactoring-agent directory to cs-agent
Update Makefile package targets to reference ~/cs-agent instead of
~/refactoring-agent for darwin arm64 and x64 builds.
chore: fix local builds with local CWF
refactor: reduce complexity in AgentRefactoringService
Extract spawnAgentProcess, setupStderrHandler, setupStdoutHandler from
invokeAgent to reduce cyclomatic complexity. Refactor applyChanges to
use filter/reduce pattern to eliminate nested conditionals. Fix
post-refactoring tests by initializing git repo and mocking extension path.
fix: clean up test-git-repo-integration dir reliably
Add global mocha afterAll hook to clean up test directories after all
tests complete. This catches async operations that recreate directories
after per-suite teardowns run.
test: update agent output fixtures to use enum values
Match the new cs-agent schema that uses fixed enum values for
summary and reasoning instead of freeform strings.
chore: expand CLAUDE.md
feat: use temp directory for agent JSON I/O files
Move render-code-fix input/output JSON files to a temp directory
instead of the repository root to avoid cluttering user projects.
The agent's io_json_dir config option specifies where the reporter
plugin reads and writes these files.
test: verify agent reads input from io_json_dir
Add assertions that prove the agent actually reads from the temp
directory: task_id must match between input and output, and when
fix_proposed, changes must reference the input file.
feat: add settings UI for agent model selection
Add
codescene.agentModeldropdown setting with predefined models andcodescene.agentModelOverridetext field for custom model identifiers.The override takes precedence when non-blank.
chore: expand CLAUDE.md
feat: add providerOptions setting for agent configuration
Replace hardcoded opencode_config with user-configurable key-value
pairs in VSCode settings. Keys come from a fixed list (amazon-bedrock:profile,
amazon-bedrock:region) and render as an editable table in the settings UI.
feat: expand Anthropic models
feat: add Test connectivity button in settings
Add a clickable link in the providerOptions setting description that
runs cs-agent's check-connectivity skill to verify provider credentials
are correctly configured.
fix: pass workspace folder as cwd in connectivity test
The connectivity test was failing with "not inside a git repository"
because
spawn()wasn't given acwd, inheriting whatever directoryVS Code was running from.
chore: expand CLAUDE.md
fix: prevent connectivity test from hanging indefinitely
The spawned cs-agent process was blocking because stdin remained open.
Adding
stdio: ['ignore', 'pipe', 'pipe']closes stdin and allows theprocess to complete. Also adds timeout (60s default), abort signal
support, and makes the progress notification cancellable.
Integration test requires
CS_AGENT_AWS_PROFILEandCS_AGENT_AWS_REGIONenvironment variables.
feat: add OpenAI, Anthropic, and Google provider auth options
Extend
providerOptionsto support API key authentication for directprovider access alongside the existing Amazon Bedrock profile/region
options. The
agentModeldropdown now includes models from all fourproviders.
Provider options map to cs-agent's expected JSON structure:
anthropic:api-key->anthropic_api_keyopenai:api-key->openai_api_keygoogle:api-key->google_api_keyfeat: add tracking config to cs-agent environment
Include
tracking.environment: "codescene-vscode"in the agent configso analytics and MCP events identify this extension as the source.
refactor: separate provider credentials from bedrock config
Split the configuration building logic to distinguish between API key
credentials (Anthropic, OpenAI, Google) and Bedrock-specific options
(profile, region). The
buildProviderCredentialsfunction now handlesAPI keys and populates the top-level
providerfield inAgentConfig, whilebuildOpencodeConfigremains focused solely onBedrock configuration within
opencode_config.provider.refactor: simplify conditional in buildProviderCredentials
Extract the complex conditional into a
hasAnyApiKeyhelper function toreduce cyclomatic complexity and improve readability.
fix: infer agent model from provider credentials
When using OpenAI or Anthropic API keys without an explicit model
selection, the connectivity test was failing because it used the
default Bedrock model. Now getAgentModel() checks providerOptions
and returns a matching model (openai/gpt-4o or anthropic/claude-sonnet-4-0)
when API keys are configured.
fix: infer model from provider keys even with default agentModel
When agentModel had its package.json default value, the model inference
from providerOptions was never reached. Now getAgentModel() checks if
agentModel equals the default before returning it, allowing OpenAI or
Anthropic keys to select the appropriate model automatically.
fix: treat content-identical refactorings as failures
When the agent returns a proposed refactoring but the resulting code is
identical to the original, override the response to
unable_to_fixwithlow confidence. This prevents no-op refactorings from being presented
as successful suggestions.