[codemode]: fixing string escape corruption, enable top-level control flow in starlark, refining the prompt of executecode tool#2206
Merged
Pratham-Mishra04 merged 1 commit intov1.5.0from Mar 24, 2026
Merged
Conversation
|
|
Contributor
|
Caution Review failedThe pull request is closed. ℹ️ Recent review info⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Run ID: 📒 Files selected for processing (7)
📝 WalkthroughSummary by CodeRabbit
WalkthroughSwitched Starlark execution to a Python-like dialect via ExecFileOptions/syntax.FileOptions, preserved raw string escapes, and replaced ad-hoc tool-name normalization with canonical-name resolution and compatibility aliases across execution, listing, docs, and tests. Changes
Estimated code review effort🎯 4 (Complex) | ⏱️ ~45 minutes Poem
🚥 Pre-merge checks | ✅ 2 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (2 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Comment |
Contributor
Author
This stack of pull requests is managed by Graphite. Learn more about stacking. |
Contributor
🧪 Test Suite AvailableThis PR can be tested by a repository admin. |
Madhuvod
changed the title
Madhuvod
force-pushed
the
03-22-_codemode_fix_string_escape_corruption_enable_top-level_control_flow_refining_the_prompt_of_executecode_tool
branch
from
bf7e9dc to
6af0f38
Compare
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Actionable comments posted: 1
🧹 Nitpick comments (1)
core/mcp/codemode/starlark/starlark_test.go (1)
495-505: Consider extracting shared dialect config to avoid sync drift.The
starlarkOpts()helper duplicates the dialect configuration fromexecutecode.go. While the comment acknowledges this, any changes to the production config could silently cause test/production divergence.Consider exporting the options from a shared location or adding a test that verifies the configurations match.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@core/mcp/codemode/starlark/starlark_test.go` around lines 495 - 505, Extract the shared dialect config into a single exported helper and make the test use it (or add a test that compares them): move the FileOptions literal from starlark_test.go into an exported function/variable in the production code (the same place in executecode.go that currently defines the dialect) and have starlarkOpts() delegate to that exported helper, or alternatively add a unit test that compares starlarkOpts() to the production options in executecode.go to fail if they diverge; reference the existing starlarkOpts() helper and the dialect options in executecode.go when making the change.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@core/mcp/codemode/starlark/utils.go`:
- Around line 194-200: The hint example uses the unsupported "is" operator;
update the appended hint lines in the block that checks errorMessage (the
contains checks for "got try"/"got except"/"got finally"/"got raise") to use ==
for comparisons (e.g., change "if result is None..." to "if result == None..."
or another supported equality form) and ensure the example consistently avoids
the "is" operator in any appended hints added to the hints slice; modify the
strings passed to hints = append(...) accordingly so the example matches the
executor's "NO is operator" rule.
---
Nitpick comments:
In `@core/mcp/codemode/starlark/starlark_test.go`:
- Around line 495-505: Extract the shared dialect config into a single exported
helper and make the test use it (or add a test that compares them): move the
FileOptions literal from starlark_test.go into an exported function/variable in
the production code (the same place in executecode.go that currently defines the
dialect) and have starlarkOpts() delegate to that exported helper, or
alternatively add a unit test that compares starlarkOpts() to the production
options in executecode.go to fail if they diverge; reference the existing
starlarkOpts() helper and the dialect options in executecode.go when making the
change.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro
Run ID: 72980382-61ee-4816-8d9b-36889bba4d6d
📒 Files selected for processing (3)
core/mcp/codemode/starlark/executecode.gocore/mcp/codemode/starlark/starlark_test.gocore/mcp/codemode/starlark/utils.go
Madhuvod
force-pushed
the
03-22-_codemode_fix_string_escape_corruption_enable_top-level_control_flow_refining_the_prompt_of_executecode_tool
branch
from
6af0f38 to
02d8034
Compare
Contributor
There was a problem hiding this comment.
Actionable comments posted: 1
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
core/mcp/codemode/starlark/executecode.go (1)
313-338:⚠️ Potential issue | 🔴 CriticalWire deadline into thread cancellation;
thread.SetLocal()does not auto-stop on context timeout.
thread.SetLocal("context", timeoutCtx)only stores the context in thread-local storage—it does not interrupt execution when the deadline expires. WithWhile,TopLevelControl, andRecursionenabled,while True: passor infinite recursion can block the request indefinitely past the intended timeout.Use
thread.Cancel()orthread.SetMaxExecutionSteps()to enforce execution limits. Spawn a cancellation goroutine that waits onctx.Done()and callsthread.Cancel(), and ensure nested tool calls inherit the same timeout bound.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@core/mcp/codemode/starlark/executecode.go` around lines 313 - 338, The thread currently stores timeoutCtx via thread.SetLocal("context", timeoutCtx) but that doesn't stop execution on deadline; spawn a goroutine after creating timeoutCtx that waits for timeoutCtx.Done() (or ctx.Done()) and then calls thread.Cancel(), and additionally set a conservative execution step limit via thread.SetMaxExecutionSteps(...) to guard against busy loops; ensure any nested tool invocations read the same timeoutCtx from thread locals so they inherit the bound; apply these changes around the thread creation (symbols: timeoutCtx, cancel, thread, thread.SetLocal, thread.Cancel, thread.SetMaxExecutionSteps, starlark.ExecFileOptions).
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@core/mcp/codemode/starlark/executecode.go`:
- Around line 57-59: Update the user-facing description string for the Python
(Starlark) "executeToolCode" parameter so it no longer instructs "ALWAYS retry
if code fails"; instead state that retries should only be attempted for
read-only or idempotent operations or after the model has inspected earlier tool
calls to confirm no stateful side effects occurred. Locate the description text
in the executeToolCode schema (the "description" value shown in the diff) and
change the guidance to: recommend cautious retrying limited to
idempotent/read-only flows and advise the model to check whether any prior tool
calls mutated state before rerunning the block.
---
Outside diff comments:
In `@core/mcp/codemode/starlark/executecode.go`:
- Around line 313-338: The thread currently stores timeoutCtx via
thread.SetLocal("context", timeoutCtx) but that doesn't stop execution on
deadline; spawn a goroutine after creating timeoutCtx that waits for
timeoutCtx.Done() (or ctx.Done()) and then calls thread.Cancel(), and
additionally set a conservative execution step limit via
thread.SetMaxExecutionSteps(...) to guard against busy loops; ensure any nested
tool invocations read the same timeoutCtx from thread locals so they inherit the
bound; apply these changes around the thread creation (symbols: timeoutCtx,
cancel, thread, thread.SetLocal, thread.Cancel, thread.SetMaxExecutionSteps,
starlark.ExecFileOptions).
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro
Run ID: 69ef1c8c-ecf4-47d9-8bdc-d5003a010ceb
📒 Files selected for processing (4)
.gitignorecore/mcp/codemode/starlark/executecode.gocore/mcp/codemode/starlark/starlark_test.gocore/mcp/codemode/starlark/utils.go
✅ Files skipped from review due to trivial changes (1)
- .gitignore
🚧 Files skipped from review as they are similar to previous changes (2)
- core/mcp/codemode/starlark/starlark_test.go
- core/mcp/codemode/starlark/utils.go
Madhuvod
force-pushed
the
03-22-_codemode_fix_string_escape_corruption_enable_top-level_control_flow_refining_the_prompt_of_executecode_tool
branch
from
02d8034 to
4a6f8f3
Compare
Madhuvod
force-pushed
the
03-22-_codemode_fix_string_escape_corruption_enable_top-level_control_flow_refining_the_prompt_of_executecode_tool
branch
from
4a6f8f3 to
7aa8bd7
Compare
Madhuvod
force-pushed
the
03-22-_codemode_fix_string_escape_corruption_enable_top-level_control_flow_refining_the_prompt_of_executecode_tool
branch
from
7aa8bd7 to
331cd83
Compare
Madhuvod
changed the base branch from
graphite-base/2206
to
03-23-fix_send_back_accumulated_usage_in_mcp_agent_mode
18 tasks
Contributor
There was a problem hiding this comment.
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (4)
core/mcp/codemode/starlark/getdocs.go (1)
143-164:⚠️ Potential issue | 🟡 MinorgetToolDocs still tells the model top-level loops are unsupported.
generateTypeDefinitionsstill says top-levelfor/if/whilemust be wrapped in a function, butexecuteCodenow enablessyntax.FileOptions.TopLevelControland the new dialect tests assert those constructs work at file scope. Leaving this text stale will steer the model away from the feature this PR adds.📝 Suggested doc fix
- sb.WriteString("# STARLARK DIFFERENCE FROM PYTHON:\n") - sb.WriteString("# for/if/while at top level MUST be inside a function.\n") - sb.WriteString("# Wrap loops: def main(): for x in items: ... then result = main()\n") + sb.WriteString("# STARLARK DIFFERENCE FROM PYTHON:\n") + sb.WriteString("# Top-level if/for/while are supported.\n") + sb.WriteString("# Wrap logic in a function only when you want reuse or local scoping.\n")🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@core/mcp/codemode/starlark/getdocs.go` around lines 143 - 164, The generated header in getToolDocs (getdocs.go) is stale: update the STARLARK DIFFERENCE FROM PYTHON text emitted by generateTypeDefinitions/getToolDocs to remove the line "for/if/while at top level MUST be inside a function." and instead reflect that top-level control flow is supported when enabled (see executeCode enabling syntax.FileOptions.TopLevelControl); adjust the wording to mention that top-level for/if/while is allowed when TopLevelControl is enabled so models are not discouraged from using the new feature.core/mcp/codemode/starlark/executecode.go (1)
257-305:⚠️ Potential issue | 🟠 MajorDetect canonical-name collisions before populating
structMembers.
getCanonicalToolNamenow collapses prefix/case differences, butstructMembers[parsedToolName] = toolFuncwill overwrite any earlier tool that normalizes to the same exported name, and the alias branch only avoids one side of the collision. For runtime-provided MCP servers, that makes one tool silently disappear and leaves the canonical.pyi/ docs surface ambiguous.Please fail fast or disambiguate when two tools map to the same canonical or compatibility identifier.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@core/mcp/codemode/starlark/executecode.go` around lines 257 - 305, Before inserting a tool into structMembers, detect name collisions for parsedToolName and compatibilityAlias and fail fast or disambiguate: use getCanonicalToolName and getCompatibilityToolAlias to compute parsedToolName and compatibilityAlias, then check structMembers for an existing entry under parsedToolName (and under compatibilityAlias if isValidStarlarkIdentifier(compatibilityAlias)) and if an entry exists whose underlying originalToolName differs from the current originalToolName, return an error (or otherwise surface the collision) instead of overwriting; only when no conflict is found assign structMembers[parsedToolName] = toolFunc and then add the compatibility alias if it doesn’t conflict.core/mcp/agent.go (1)
211-250:⚠️ Potential issue | 🟠 MajorShare one code-mode tool-name canonicalizer between execution and auto-exec validation.
executeToolCode,listToolFiles, andreadToolFilenow expose canonical names fromgetCanonicalToolName(...), but agent-mode validation still compares extractedserver.tool()calls againstbuildAllowedAutoExecutionTools()'s legacyparseToolName(strings.ReplaceAll(...))transform. For prefixed/case-folded names, the model can call the exact identifier we document and still get classified as non-auto-executable.Please move the code-mode identifier normalization into one shared helper and use it for both binding/docs and
allowedAutoExecutionToolsvalidation.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@core/mcp/agent.go` around lines 211 - 250, The validation mismatch is caused by different canonicalization paths: getCanonicalToolName is used by executeToolCode, listToolFiles, and readToolFile while buildAllowedAutoExecutionTools still uses parseToolName, causing isToolCallAllowedForCodeMode to misclassify canonical names; create a single helper (e.g., canonicalizeCodeModeToolName) that wraps getCanonicalToolName semantics and replace calls to parseToolName in buildAllowedAutoExecutionTools as well as normalize extracted tool calls before calling isToolCallAllowedForCodeMode (normalize both extractedToolCall.serverName/extractedToolCall.toolName and the allowedAutoExecutionTools entries so comparisons use the shared canonical form used by executeToolCode, listToolFiles, readToolFile and getCanonicalToolName).core/mcp/codemode/starlark/readfile.go (1)
140-149:⚠️ Potential issue | 🟡 MinorKeep the error suggestions in the same
servers/...format we advertise.After these description changes,
listToolFilesandfileNameDescriptiontell the model to use paths likeservers/calculator/add.pyi, but the tool-not-found and server-not-found branches still suggestcalculator/add.pyiandcalculator.pyi. That nudges retries toward an invalid filename.🛠 Suggested fix
- errorMsg += fmt.Sprintf(" - %s/%s.pyi\n", clientName, t) + errorMsg += fmt.Sprintf(" - servers/%s/%s.pyi\n", clientName, t) ... - availableFiles = append(availableFiles, fmt.Sprintf("%s.pyi", name)) + availableFiles = append(availableFiles, fmt.Sprintf("servers/%s.pyi", name)) ... - availableFiles = append(availableFiles, fmt.Sprintf("%s/%s.pyi", name, getCanonicalToolName(name, tool.Function.Name))) + availableFiles = append(availableFiles, fmt.Sprintf("servers/%s/%s.pyi", name, getCanonicalToolName(name, tool.Function.Name)))Also applies to: 167-176
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@core/mcp/codemode/starlark/readfile.go` around lines 140 - 149, The error messages are suggesting filenames without the advertised servers/... prefix; update the construction of the tool-not-found and server-not-found error strings to use the canonical "servers/<server>/<file>.pyi" format. Specifically, in the block that builds errorMsg when foundTool == nil (using variables foundTool, availableTools, tools, tool.Function.Name, getCanonicalToolName) change the appended lines to include the full path like "servers/%s/%s.pyi" for each available tool; also update the analogous server-not-found branch (the other block around lines 167-176) to suggest "servers/%s.pyi" for the server-level hint so both suggestions match listToolFiles/fileNameDescription conventions.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Outside diff comments:
In `@core/mcp/agent.go`:
- Around line 211-250: The validation mismatch is caused by different
canonicalization paths: getCanonicalToolName is used by executeToolCode,
listToolFiles, and readToolFile while buildAllowedAutoExecutionTools still uses
parseToolName, causing isToolCallAllowedForCodeMode to misclassify canonical
names; create a single helper (e.g., canonicalizeCodeModeToolName) that wraps
getCanonicalToolName semantics and replace calls to parseToolName in
buildAllowedAutoExecutionTools as well as normalize extracted tool calls before
calling isToolCallAllowedForCodeMode (normalize both
extractedToolCall.serverName/extractedToolCall.toolName and the
allowedAutoExecutionTools entries so comparisons use the shared canonical form
used by executeToolCode, listToolFiles, readToolFile and getCanonicalToolName).
In `@core/mcp/codemode/starlark/executecode.go`:
- Around line 257-305: Before inserting a tool into structMembers, detect name
collisions for parsedToolName and compatibilityAlias and fail fast or
disambiguate: use getCanonicalToolName and getCompatibilityToolAlias to compute
parsedToolName and compatibilityAlias, then check structMembers for an existing
entry under parsedToolName (and under compatibilityAlias if
isValidStarlarkIdentifier(compatibilityAlias)) and if an entry exists whose
underlying originalToolName differs from the current originalToolName, return an
error (or otherwise surface the collision) instead of overwriting; only when no
conflict is found assign structMembers[parsedToolName] = toolFunc and then add
the compatibility alias if it doesn’t conflict.
In `@core/mcp/codemode/starlark/getdocs.go`:
- Around line 143-164: The generated header in getToolDocs (getdocs.go) is
stale: update the STARLARK DIFFERENCE FROM PYTHON text emitted by
generateTypeDefinitions/getToolDocs to remove the line "for/if/while at top
level MUST be inside a function." and instead reflect that top-level control
flow is supported when enabled (see executeCode enabling
syntax.FileOptions.TopLevelControl); adjust the wording to mention that
top-level for/if/while is allowed when TopLevelControl is enabled so models are
not discouraged from using the new feature.
In `@core/mcp/codemode/starlark/readfile.go`:
- Around line 140-149: The error messages are suggesting filenames without the
advertised servers/... prefix; update the construction of the tool-not-found and
server-not-found error strings to use the canonical
"servers/<server>/<file>.pyi" format. Specifically, in the block that builds
errorMsg when foundTool == nil (using variables foundTool, availableTools,
tools, tool.Function.Name, getCanonicalToolName) change the appended lines to
include the full path like "servers/%s/%s.pyi" for each available tool; also
update the analogous server-not-found branch (the other block around lines
167-176) to suggest "servers/%s.pyi" for the server-level hint so both
suggestions match listToolFiles/fileNameDescription conventions.
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro
Run ID: 89a73908-e18a-4e59-9323-7bacb394d6dc
📒 Files selected for processing (8)
.gitignorecore/mcp/agent.gocore/mcp/codemode/starlark/executecode.gocore/mcp/codemode/starlark/getdocs.gocore/mcp/codemode/starlark/listfiles.gocore/mcp/codemode/starlark/readfile.gocore/mcp/codemode/starlark/starlark_test.gocore/mcp/codemode/starlark/utils.go
✅ Files skipped from review due to trivial changes (1)
- .gitignore
🚧 Files skipped from review as they are similar to previous changes (1)
- core/mcp/codemode/starlark/utils.go
Madhuvod
force-pushed
the
03-22-_codemode_fix_string_escape_corruption_enable_top-level_control_flow_refining_the_prompt_of_executecode_tool
branch
2 times, most recently
from
e5dc96a to
8ff1fdd
Compare
Contributor
There was a problem hiding this comment.
Actionable comments posted: 4
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
core/mcp/codemode/starlark/executecode.go (1)
323-347:⚠️ Potential issue | 🟠 MajorAdd a goroutine to actually cancel the Starlark thread when the timeout deadline expires.
thread.SetLocal("context", timeoutCtx)only stores the context in thread-local memory—it does not interruptExecFileOptionswhen the deadline expires. According to starlark-go's design, timeout enforcement requires either a separate goroutine that callsthread.Cancel()when the context deadline is reached, or callingthread.SetMaxExecutionSteps()to bound execution. WithWhileandRecursionenabled, pure Starlark loops likewhile True: passcan now run indefinitely and pin the request until something external kills it. Wire the timeout into the actual thread cancellation mechanism before enabling these features.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@core/mcp/codemode/starlark/executecode.go` around lines 323 - 347, The thread-local context alone doesn't stop starlark.ExecFileOptions; add cancellation logic that calls thread.Cancel() when the timeoutCtx is done (or set a max-step bound via thread.SetMaxExecutionSteps()) so long-running constructs (enabled by FileOptions.While and Recursion) can't hang the request; specifically, after creating timeoutCtx and thread and after thread.SetLocal("context", timeoutCtx) spawn a goroutine that waits on timeoutCtx.Done() and calls thread.Cancel(), keeping the existing cancel() defer, so timeoutCtx from s.getToolExecutionTimeout() actually interrupts the starlark.Thread used by starlark.ExecFileOptions.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@core/mcp/codemode/starlark/readfile.go`:
- Around line 24-40: The fallback virtual path suggestions in readfile.go
currently return examples like "client.pyi" / "client/tool.pyi" that conflict
with your documented VFS format; update the string values used for
fileNameDescription and any fallback messages in the same file (including the
other occurrence around the second block using
fileNameDescription/toolDescription) to prefix those suggestions with "servers/"
(e.g., "servers/<serverName>.pyi" and "servers/<serverName>/<toolName>.pyi") so
the documented format, list output, and error suggestions all use the same
virtual path convention.
In `@core/mcp/codemode/starlark/starlark_test.go`:
- Around line 217-224: The map literal for clients is assigning a value of type
*schemas.MCPClientState but uses a plain struct literal; wrap the value with a
pointer constructor so the map stores &schemas.MCPClientState{...} instead of
schemas.MCPClientState{...}; update the entry where clientName is the key and
the embedded ExecutionConfig uses &schemas.MCPClientConfig{...} as needed to
match pointer types (symbols: clients, MCPClientState, MCPClientConfig,
clientName).
In `@core/mcp/codemode/starlark/utils.go`:
- Around line 360-391: Tool canonicalization (getCanonicalToolName /
stripClientPrefix) can collapse distinct tool names into identical callable
identifiers and cause overwrites in executeCode; detect these collisions per
client and fail or disambiguate before exposing bindings. In the code path that
gathers tools for a client (call site: executeCode), build a map from canonical
:= getCanonicalToolName(clientName, originalToolName) to the list of
originalToolName values, check for duplicates, and either return an error
listing the conflicting original names or deterministically rename collisions
(e.g. append a stable suffix derived from the originalToolName or a short hash)
and update any compatibility aliases (getCompatibilityToolAlias) accordingly so
bindings and .pyi output remain unique. Ensure matchesToolReference semantics
are preserved by mapping lookup to the disambiguated canonical identifiers.
- Around line 201-206: The current greedy regex in the hinting block
misidentifies the undefined symbol; update the parsing to try multiple, more
specific regexes in order: first match patterns like `name 'foo' is not defined`
(capture inside quotes), then match Starlark-style `...: undefined: <ident>`
(capture the identifier after `undefined:`), and only then fall back to a
generic word capture; use the first successful capture to set undefinedVar and
append the same hints. Apply this change in the same block where `re :=
regexp.MustCompile(\`(\w+).*(?:undefined|not defined)\`)` is created (the
undefined-variable handling code that builds `hints` and `undefinedVar`).
---
Outside diff comments:
In `@core/mcp/codemode/starlark/executecode.go`:
- Around line 323-347: The thread-local context alone doesn't stop
starlark.ExecFileOptions; add cancellation logic that calls thread.Cancel() when
the timeoutCtx is done (or set a max-step bound via
thread.SetMaxExecutionSteps()) so long-running constructs (enabled by
FileOptions.While and Recursion) can't hang the request; specifically, after
creating timeoutCtx and thread and after thread.SetLocal("context", timeoutCtx)
spawn a goroutine that waits on timeoutCtx.Done() and calls thread.Cancel(),
keeping the existing cancel() defer, so timeoutCtx from
s.getToolExecutionTimeout() actually interrupts the starlark.Thread used by
starlark.ExecFileOptions.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro
Run ID: e153641d-9eb0-47ed-bd01-b2eeeb233352
📒 Files selected for processing (7)
core/mcp/agent.gocore/mcp/codemode/starlark/executecode.gocore/mcp/codemode/starlark/getdocs.gocore/mcp/codemode/starlark/listfiles.gocore/mcp/codemode/starlark/readfile.gocore/mcp/codemode/starlark/starlark_test.gocore/mcp/codemode/starlark/utils.go
✅ Files skipped from review due to trivial changes (1)
- core/mcp/codemode/starlark/getdocs.go
🚧 Files skipped from review as they are similar to previous changes (2)
- core/mcp/agent.go
- core/mcp/codemode/starlark/listfiles.go
Madhuvod
force-pushed
the
03-22-_codemode_fix_string_escape_corruption_enable_top-level_control_flow_refining_the_prompt_of_executecode_tool
branch
from
8ff1fdd to
631fb8a
Compare
Contributor
There was a problem hiding this comment.
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (3)
core/mcp/codemode/starlark/executecode.go (2)
347-357:⚠️ Potential issue | 🟡 MinorClassify parser failures as syntax/compile errors.
The new
got try/got except/got raisecases fail before execution starts, but this branch still reportsruntimeunless the message literally containssyntax error. That makeserrors.kindinconsistent with the new hints and with the existingExecutionErrorTypeCompile/ExecutionErrorTypeSyntaxmodel.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@core/mcp/codemode/starlark/executecode.go` around lines 347 - 357, The error handling after starlark.ExecFileOptions currently defaults to ExecutionErrorTypeRuntime and only sets ExecutionErrorTypeSyntax when the message literally contains "syntax error"; update this block in executecode.go (around starlark.ExecFileOptions, errorMessage, generatePythonErrorHints) to detect parser/compile failures (e.g., messages containing parser tokens like "got try", "got except", "got raise" or other starlark parser indicators) and set errorKind to ExecutionErrorTypeCompile or ExecutionErrorTypeSyntax as appropriate instead of Runtime so errors.kind matches generatePythonErrorHints and the compile/syntax model.
322-347:⚠️ Potential issue | 🔴 CriticalApply the deadline to the interpreter itself.
timeoutCtxis only stored on the thread; nothing in this path checks it duringExecFileOptions, and the builtin still callscallMCPToolwith the outerctx. WithTopLevelControl,While, andRecursionnow enabled,while True:or runaway recursion can pin the request goroutine indefinitely without ever timing out.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@core/mcp/codemode/starlark/executecode.go` around lines 322 - 347, The interpreter deadline isn't applied during execution: use the timeoutCtx produced by getToolExecutionTimeout()/context.WithTimeout instead of the outer ctx when running and when invoking builtins so the interpreter can be cancelled; specifically, set the thread local value (thread.SetLocal("context", timeoutCtx)) and pass timeoutCtx into starlark.ExecFileOptions (replace the outer ctx usage) and ensure the builtin that calls callMCPTool uses timeoutCtx (or derives its ctx from the thread local) rather than the outer ctx so top-level while/recursion can be preempted by the timeout.core/mcp/codemode/starlark/utils.go (1)
201-226:⚠️ Potential issue | 🟡 MinorUpdate the stale “not within a function” hint.
This helper now handles the new dialect errors, but the later branch still tells users to wrap top-level
for/if/whilein a function.executeCodenow enables top-level control flow, so that guidance can point users at the wrong fix.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@core/mcp/codemode/starlark/utils.go` around lines 201 - 226, The "not within a function" branch (the strings.Contains(errorMessage, "not within a function") case) must be updated: remove the advice to wrap top-level for/if/while in a function and instead append hints that top-level control flow is supported, suggest checking for dialect/syntax issues or wrong indentation, and provide a short valid example of top-level usage (e.g., a top-level for loop calling server.get and accumulating results) and how to use serverKeys if relevant; update the hints slice (hints) and reference serverKeys and errorMessage when composing the new messages so users see correct guidance for the current executeCode behavior.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Outside diff comments:
In `@core/mcp/codemode/starlark/executecode.go`:
- Around line 347-357: The error handling after starlark.ExecFileOptions
currently defaults to ExecutionErrorTypeRuntime and only sets
ExecutionErrorTypeSyntax when the message literally contains "syntax error";
update this block in executecode.go (around starlark.ExecFileOptions,
errorMessage, generatePythonErrorHints) to detect parser/compile failures (e.g.,
messages containing parser tokens like "got try", "got except", "got raise" or
other starlark parser indicators) and set errorKind to ExecutionErrorTypeCompile
or ExecutionErrorTypeSyntax as appropriate instead of Runtime so errors.kind
matches generatePythonErrorHints and the compile/syntax model.
- Around line 322-347: The interpreter deadline isn't applied during execution:
use the timeoutCtx produced by getToolExecutionTimeout()/context.WithTimeout
instead of the outer ctx when running and when invoking builtins so the
interpreter can be cancelled; specifically, set the thread local value
(thread.SetLocal("context", timeoutCtx)) and pass timeoutCtx into
starlark.ExecFileOptions (replace the outer ctx usage) and ensure the builtin
that calls callMCPTool uses timeoutCtx (or derives its ctx from the thread
local) rather than the outer ctx so top-level while/recursion can be preempted
by the timeout.
In `@core/mcp/codemode/starlark/utils.go`:
- Around line 201-226: The "not within a function" branch (the
strings.Contains(errorMessage, "not within a function") case) must be updated:
remove the advice to wrap top-level for/if/while in a function and instead
append hints that top-level control flow is supported, suggest checking for
dialect/syntax issues or wrong indentation, and provide a short valid example of
top-level usage (e.g., a top-level for loop calling server.get and accumulating
results) and how to use serverKeys if relevant; update the hints slice (hints)
and reference serverKeys and errorMessage when composing the new messages so
users see correct guidance for the current executeCode behavior.
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro
Run ID: 25b5a212-2ea7-4aed-8d03-f532da3f2bdf
📒 Files selected for processing (7)
core/mcp/agent.gocore/mcp/codemode/starlark/executecode.gocore/mcp/codemode/starlark/getdocs.gocore/mcp/codemode/starlark/listfiles.gocore/mcp/codemode/starlark/readfile.gocore/mcp/codemode/starlark/starlark_test.gocore/mcp/codemode/starlark/utils.go
🚧 Files skipped from review as they are similar to previous changes (3)
- core/mcp/agent.go
- core/mcp/codemode/starlark/readfile.go
- core/mcp/codemode/starlark/getdocs.go
akshaydeo
approved these changes
Mar 24, 2026
Collaborator
Merge activity
|
Pratham-Mishra04
changed the base branch from
03-23-fix_send_back_accumulated_usage_in_mcp_agent_mode
to
graphite-base/2206
…ow, refining the prompt of executecode tool
Pratham-Mishra04
force-pushed
the
03-22-_codemode_fix_string_escape_corruption_enable_top-level_control_flow_refining_the_prompt_of_executecode_tool
branch
from
631fb8a to
f34d0a4
Compare
Pratham-Mishra04
deleted the
03-22-_codemode_fix_string_escape_corruption_enable_top-level_control_flow_refining_the_prompt_of_executecode_tool
branch
akshaydeo
added a commit
that referenced
this pull request
Apr 18, 2026
* refactor: standardize empty array conventions for VK Provider & MCP Configs, and makes Provider Config weight optional for routing (#1932)
## Summary
Changes Virtual Key provider and MCP configurations from "allow-all by default" to "deny-by-default" security model. Virtual Keys now require explicit provider and MCP client configurations to allow access, improving security posture.
## Changes
- **Provider Configs**: Empty `provider_configs` now blocks all providers instead of allowing all
- **MCP Configs**: Empty `mcp_configs` now blocks all MCP tools instead of allowing all
- **Weight Field**: Changed provider `weight` from required `float64` to optional `*float64` - null weight excludes provider from weighted routing
- **Migration**: Added automatic backfill migration to preserve existing Virtual Key behavior by adding all available providers/MCP clients to VKs with empty configs
- **Documentation**: Updated all references to reflect new deny-by-default behavior
- **UI Updates**: Modified Virtual Key creation/editing interface to reflect new behavior and weight handling
## Type of change
- [x] Feature
- [x] Refactor
- [x] Documentation
## Affected areas
- [x] Core (Go)
- [x] Transports (HTTP)
- [x] Plugins
- [x] UI (Next.js)
- [x] Docs
## How to test
Test Virtual Key creation and provider/MCP access:
```sh
# Core/Transports
go version
go test ./...
# Test Virtual Key with no provider configs blocks requests
curl -X POST http://localhost:8080/v1/chat/completions \
-H "Authorization: Bearer sk-bf-empty-vk" \
-H "Content-Type: application/json" \
-d '{"model": "gpt-4", "messages": [{"role": "user", "content": "test"}]}'
# Should return error about no providers configured
# Test Virtual Key with provider configs allows requests
curl -X POST http://localhost:8080/v1/chat/completions \
-H "Authorization: Bearer sk-bf-configured-vk" \
-H "Content-Type: application/json" \
-d '{"model": "gpt-4", "messages": [{"role": "user", "content": "test"}]}'
# Should work normally
# UI
cd ui
pnpm i || npm i
pnpm test || npm test
pnpm build || npm run build
```
## Breaking changes
- [x] Yes
**Impact**: Existing Virtual Keys with empty `provider_configs` or `mcp_configs` would be blocked after this change.
**Migration**: Automatic migration `migrationBackfillEmptyVirtualKeyConfigs` runs on startup to backfill existing Virtual Keys with all available providers/MCP clients, preserving current behavior. New Virtual Keys created after this change will use deny-by-default.
## Security considerations
This change significantly improves security posture by requiring explicit configuration of allowed providers and MCP tools for Virtual Keys. The automatic migration ensures no disruption to existing deployments while new Virtual Keys benefit from the more secure default behavior.
## Checklist
- [x] I read `docs/contributing/README.md` and followed the guidelines
- [x] I added/updated tests where appropriate
- [x] I updated documentation where needed
- [x] I verified builds succeed (Go and UI)
- [x] I verified the CI pipeline passes locally if applicable
* feat: add MCP auto tool injection toggle (#1933)
## Summary
Adds a new configuration option `DisableAutoToolInject` to the MCP (Model Context Protocol) system that allows disabling automatic tool injection into requests. When enabled, MCP tools are only included when explicitly requested via context headers or filters, providing more granular control over tool availability.
## Changes
- Added `DisableAutoToolInject` field to `MCPToolManagerConfig` schema with runtime update support
- Implemented atomic boolean storage in `ToolsManager` to safely handle concurrent access
- Added logic in `ParseAndAddToolsToRequest` to respect the disable flag and only inject tools when explicit context filters are present
- Extended configuration management with database migration, UI controls, and API endpoints
- Added hot-reload capability through `UpdateMCPDisableAutoToolInject` methods across the stack
- Updated UI with a toggle switch and clear documentation about the feature's behavior
## Type of change
- [x] Feature
- [ ] Bug fix
- [ ] Refactor
- [ ] Documentation
- [ ] Chore/CI
## Affected areas
- [x] Core (Go)
- [x] Transports (HTTP)
- [ ] Providers/Integrations
- [ ] Plugins
- [x] UI (Next.js)
- [ ] Docs
## How to test
Validate the new MCP auto tool injection toggle:
```sh
# Core/Transports
go version
go test ./...
# UI
cd ui
pnpm i || npm i
pnpm test || npm test
pnpm build || npm run build
```
Test the feature:
1. Configure MCP clients and tools
2. Enable "Disable Auto Tool Injection" in the MCP configuration UI
3. Make requests without explicit tool headers - tools should not be injected
4. Make requests with `x-bf-mcp-include-tools` header - tools should be injected
5. Verify hot-reload works by toggling the setting without server restart
## Screenshots/Recordings
UI changes include a new toggle switch in the MCP configuration view with descriptive text explaining when tools are injected based on explicit headers.
## Breaking changes
- [ ] Yes
- [x] No
This is a backward-compatible addition with a default value of `false` (auto injection enabled).
## Related issues
This addresses the need for more granular control over MCP tool injection behavior in request processing.
## Security considerations
The feature provides better control over tool exposure by allowing administrators to require explicit opt-in for tool injection, potentially reducing unintended tool access.
## Checklist
- [x] I read `docs/contributing/README.md` and followed the guidelines
- [x] I added/updated tests where appropriate
- [x] I updated documentation where needed
- [x] I verified builds succeed (Go and UI)
- [x] I verified the CI pipeline passes locally if applicable
* feat: VK MCP config now works as an AllowList (#1940)
## Summary
This PR implements MCP tool governance by enforcing virtual key MCP configurations as an execution-time allow-list. When virtual keys have empty MCPConfigs, all MCP tools are denied. When non-empty, each tool is validated against the configured allow-list at both inference time and MCP tool execution.
## Changes
- **Context parameter updates**: Changed MCP-related functions to use `*schemas.BifrostContext` instead of `context.Context` to enable tool tracking
- **Tool tracking**: Added `BifrostContextKeyMCPAddedTools` context key to track which MCP tools are added to requests
- **Governance enforcement**: Virtual key MCP configurations now act as execution-time allow-lists with validation in both `PreMCPHook` and `evaluateGovernanceRequest`
- **Auto-injection control**: Added `DisableAutoToolInject` configuration option that respects the toggle and skips auto-injection when headers are already set by callers
- **Decision type**: Added `DecisionMCPToolBlocked` for MCP tool governance violations
- **UI improvements**: Updated MCP view description and sidebar item naming for better clarity
## Type of change
- [x] Feature
- [ ] Bug fix
- [ ] Refactor
- [ ] Documentation
- [ ] Chore/CI
## Affected areas
- [x] Core (Go)
- [x] Transports (HTTP)
- [ ] Providers/Integrations
- [x] Plugins
- [x] UI (Next.js)
- [ ] Docs
## How to test
Test MCP tool governance with virtual keys:
```sh
# Core/Transports
go version
go test ./...
# Test with virtual key having empty MCPConfigs (should deny all MCP tools)
curl -X POST /v1/chat/completions \
-H "x-bf-virtual-key: test-vk-empty-mcp" \
-d '{"model": "gpt-4", "messages": [{"role": "user", "content": "test"}]}'
# Test with virtual key having specific MCP tool allowlist
curl -X POST /v1/chat/completions \
-H "x-bf-virtual-key: test-vk-with-mcp" \
-d '{"model": "gpt-4", "messages": [{"role": "user", "content": "test"}]}'
# Test disable auto tool inject configuration
curl -X PUT /v1/config/mcp/disable-auto-tool-inject \
-d '{"disable": true}'
# UI
cd ui
pnpm i || npm i
pnpm test || npm test
pnpm build || npm run build
```
New configuration options:
- `disable_auto_tool_inject`: Boolean flag to disable automatic MCP tool injection
- Virtual key `MCPConfigs`: Array of MCP client configurations that act as allow-lists
## Screenshots/Recordings
UI changes include updated MCP configuration view with clearer descriptions for the disable auto tool injection toggle and improved sidebar navigation labels.
## Breaking changes
- [x] Yes
- [ ] No
**Impact**: MCP-related function signatures now require `*schemas.BifrostContext` instead of `context.Context`. Virtual keys with empty MCPConfigs will now deny all MCP tools by default.
**Migration**: Update any custom MCP integrations to use the new context parameter type. Configure MCPConfigs on virtual keys that need MCP tool access.
## Related issues
Implements MCP tool governance and execution-time validation for virtual key configurations.
## Security considerations
- **Access control**: Virtual key MCP configurations now enforce strict allow-lists for tool execution
- **Context isolation**: Tool tracking is isolated per request context to prevent cross-request leakage
- **Validation**: Both pre-execution and execution-time validation prevent unauthorized tool access
## Checklist
- [x] I read `docs/contributing/README.md` and followed the guidelines
- [x] I added/updated tests where appropriate
- [x] I updated documentation where needed
- [x] I verified builds succeed (Go and UI)
- [x] I verified the CI pipeline passes locally if applicable
* refactor: standardize empty array conventions for VK Provider Config Allowed Keys (#2006)
## Summary
Migrates VK provider config allowed keys from implicit allow-all semantics to explicit deny-by-default behavior. Adds `AllowAllKeys` boolean field to enable granular key access control while maintaining backward compatibility.
## Changes
- Added `AllowAllKeys` boolean field to `TableVirtualKeyProviderConfig` with database migration
- Backfilled existing configs with `allow_all_keys=true` to preserve current behavior
- Updated key resolution logic: empty keys now denies all access, `["*"]` wildcard allows all keys
- Modified governance resolver to set empty `includeOnlyKeys` slice when no keys are configured
- Enhanced HTTP handlers to recognize `["*"]` wildcard and set `AllowAllKeys` flag appropriately
- Updated UI to display "Allow All Keys" option and show deny-by-default messaging
- Added JSON unmarshaling support for `["*"]` wildcard in config files
## Type of change
- [x] Feature
- [ ] Bug fix
- [ ] Refactor
- [ ] Documentation
- [ ] Chore/CI
## Affected areas
- [x] Core (Go)
- [x] Transports (HTTP)
- [ ] Providers/Integrations
- [x] Plugins
- [x] UI (Next.js)
- [ ] Docs
## How to test
Validate the migration and new key access control behavior:
```sh
# Core/Transports
go version
go test ./...
# Test migration runs successfully
go run main.go migrate
# UI
cd ui
pnpm i || npm i
pnpm test || npm test
pnpm build || npm run build
```
Test scenarios:
1. Create VK with empty `key_ids` - should deny all keys
2. Create VK with `key_ids: ["*"]` - should allow all keys
3. Create VK with specific key IDs - should allow only those keys
4. Verify existing VKs maintain their current behavior after migration
## Screenshots/Recordings
UI now shows:
- "Allow All Keys" option in key selection dropdown
- "No keys allowed" vs "All keys allowed" status indicators
- "No providers configured (deny-by-default)" messaging
## Breaking changes
- [ ] Yes
- [x] No
The migration preserves existing behavior by setting `allow_all_keys=true` for configs that previously had no keys specified.
## Related issues
Part of VK access control enhancement initiative.
## Security considerations
Improves security posture by implementing deny-by-default semantics for key access. Existing deployments maintain current access patterns through automatic backfill migration.
## Checklist
- [x] I read `docs/contributing/README.md` and followed the guidelines
- [x] I added/updated tests where appropriate
- [x] I updated documentation where needed
- [x] I verified builds succeed (Go and UI)
- [x] I verified the CI pipeline passes locally if applicable
* refactor: standardize empty array conventions for allowed models (#2113)
## Summary
Standardizes empty array conventions across Bifrost to implement deny-by-default security semantics. Previously, empty arrays for `allowed_models` and `Models` fields meant "allow all", creating potential security gaps. Now `["*"]` explicitly means "allow all" while empty arrays mean "deny all".
## Changes
- **Core Logic**: Updated model filtering in `bifrost.go` and `selectKeyFromProviderForModel` to treat empty `Models` arrays as deny-all and `["*"]` as allow-all
- **Database Migration**: Added `migrationBackfillAllowedModelsWildcard` to convert existing empty arrays to `["*"]` preserving current behavior for existing records
- **Model Catalog**: Updated `IsModelAllowedForProvider` to use wildcard semantics with deny-by-default fallback
- **Schema Defaults**: Changed default `Models` value from `[]` to `["*"]` in table definitions and form schemas
- **UI Components**: Enhanced `ModelMultiselect` with `allowAllOption` prop and updated virtual key forms to handle wildcard selection
- **Documentation**: Updated JSON schemas, comments, and tooltips to reflect new conventions
- **Governance**: Updated provider config filtering logic to use new wildcard semantics
- **Server Bootstrap**: Added wildcard filtering when loading models to prevent literal "*" from appearing as a model name
## Type of change
- [x] Refactor
- [ ] Bug fix
- [ ] Feature
- [ ] Documentation
- [ ] Chore/CI
## Affected areas
- [x] Core (Go)
- [x] Transports (HTTP)
- [x] Providers/Integrations
- [x] Plugins
- [x] UI (Next.js)
- [x] Docs
## How to test
Validate the migration and new semantics:
```sh
# Core/Transports
go version
go test ./...
# UI
cd ui
pnpm i || npm i
pnpm test || npm test
pnpm build || npm run build
```
Test scenarios:
1. Create new virtual keys - should default to `["*"]` for allowed models
2. Create new provider keys - should default to `["*"]` for models
3. Verify existing keys with empty arrays are migrated to `["*"]`
4. Test that empty arrays now deny all models/keys as expected
5. Verify UI shows "All models allowed" for wildcard and "No models (deny all)" for empty arrays
## Screenshots/Recordings
UI changes include:
- Model multiselect now shows "Allow All Models" option
- Virtual key details display "All Models" badge for wildcard vs "No models (deny all)" for empty
- Form placeholders updated to reflect new semantics
## Breaking changes
- [x] Yes
- [ ] No
**Migration Impact**: The database migration automatically converts existing empty `allowed_models` and `models_json` arrays to `["*"]`, preserving current behavior. However, any new configurations with empty arrays will now deny access instead of allowing all. Applications relying on "empty = allow all" semantics must be updated to use `["*"]` explicitly.
## Related issues
Part of security hardening initiative to implement explicit allow-lists and deny-by-default semantics across Bifrost configuration.
## Security considerations
This change significantly improves security posture by:
- Eliminating ambiguous "empty means allow all" semantics
- Implementing explicit deny-by-default for new configurations
- Requiring intentional wildcard usage via `["*"]` for broad access
- Maintaining backward compatibility through automatic migration
## Checklist
- [x] I read `docs/contributing/README.md` and followed the guidelines
- [x] I added/updated tests where appropriate
- [x] I updated documentation where needed
- [x] I verified builds succeed (Go and UI)
- [x] I verified the CI pipeline passes locally if applicable
* refactor: replace string slices with WhiteList for allowlist fields (#2125)
## Summary
Introduces a new `WhiteList` type to standardize whitelist behavior across the codebase, replacing manual slice operations and string comparisons with semantic methods for handling allow/deny lists.
## Changes
- Added `WhiteList` type with methods `IsAllowed()`, `IsUnrestricted()`, `IsEmpty()`, `Contains()`, and `Validate()`
- Replaced `[]string` fields with `WhiteList` for model restrictions, tool filtering, and key access controls
- Updated all whitelist logic to use semantic methods instead of manual `slices.Contains()` checks
- Added validation to ensure wildcards ("*") aren't mixed with specific values and prevent duplicates
- Improved case-insensitive matching for whitelist comparisons
## Type of change
- [x] Refactor
## Affected areas
- [x] Core (Go)
- [x] Transports (HTTP)
- [x] Plugins
## How to test
Verify that whitelist behavior remains consistent across all affected components:
```sh
# Core/Transports
go version
go test ./...
# Test specific whitelist scenarios:
# - Empty lists deny all access
# - ["*"] allows all access
# - Specific lists only allow listed items
# - Mixed wildcards and specific items are rejected
# - Duplicate entries are rejected
```
Test key model filtering, MCP tool execution, and virtual key configurations to ensure whitelist logic works correctly.
## Screenshots/Recordings
N/A
## Breaking changes
- [ ] Yes
- [x] No
The `WhiteList` type maintains the same JSON serialization format as `[]string`, so existing configurations remain compatible.
## Related issues
N/A
## Security considerations
Improves security by standardizing deny-by-default behavior and adding validation to prevent misconfigured whitelists that could inadvertently grant excessive permissions.
## Checklist
- [x] I read `docs/contributing/README.md` and followed the guidelines
- [x] I added/updated tests where appropriate
- [x] I updated documentation where needed
- [x] I verified builds succeed (Go and UI)
- [x] I verified the CI pipeline passes locally if applicable
* feat: add request-level extra headers support for MCP tool execution (#2126)
## Summary
This PR adds support for request-level extra headers in MCP tool execution, allowing callers to forward specific headers to MCP servers at runtime based on a per-client allowlist configuration.
## Changes
- Added `AllowedExtraHeaders` field to MCP client configuration with allowlist semantics (empty array = deny all, `["*"]` = allow all)
- Introduced `BifrostContextKeyMCPExtraHeaders` context key to track headers forwarded to MCP tools
- Created `core/mcp/utils` package with `GetHeadersForToolExecution` function to merge static and dynamic headers
- Updated MCP tool execution in both regular tool manager and Starlark code mode to use the new header forwarding system
- Added database migration for `allowed_extra_headers_json` column in MCP client table
- Updated UI to include allowed extra headers configuration in MCP client management
- Enhanced auth demo server example to demonstrate tool-execution level authentication patterns
## Type of change
- [x] Feature
## Affected areas
- [x] Core (Go)
- [x] Transports (HTTP)
- [x] UI (Next.js)
## How to test
1. Configure an MCP client with allowed extra headers:
```json
{
"name": "test-client",
"connection_string": "http://localhost:3002/",
"auth_type": "headers",
"headers": {
"X-API-Key": "connection-secret"
},
"allowed_extra_headers": ["X-Tool-Token"],
"tools_to_execute": ["*"]
}
```
2. Make requests with extra headers that should be forwarded:
```bash
curl -X POST http://localhost:8080/v1/chat/completions \
-H "Authorization: Bearer your-key" \
-H "X-Tool-Token: tool-execution-secret" \
-d '{
"model": "gpt-4",
"messages": [{"role": "user", "content": "Use the secret_data tool"}],
"tools": [{"type": "function", "function": {"name": "secret_data"}}]
}'
```
3. Test the auth demo server:
```bash
cd examples/mcps/auth-demo-server
go run main.go
# Server demonstrates two-tier auth: connection-level (X-API-Key) and tool-level (X-Tool-Token)
```
4. Run tests:
```sh
go test ./core/mcp/...
go test ./transports/bifrost-http/...
cd ui
pnpm test
pnpm build
```
## Breaking changes
- [ ] Yes
- [x] No
This is a backward-compatible addition. Existing MCP clients will have empty `allowed_extra_headers` (deny all extra headers) which maintains current behavior.
## Security considerations
- Extra headers are filtered through a strict allowlist per MCP client
- Security denylist prevents auth header overrides via extra headers
- Two-tier authentication pattern demonstrated: connection-level + tool-execution level
- Headers are only forwarded to MCP servers that explicitly allow them
## Checklist
- [x] I read `docs/contributing/README.md` and followed the guidelines
- [x] I added/updated tests where appropriate
- [x] I updated documentation where needed
- [x] I verified builds succeed (Go and UI)
- [x] I verified the CI pipeline passes locally if applicable
* fix: apply MCP tool filtering headers to tools/list response when using bifrost as MCP gateway (#2127)
## Summary
Adds support for `x-bf-mcp-include-clients` and `x-bf-mcp-include-tools` request headers to filter MCP tools/list response when using Bifrost as an MCP gateway. This ensures that tool filtering is respected at the MCP protocol level, not just during inference.
## Changes
- Implemented dynamic tool filtering in MCP server handlers that respects per-request include headers
- Added `makeIncludeClientsFilter()` function that filters tools based on request context values
- Registered the tool filter on both global and virtual key MCP servers during initialization
- Updated documentation to clarify that `mcp-include-tools` requires `clientName-toolName` format
- Enhanced examples in documentation to show proper tool naming format
## Type of change
- [x] Bug fix
- [ ] Feature
- [ ] Refactor
- [ ] Documentation
- [ ] Chore/CI
## Affected areas
- [x] Core (Go)
- [x] Transports (HTTP)
- [ ] Providers/Integrations
- [ ] Plugins
- [ ] UI (Next.js)
- [x] Docs
## How to test
Test MCP gateway functionality with tool filtering:
```sh
# Test tools/list filtering with include-tools header
curl --location 'http://localhost:8080/mcp/tools/list' \
--header 'x-bf-mcp-include-tools: gmail-send_email,filesystem-read_file' \
--header 'Authorization: Bearer your-vk-here'
# Test tools/list filtering with include-clients header
curl --location 'http://localhost:8080/mcp/tools/list' \
--header 'x-bf-mcp-include-clients: gmail,filesystem' \
--header 'Authorization: Bearer your-vk-here'
# Verify chat completions still respect the same headers
curl --location 'http://localhost:8080/v1/chat/completions' \
--header 'x-bf-mcp-include-tools: gmail-send_email' \
--header 'Content-Type: application/json' \
--data '{
"model": "openai/gpt-4o-mini",
"messages": [{"role": "user", "content": "What tools are available?"}]
}'
```
## Screenshots/Recordings
N/A
## Breaking changes
- [ ] Yes
- [x] No
## Related issues
N/A
## Security considerations
The tool filtering mechanism ensures that virtual key restrictions are properly enforced at the MCP protocol level, preventing unauthorized access to tools that should be filtered out based on request headers.
## Checklist
- [x] I read `docs/contributing/README.md` and followed the guidelines
- [x] I added/updated tests where appropriate
- [x] I updated documentation where needed
- [x] I verified builds succeed (Go and UI)
- [x] I verified the CI pipeline passes locally if applicable
* refactor: parallelize model listing for providers to speed up startup time (#2151)
## Summary
Parallelizes model listing operations for providers during server startup and provider reloading to significantly reduce initialization time. Previously, model listing was performed sequentially for each provider, causing slower startup times especially when multiple providers were configured.
## Changes
- Added concurrent execution using goroutines and sync.WaitGroup for model listing operations in three key functions: `ReloadProvider`, `ForceReloadPricing`, and `Bootstrap`
- In `ReloadProvider`, both filtered and unfiltered model listing requests now run concurrently for the same provider
- In `ForceReloadPricing` and `Bootstrap`, model listing for different providers now runs in parallel instead of sequentially
- Moved provider key retrieval earlier in `ReloadProvider` to ensure it happens before concurrent model listing
- Added proper context cancellation with defer statements for bifrost contexts
## Type of change
- [x] Refactor
## Affected areas
- [x] Transports (HTTP)
## How to test
Test server startup time with multiple providers configured to verify the performance improvement:
```sh
# Core/Transports
go version
go test ./...
# Test with multiple providers configured
# Measure startup time before and after the change
time go run main.go
```
Configure multiple providers in your bifrost configuration and observe faster startup times, especially noticeable when providers have high latency or many models.
## Screenshots/Recordings
N/A
## Breaking changes
- [ ] Yes
- [x] No
## Related issues
N/A
## Security considerations
No security implications. The change maintains the same authentication and authorization patterns while improving performance through parallelization.
## Checklist
- [x] I read `docs/contributing/README.md` and followed the guidelines
- [x] I added/updated tests where appropriate
- [x] I updated documentation where needed
- [x] I verified builds succeed (Go and UI)
- [x] I verified the CI pipeline passes locally if applicable
* fix: reorder migrations and set AllowAllKeys to true for virtual key provider configs (#2158)
## Summary
Fixes database migration ordering issue and ensures virtual key configurations are properly initialized with the AllowAllKeys field set to true.
## Changes
- Reordered database migrations to execute `migrationAddAllowAllKeysToProviderConfig` before `migrationBackfillEmptyVirtualKeyConfigs` to ensure the AllowAllKeys column exists before backfilling
- Added `AllowAllKeys: true` to provider configurations created during virtual key backfill migration to enable unrestricted key access by default
## Type of change
- [x] Bug fix
- [ ] Feature
- [ ] Refactor
- [ ] Documentation
- [ ] Chore/CI
## Affected areas
- [x] Core (Go)
- [ ] Transports (HTTP)
- [ ] Providers/Integrations
- [ ] Plugins
- [ ] UI (Next.js)
- [ ] Docs
## How to test
Verify that database migrations run successfully and virtual key configurations are created with proper defaults:
```sh
# Core/Transports
go version
go test ./...
```
Test migration ordering by running against a fresh database to ensure no column reference errors occur.
## Screenshots/Recordings
N/A
## Breaking changes
- [ ] Yes
- [x] No
## Related issues
N/A
## Security considerations
This change enables unrestricted key access by default for virtual key configurations, which may have security implications depending on the intended access control model.
## Checklist
- [ ] I read `docs/contributing/README.md` and followed the guidelines
- [ ] I added/updated tests where appropriate
- [ ] I updated documentation where needed
- [ ] I verified builds succeed (Go and UI)
- [ ] I verified the CI pipeline passes locally if applicable
* feat: implement scoped pricing override
* refactor: custom pricing refactor
* fix: resolve merge conflicts in config loading and governance functions (#2230)
## Summary
Resolves Git merge conflicts in the bifrost-http configuration loading code by cleaning up duplicate function definitions and consolidating the configuration initialization flow.
## Changes
- Removed Git merge conflict markers and duplicate code blocks from `LoadConfig` function
- Consolidated governance configuration loading by keeping both `loadGovernanceConfigFromFile` and `loadGovernanceConfig` functions with distinct purposes
- Removed duplicate `convertSchemasMCPClientConfigToTable` function definition
- Moved pricing overrides initialization logic to `initFrameworkConfig` function for better organization
- Cleaned up extensive duplicate default configuration loading code that was causing merge conflicts
- Changed error handling for pricing overrides from returning error to logging warning
## Type of change
- [x] Bug fix
- [ ] Feature
- [ ] Refactor
- [ ] Documentation
- [ ] Chore/CI
## Affected areas
- [x] Core (Go)
- [x] Transports (HTTP)
- [ ] Providers/Integrations
- [ ] Plugins
- [ ] UI (Next.js)
- [ ] Docs
## How to test
Verify that configuration loading works correctly without merge conflicts:
```sh
# Core/Transports
go version
go test ./...
go build ./transports/bifrost-http/...
```
Test configuration loading with various scenarios:
- Config file present
- Config file absent (default loading)
- Store-based configuration
- Governance and MCP configuration loading
## Screenshots/Recordings
N/A
## Breaking changes
- [ ] Yes
- [x] No
## Related issues
N/A
## Security considerations
No security implications - this is a merge conflict resolution that maintains existing functionality.
## Checklist
- [x] I read `docs/contributing/README.md` and followed the guidelines
- [x] I added/updated tests where appropriate
- [x] I updated documentation where needed
- [x] I verified builds succeed (Go and UI)
- [x] I verified the CI pipeline passes locally if applicable
* feat: add Stability AI model support for Bedrock image generation (#2180)
## Summary
Adds support for Stability AI image generation models (stability.stable-image-*) to the Bedrock provider, enabling text-to-image generation with models like stability.stable-image-core-v1:1 and stability.stable-image-ultra-v1:1.
## Changes
- Added `isStabilityAIModel()` function to detect Stability AI models by "stability." prefix
- Created `ToStabilityAIImageGenerationRequest()` to convert Bifrost requests to Stability AI's flat request format
- Implemented `StabilityAIImageGenerationRequest` type with support for prompt, mode, aspect_ratio, output_format, seed, and negative_prompt parameters
- Added conditional routing in `ImageGeneration()` to use Stability AI request format when appropriate
- Extended known fields for image generation parameters to include "aspect_ratio" and "input_images"
- Updated documentation comment to reflect Stability AI model support
## Type of change
- [x] Feature
- [ ] Bug fix
- [ ] Refactor
- [ ] Documentation
- [ ] Chore/CI
## Affected areas
- [x] Core (Go)
- [x] Transports (HTTP)
- [x] Providers/Integrations
- [ ] Plugins
- [ ] UI (Next.js)
- [ ] Docs
## How to test
Test Stability AI image generation through the Bedrock provider:
```sh
# Core/Transports
go version
go test ./...
# Test with a Stability AI model
curl -X POST http://localhost:8080/v1/images/generations \
-H "Content-Type: application/json" \
-H "Authorization: Bearer your-key" \
-d '{
"model": "stability.stable-image-core-v1:1",
"prompt": "A beautiful sunset over mountains",
"aspect_ratio": "16:9",
"output_format": "PNG"
}'
```
Ensure AWS credentials are configured for Bedrock access and the Stability AI models are available in your region.
## Screenshots/Recordings
N/A
## Breaking changes
- [ ] Yes
- [x] No
## Related issues
N/A
## Security considerations
No additional security implications beyond existing Bedrock provider authentication and AWS credential handling.
## Checklist
- [x] I read `docs/contributing/README.md` and followed the guidelines
- [x] I added/updated tests where appropriate
- [x] I updated documentation where needed
- [x] I verified builds succeed (Go and UI)
- [x] I verified the CI pipeline passes locally if applicable
* feat: add Stability AI image edit models support to Bedrock provider (#2225)
## Summary
Adds support for Stability AI image editing models in the Bedrock provider, expanding image editing capabilities beyond the existing Titan and Nova Canvas models.
## Changes
- Added `getStabilityAIEditTaskType()` function to infer edit task types from Stability AI model names (inpaint, outpaint, recolor, search-replace, erase-object, remove-bg, control-sketch, control-structure, style-guide, style-transfer, upscale-creative, upscale-conservative, upscale-fast)
- Created `ToStabilityAIImageEditRequest()` function to convert Bifrost requests to Stability AI's flat JSON format, with task-specific field validation
- Added `StabilityAIImageEditRequest` struct with comprehensive field support for all Stability AI edit operations
- Enhanced `BedrockImageGenerationResponse` with Seeds and FinishReasons fields for Stability AI compatibility
- Modified `ImageEdit()` method to route requests to appropriate conversion function based on model type
- Updated documentation to reflect expanded model support
## Type of change
- [x] Feature
- [ ] Bug fix
- [ ] Refactor
- [ ] Documentation
- [ ] Chore/CI
## Affected areas
- [x] Core (Go)
- [ ] Transports (HTTP)
- [x] Providers/Integrations
- [ ] Plugins
- [ ] UI (Next.js)
- [ ] Docs
## How to test
Test with various Stability AI edit models through the Bedrock provider:
```sh
# Core/Transports
go version
go test ./...
# Test image editing with Stability AI models
# Example: stable-image-inpaint, stable-outpaint, stable-creative-upscale, etc.
```
Verify that task-specific parameters are correctly mapped and invalid fields are filtered out based on the detected task type.
## Screenshots/Recordings
N/A - Backend functionality only
## Breaking changes
- [ ] Yes
- [x] No
## Related issues
N/A
## Security considerations
Image data is handled as base64-encoded strings. Mask and image parameters are properly validated before processing.
## Checklist
- [x] I read `docs/contributing/README.md` and followed the guidelines
- [x] I added/updated tests where appropriate
- [x] I updated documentation where needed
- [x] I verified builds succeed (Go and UI)
- [x] I verified the CI pipeline passes locally if applicable
* fix: send back accumulated usage in MCP agent mode (#2246)
## Summary
This PR fixes token usage tracking in MCP agent mode by accumulating usage across all LLM calls in the agent loop and returning the total usage in the final response.
## Changes
- Added usage accumulation logic in the MCP agent execution loop to track token consumption across multiple LLM calls
- Implemented `mergeUsage` function to combine token counts and costs from multiple `BifrostLLMUsage` values, handling all detail sub-fields including prompt tokens, completion tokens, and cost breakdowns
- Extended agent API adapters with `extractUsage` and `applyUsage` methods to handle usage extraction and application for both Chat API and Responses API
- Applied accumulated usage to the final response before returning it to the client
## Type of change
- [x] Bug fix
- [ ] Feature
- [ ] Refactor
- [ ] Documentation
- [ ] Chore/CI
## Affected areas
- [x] Core (Go)
- [x] Transports (HTTP)
- [ ] Providers/Integrations
- [ ] Plugins
- [ ] UI (Next.js)
- [ ] Docs
## How to test
Test MCP agent mode with multiple tool calls to verify usage accumulation:
```sh
# Core/Transports
go version
go test ./...
# Test MCP agent mode with multiple LLM calls
# Verify that the returned usage reflects the sum of all calls in the agent loop
# Check that both token counts and cost details are properly accumulated
```
## Screenshots/Recordings
N/A
## Breaking changes
- [ ] Yes
- [x] No
## Related issues
N/A
## Security considerations
No security implications - this change only affects usage tracking and reporting.
## Checklist
- [ ] I read `docs/contributing/README.md` and followed the guidelines
- [ ] I added/updated tests where appropriate
- [ ] I updated documentation where needed
- [ ] I verified builds succeed (Go and UI)
- [ ] I verified the CI pipeline passes locally if applicable
* [codemode]: fixing string escape corruption, enable top-level control flow in starlark, refining the prompt of executecode tool (#2206)
## Changes
- **Enhanced Starlark dialect configuration**: Enabled top-level control flow statements (if/for/while), while loops, set() builtin, global variable reassignment, and recursive functions for a more Python-like experience
- **Improved string escape handling**: Removed automatic `\n` to newline conversion, allowing Starlark's native string escape processing to handle `\n`, `\t`, and other escape sequences correctly
- **Updated tool description**: Streamlined the executeToolCode tool description with clearer syntax notes, explicit documentation of Starlark differences from Python (no try/except, no classes, no imports, no f-strings), and emphasis on fresh isolated scope per execution
- **Enhanced error hints**: Added specific error messages for unsupported Python features like try/except/finally/raise, with guidance on alternative approaches and scope persistence warnings
- **Comprehensive test coverage**: Added tests for dialect options, string escape preservation, unsupported feature detection, and end-to-end JSON deserialization scenarios
## Type of change
- [ ] Feature
- [ ] Bug fix
- [x] Refactor
- [ ] Documentation
- [ ] Chore/CI
## Affected areas
- [x] Core (Go) - Starlark CodeMode improvements
- [ ] Transports (HTTP)
- [ ] Providers/Integrations
- [ ] Plugins
- [ ] UI (Next.js)
- [ ] Docs
## How to test
Test the enhanced Starlark features with MCP CodeMode:
```sh
# Test dialect options (top-level control flow, while loops, etc.)
make test-mcp TESTCASE=TestStarlarkDialectOptions
# Test string escape handling
make test-mcp PATTERN=TestStarlarkStringEscape
# Test unsupported feature detection
make test-mcp PATTERN=TestStarlarkUnsupportedFeatures
```
## Breaking changes
- [ ] Yes
- [x] No
The Starlark changes are additive and maintain backward compatibility while enabling more Python-like syntax.
## Security considerations
Starlark CodeMode maintains its existing sandboxing with no additional network or filesystem access. The dialect enhancements only affect language features within the existing security boundary.
* logging in plugins (#2215)
## Summary
Reorders middleware initialization in the Bifrost HTTP server to ensure tracing middleware is added before transport interceptor middleware in the inference pipeline.
## Changes
- Moved tracing middleware initialization and setup earlier in the bootstrap process
- Reordered middleware registration so tracing middleware is prepended before transport interceptor middleware
- Updated comments to clarify the middleware ordering logic and rationale
The change ensures that tracing context and trace IDs are properly established before other middleware components process requests.
## Type of change
- [x] Bug fix
- [ ] Feature
- [ ] Refactor
- [ ] Documentation
- [ ] Chore/CI
## Affected areas
- [ ] Core (Go)
- [x] Transports (HTTP)
- [ ] Providers/Integrations
- [ ] Plugins
- [ ] UI (Next.js)
- [ ] Docs
## How to test
Verify that tracing middleware executes before transport interceptor middleware by checking trace logs and middleware execution order.
```sh
# Core/Transports
go version
go test ./...
```
Test with tracing enabled to ensure trace IDs are properly set in context before subsequent middleware processing.
## Screenshots/Recordings
N/A
## Breaking changes
- [ ] Yes
- [x] No
## Related issues
N/A
## Security considerations
No security implications - this is a middleware ordering change that affects observability components.
## Checklist
- [ ] I read `docs/contributing/README.md` and followed the guidelines
- [ ] I added/updated tests where appropriate
- [ ] I updated documentation where needed
- [ ] I verified builds succeed (Go and UI)
- [ ] I verified the CI pipeline passes locally if applicable
* fix: handling text, vtt, srt response format for transcriptions (#2102)
* feat: add virtual key access management for MCP clients (#2255)
## Summary
Adds virtual key access management to MCP client configuration, allowing administrators to control which virtual keys can access specific MCP servers and which tools they can execute on a per-VK basis.
## Changes
- Added `vk_configs` field to MCP client update API that accepts an array of virtual key configurations
- Each VK config specifies a virtual key ID and the tools it's allowed to execute on that MCP server
- When `vk_configs` is provided, it atomically replaces all existing VK assignments for the MCP client
- Added database method `GetVirtualKeyMCPConfigsByMCPClientID` to retrieve VK configs by MCP client
- Updated OpenAPI documentation to describe the new VK configuration functionality
- Enhanced UI with virtual key access management section in the MCP client sheet
- Added Go SDK context keys for MCP tool filtering: `MCPContextKeyIncludeClients`, `MCPContextKeyIncludeTools`, and `BifrostContextKeyMCPExtraHeaders`
- Updated context keys documentation with comprehensive MCP configuration examples
## Type of change
- [x] Feature
## Affected areas
- [x] Core (Go)
- [x] Transports (HTTP)
- [x] UI (Next.js)
- [x] Docs
## How to test
1. Create an MCP client with tools available
2. Create virtual keys in the system
3. Update the MCP client with VK configurations:
```sh
curl -X PUT /api/mcp/client/{id} \
-H "Content-Type: application/json" \
-d '{
"name": "test-client",
"vk_configs": [
{
"virtual_key_id": "vk-123",
"tools_to_execute": ["*"]
},
{
"virtual_key_id": "vk-456",
"tools_to_execute": ["read_file", "write_file"]
}
]
}'
```
4. Verify VK assignments are created/updated in the database
5. Test the UI by opening an MCP client sheet and managing virtual key access
```sh
# Core/Transports
go version
go test ./...
# UI
cd ui
pnpm i || npm i
pnpm test || npm test
pnpm build || npm run build
```
## Screenshots/Recordings
The UI now includes a "Virtual Key Access" section in the MCP client configuration sheet where administrators can:
- Add virtual keys to grant access to the MCP server
- Configure which specific tools each virtual key can execute
- Remove virtual key access entirely
## Breaking changes
- [ ] Yes
- [x] No
## Related issues
This enables fine-grained access control for MCP servers at the virtual key level, complementing the existing governance and budgeting features.
## Security considerations
- VK access controls are enforced through the governance plugin during MCP tool execution
- The atomic replacement of VK assignments prevents partial updates that could leave the system in an inconsistent state
- Tool-level restrictions allow principle of least privilege by limiting which MCP tools each virtual key can access
## Checklist
- [x] I read `docs/contributing/README.md` and followed the guidelines
- [x] I added/updated tests where appropriate
- [x] I updated documentation where needed
- [x] I verified builds succeed (Go and UI)
- [x] I verified the CI pipeline passes locally if applicable
* feat: adds option to allow MCP clients to run on all virtual keys (#2258)
## Summary
Adds a new `AllowOnAllVirtualKeys` configuration option for MCP clients that enables them to be accessible to all virtual keys without requiring explicit per-key assignment. When enabled, all tools from the MCP client are available to every virtual key.
## Changes
- Added `AllowOnAllVirtualKeys` boolean field to `MCPClientConfig` schema and database table
- Updated MCP client manager to handle the new field during client updates
- Modified governance plugin to check for clients with `AllowOnAllVirtualKeys` enabled and automatically include their tools for all virtual keys
- Added database migration to add the new column to `TableMCPClient`
- Updated UI to include a toggle for the new setting with tooltip explanation
- Added OpenAPI documentation for the new field
- Updated configuration store methods to persist and retrieve the new field
## Type of change
- [x] Feature
- [ ] Bug fix
- [ ] Refactor
- [ ] Documentation
- [ ] Chore/CI
## Affected areas
- [x] Core (Go)
- [x] Transports (HTTP)
- [ ] Providers/Integrations
- [x] Plugins
- [x] UI (Next.js)
- [x] Docs
## How to test
1. Create or update an MCP client with `allow_on_all_virtual_keys: true`
2. Verify that the client's tools are available to all virtual keys without explicit assignment
3. Test that the governance plugin correctly allows tools from such clients
4. Verify the UI toggle works correctly in the MCP client edit sheet
```sh
# Core/Transports
go version
go test ./...
# UI
cd ui
pnpm i || npm i
pnpm test || npm test
pnpm build || npm run build
```
The new configuration field `allow_on_all_virtual_keys` defaults to `false` to maintain backward compatibility.
## Screenshots/Recordings
If UI changes, add before/after screenshots or short clips.
## Breaking changes
- [ ] Yes
- [x] No
This is a backward-compatible addition with the new field defaulting to `false`.
## Related issues
Link related issues and discussions. Example: Closes #123
## Security considerations
This feature reduces access control granularity by allowing MCP clients to bypass virtual key restrictions when enabled. Administrators should carefully consider which MCP clients should have this permission as it grants broad access across all virtual keys.
## Checklist
- [x] I read `docs/contributing/README.md` and followed the guidelines
- [x] I added/updated tests where appropriate
- [x] I updated documentation where needed
- [x] I verified builds succeed (Go and UI)
- [x] I verified the CI pipeline passes locally if applicable
* feat: add provider keys CRUD to configstore and in-memory store (#2159)
## Summary
Adds dedicated CRUD operations for individual provider keys at the data layer
(configstore interface + RDB implementation) and in-memory store. This enables
key-level operations without replacing the entire provider key set, which is
required for the new `/api/providers/{provider}/keys/*` endpoints.
## Changes
- Added `GetProviderKeys`, `GetProviderKey`, `CreateProviderKey`,
`UpdateProviderKey`, `DeleteProviderKey` to `ConfigStore` interface
- Implemented all five methods in `RDBConfigStore` with proper GORM queries,
error handling, and `ErrNotFound` propagation
- Extracted `schemaKeyFromTableKey` and `tableKeyFromSchemaKey` helpers to
deduplicate key conversion logic (previously inlined in `GetProvidersConfig`
and `GetProviderConfig`)
- Added `AddProviderKey`, `UpdateProviderKey`, `RemoveProviderKey` to in-memory
`Config` with mutex locking, DB persistence, and rollback on client update
failure
- Added `GetProviderKeysRaw`, `GetProviderKeysRedacted`, `GetProviderKeyRaw`,
`GetProviderKeyRedacted` read methods
- Added `TestProviderKeyCRUD` and `TestProviderKeyCRUD_ProviderMustExist`
integration tests
- Updated `MockConfigStore` with all five new interface methods
## Type of change
- [x] Feature
- [ ] Bug fix
- [ ] Refactor
- [ ] Documentation
- [ ] Chore/CI
## Affected areas
- [x] Core (Go)
- [x] Transports (HTTP)
- [ ] Providers/Integrations
- [ ] Plugins
- [ ] UI (Next.js)
- [ ] Docs
## How to test
```sh
# Run configstore tests
go test ./framework/configstore/... -v -run TestProviderKeyCRUD
# Run config tests (mock store)
go test ./transports/bifrost-http/lib/... -v
```
## Screenshots/Recordings
N/A
## Breaking changes
- [ ] Yes
- [x] No
## Related issues
N/A
## Security considerations
Key values are handled through existing redaction infrastructure. No new secret
exposure paths introduced.
## Checklist
- [x] I read `docs/contributing/README.md` and followed the guidelines
- [x] I added/updated tests where appropriate
- [ ] I updated documentation where needed
- [x] I verified builds succeed (Go and UI)
- [ ] I verified the CI pipeline passes locally if applicable
* feat: add provider keys HTTP handlers and refactor optional keys (#2160)
## Summary
Adds HTTP handlers for the dedicated provider keys CRUD endpoints and removes
`keys` from provider API responses and payloads. Keys are now exclusively
managed via `/api/providers/{provider}/keys/*`. Also fixes a context timeout bug
in `ReloadProvider` where model discovery could exhaust the shared context
budget, causing subsequent DB calls to fail.
## Changes
### Provider keys handlers (`provider_keys.go`)
- New file with five handlers: `listProviderKeys`, `getProviderKey`,
`createProviderKey`, `updateProviderKey`, `deleteProviderKey`
- Includes `mergeUpdatedKey` (redacted value preservation logic used by
`updateProviderKey`)
- Key handlers enforce keyless provider validation and trigger model discovery
after mutations
### Provider handlers cleanup (`providers.go`)
- Registered new key routes: `GET/POST /api/providers/{provider}/keys`,
`GET/PUT/DELETE /api/providers/{provider}/keys/{key_id}`
- Extracted inline anonymous structs into named `providerCreatePayload` and
`providerUpdatePayload` types (without `Keys` field)
- Removed `Keys` field from `ProviderResponse`
- Switched `addProvider` from `json.Unmarshal` to `sonic.Unmarshal`
- Removed `oldConfigRedacted` fetch and the entire key merge block
(`mergeKeys`, `hasKeys`, `slices` usage) from `updateProvider`
- Removed `Keys` from `getProviderResponseFromConfig` response builder
- Removed unused `encoding/json` import
### Context timeout fix (`server.go`)
- Split shared `bfCtx` in `ReloadProvider` into separate contexts:
`filteredBfCtx` (15s) for filtered `ListModelsRequest` and `unfilteredBfCtx`
(fresh 15s) for unfiltered `ListModelsRequest`, each cancelled after use
- Changed `GetKeysByProvider` to use `context.Background()` since it's a local
DB call that shouldn't be gated by model discovery timeouts
- Added `hasNoKeys` check to emit warn-level logs instead of errors when model
discovery fails because no keys are configured
- Read in-memory key count via `GetProviderKeysRaw` for the `hasNoKeys` check
### Tests (`providers_test.go`)
- Cleared file (contained only tests for removed inline struct decoding)
## Type of change
- [x] Feature
- [x] Bug fix
- [x] Refactor
- [ ] Documentation
- [ ] Chore/CI
## Affected areas
- [ ] Core (Go)
- [x] Transports (HTTP)
- [ ] Providers/Integrations
- [ ] Plugins
- [ ] UI (Next.js)
- [ ] Docs
## How to test
```sh
# Build
go build ./transports/bifrost-http/...
# Manual: start Bifrost, then test key CRUD
curl -X POST localhost:8080/api/providers/openai/keys -d '{"name":"test-key","value":"sk-test"}'
curl localhost:8080/api/providers/openai/keys
curl -X PUT localhost:8080/api/providers/openai/keys/{key_id} -d '{"name":"updated","value":"sk-new"}'
curl -X DELETE localhost:8080/api/providers/openai/keys/{key_id}
# Verify provider endpoints no longer return keys
curl localhost:8080/api/providers/openai | jq 'has("keys")' # should be false
```
## Screenshots/Recordings
N/A
## Breaking changes
- [x] Yes
- [ ] No
Provider API responses no longer include `keys` field. Provider create/update
payloads no longer accept `keys`. Clients must use the dedicated
`/api/providers/{provider}/keys/*` endpoints for key management.
## Related issues
N/A
## Security considerations
- Key handlers use existing redaction infrastructure (`GetProviderKeyRedacted`)
before returning responses
- Keyless provider validation prevents key creation on providers that don't
support keys
## Checklist
- [x] I read `docs/contributing/README.md` and followed the guidelines
- [x] I added/updated tests where appropriate
- [ ] I updated documentation where needed
- [x] I verified builds succeed (Go and UI)
- [ ] I verified the CI pipeline passes locally if applicable
* feat: migrate frontend to dedicated provider keys API (#2161)
## Summary
Migrates the frontend from reading provider keys via `provider.keys` (removed
from provider API response in PR #2160) to the dedicated `getProviderKeys`
query and `/api/keys` endpoint. Removes `keys` from all provider TypeScript
types. Key mutations patch caches from authoritative server responses; provider
updates invalidate the `ProviderKeys` tag to refresh key statuses after model
discovery. Also adds a read-only routing rule info sheet.
## Changes
### Types (`config.ts`, `schemas.ts`)
- Removed `keys` field from `ModelProviderConfig`, `AddProviderRequest`, and
`UpdateProviderRequest`
- Added `CreateProviderKeyRequest`, `UpdateProviderKeyRequest`,
`ListProviderKeysResponse` types
### Store (`providersApi.ts`, `baseApi.ts`)
- Added `ProviderKeys` tag type to `baseApi`
- Changed `getProviderKeys`/`getProviderKey` from `Providers` tag to
`ProviderKeys` tag (avoids invalidating provider cache on key changes)
- Added `invalidatesTags: [ProviderKeys, DBKeys]` on `updateProvider` mutation
(refreshes key statuses after model discovery)
- Removed `getProvider`/`getProviders` cache patches from `createProviderKey`,
`updateProviderKey`, `deleteProviderKey` (providers no longer carry keys)
- Added duplicate-check guards on `createProviderKey` cache patches to prevent
ghost keys
- Each key mutation patches `getProviderKeys` and `getAllKeys` caches from
authoritative server response
### Components
- **`modelProviderKeysTableView.tsx`**: Already uses `useGetProviderKeysQuery`;
formatting/indentation fixes
- **`page.tsx`**: Removed `keys: []` from fallback provider object and
`createProvider` call; simplified `KeyDiscoveryFailedBadge` to only check
provider-level status (removed per-key status check since keys are no longer
on provider)
- **`routingRuleSheet.tsx`**: `TargetRow` now receives `allKeys` prop (from
`useGetAllKeysQuery`) instead of `providersData` with `.keys`; filters keys
by target provider
- **`routingRuleInfoSheet.tsx`**: New read-only sheet component that displays
routing rule details (conditions, targets with provider icons and weight bars,
fallback chain, scope, priority, timestamps)
- **`settingsPanel.tsx`**: Uses `useGetAllKeysQuery` to determine configured
providers (replaces `p.keys.length > 0` check) and derive
`providerKeyConfigs` per provider
### Other frontend changes (from prior commit, unchanged)
- Added `getProviderKeys`, `getProviderKey` RTK Query endpoints
- Added `createProviderKey`, `updateProviderKey`, `deleteProviderKey` mutations
- Added `buildProviderUpdatePayload` utility for key-free provider updates
- Migrated `providerKeyForm.tsx` to separate create/update mutations
- Updated `addNewKeySheet.tsx` props from `keyIndex` to `keyId`
- Updated all 6 provider form fragments to use `buildProviderUpdatePayload`
- Removed dead `selectedProvider.keys` sync matchers from `providerSlice.ts`
## Type of change
- [x] Feature
- [x] Refactor
- [ ] Bug fix
- [ ] Documentation
- [ ] Chore/CI
## Affected areas
- [ ] Core (Go)
- [ ] Transports (HTTP)
- [ ] Providers/Integrations
- [ ] Plugins
- [x] UI (Next.js)
- [ ] Docs
## How to test
```sh
cd ui
npm run build
npm run lint
```
Manual testing:
1. Navigate to Providers page, select a provider with keys
2. Verify keys table loads correctly from dedicated API
3. Create a new key — verify it appears immediately (no ghost/duplicate)
4. Toggle enable/disable — verify switch updates immediately
5. Edit a key — verify form pre-populates, save works
6. Delete a key — verify it disappears immediately
7. Update provider settings — verify key statuses refresh after save
8. Check sidebar badge shows provider-level discovery failures
9. Open Playground settings — verify provider/key dropdowns work
10. Open Routing Rules — verify target key selector works
11. Click a routing rule row — verify info sheet opens with correct details
(conditions, targets, fallbacks, scope, priority)
## Screenshots/Recordings
N/A — no visual changes to existing features; routing rule info sheet is new.
## Breaking changes
- [ ] Yes
- [x] No
Frontend-only changes consuming the new API shape from PR #2160.
## Related issues
N/A
## Security considerations
No new security considerations. Key values continue to be handled through
existing redaction on the backend.
## Checklist
- [x] I read `docs/contributing/README.md` and followed the guidelines
- [ ] I added/updated tests where appropriate
- [ ] I updated documentation where needed
- [x] I verified builds succeed (Go and UI)
- [ ] I verified the CI pipeline passes locally if applicable
* refactor: replace string slice with WhiteList type for model restrictions (#2282)
## Summary
Refactored model access control logic by replacing string slice with a dedicated `WhiteList` type for the `Models` field in `TableKey`. This change introduces a more structured approach to handling wildcard permissions and improves code readability.
## Changes
- Changed `Models` field type from `[]string` to `schemas.WhiteList` in `TableKey` struct
- Replaced manual wildcard checking (`model == "*"`) with `IsUnrestricted()` method calls across multiple functions
- Added missing mock method `GetVirtualKeyMCPConfigsByMCPClientIDs` to test configuration store
- Applied the refactoring consistently in `ReloadProvider`, `ForceReloadPricing`, and `Bootstrap` methods
## Type of change
- [x] Refactor
- [ ] Bug fix
- [ ] Feature
- [ ] Documentation
- [ ] Chore/CI
## Affected areas
- [x] Core (Go)
- [x] Transports (HTTP)
- [ ] Providers/Integrations
- [ ] Plugins
- [ ] UI (Next.js)
- [ ] Docs
## How to test
Verify that model access control continues to work correctly with both wildcard and specific model permissions:
```sh
# Core/Transports
go version
go test ./...
# Test specific areas affected by the changes
go test ./framework/configstore/tables/...
go test ./transports/bifrost-http/...
```
Test scenarios should include:
- Keys with wildcard permissions (`["*"]`)
- Keys with specific model restrictions
- Keys with empty model lists (deny-by-default behavior)
## Screenshots/Recordings
N/A
## Breaking changes
- [ ] Yes
- [x] No
## Related issues
N/A
## Security considerations
This refactoring maintains the existing security model for API key permissions. The deny-by-default behavior and wildcard functionality remain unchanged, just implemented through a more structured type system.
## Checklist
- [ ] I read `docs/contributing/README.md` and followed the guidelines
- [ ] I added/updated tests where appropriate
- [ ] I updated documentation where needed
- [ ] I verified builds succeed (Go and UI)
- [ ] I verified the CI pipeline passes locally if applicable
* feat: add Plus icon and responsive text to pricing override create button (#2285)
## Summary
Improves the visual design and mobile responsiveness of the pricing overrides section by adding a Plus icon to the create button and optimizing the button text for different screen sizes.
## Changes
- Added Plus icon import from lucide-react
- Enhanced the "Create Override" button with a Plus icon and responsive text that shows "New Override" on larger screens and hides text on mobile
- Adjusted container spacing by removing top margin and changing flex alignment from `items-start` to `items-center` for better visual balance
## Type of change
- [ ] Bug fix
- [x] Feature
- [ ] Refactor
- [ ] Documentation
- [ ] Chore/CI
## Affected areas
- [ ] Core (Go)
- [ ] Transports (HTTP)
- [ ] Providers/Integrations
- [ ] Plugins
- [x] UI (Next.js)
- [ ] Docs
## How to test
Navigate to the custom pricing overrides page and verify:
1. The "New Override" button displays with a Plus icon
2. On mobile screens, only the Plus icon is visible
3. On larger screens (sm and above), both icon and "New Override" text are visible
4. The button functionality remains unchanged when clicked
```sh
# UI
cd ui
pnpm i || npm i
pnpm test || npm test
pnpm build || npm run build
```
## Screenshots/Recordings
Before/after screenshots showing the button design changes and responsive behavior would be helpful.
## Breaking changes
- [x] Yes
- [ ] No
## Related issues
## Security considerations
No security implications - this is a purely visual enhancement.
## Checklist
- [ ] I read `docs/contributing/README.md` and followed the guidelines
- [ ] I added/updated tests where appropriate
- [ ] I updated documentation where needed
- [ ] I verified builds succeed (Go and UI)
- [ ] I verified the CI pipeline passes locally if applicable
* refactor: blacklist models on new convention (#2305)
## Summary
Implements comprehensive blacklist support for model filtering across all providers. This adds the ability to explicitly deny access to specific models at the key level, with blacklist rules taking precedence over allowlist rules.
## Changes
- Added `BlackList` type with semantic validation (supports wildcard "*" for block-all)
- Updated key selection logic to check both allowlist and blacklist constraints
- Modified all provider model listing functions to filter out blacklisted models
- Enhanced UI to support blacklist configuration with improved UX for wildcard selection
- Added blacklist filtering to model catalog and provider handlers
- Updated test cases to verify blacklist functionality
Key design decisions:
- Blacklist always wins over allowlist when conflicts occur
- Wildcard "*" in blacklist blocks all models for that key
- Empty blacklist blocks nothing (permissive default)
- Consistent filtering logic across all providers (Anthropic, Azure, Bedrock, Cohere, etc.)
## Type of change
- [x] Feature
- [ ] Bug fix
- [ ] Refactor
- [ ] Documentation
- [ ] Chore/CI
## Affected areas
- [x] Core (Go)
- [x] Transports (HTTP)
- [x] Providers/Integrations
- [ ] Plugins
- [x] UI (Next.js)
- [x] Docs
## How to test
Test blacklist functionality with provider keys:
```sh
# Core/Transports
go version
go test ./...
# UI
cd ui
pnpm i || npm i
pnpm test || npm test
pnpm build || npm run build
```
Example configuration to test:
```json
{
"keys": [{
"id": "test-key",
"models": ["*"],
"blacklisted_models": ["gpt-4", "claude-3"]
}]
}
```
Verify that blacklisted models are excluded from model listings and key selection.
## Screenshots/Recordings
UI now shows "Blocked Models" field with improved tooltips and wildcard handling for denying access to specific models.
## Breaking changes
- [ ] Yes
- [x] No
The `blacklisted_models` field was already present in the schema but not fully implemented. This change makes it functional without breaking existing configurations.
## Related issues
Enhances model access control capabilities for fine-grained permission management.
## Security considerations
Improves security by allowing explicit denial of access to sensitive or expensive models at the key level. Blacklist rules cannot be bypassed by allowlist configurations.
## Checklist
- [x] I read `docs/contributing/README.md` and followed the guidelines
- [x] I added/updated tests where appropriate
- [x] I updated documentation where needed
- [x] I verified builds succeed (Go and UI)
- [x] I verified the CI pipeline passes locally if applicable
* minor fix add blacklisted model field in tableKeyFromSchemaKey (#2324)
## Summary
This PR adds support for the `BlacklistedModels` field when converting schema keys to table keys in the configuration store's RDB implementation.
## Changes
- Added `BlacklistedModels: key.BlacklistedModels` field mapping in the `tableKeyFromSchemaKey` function
- Ensures that blacklisted model information is properly preserved when converting between schema and table representations
## Type of change
- [ ] Bug fix
- [x] Feature
- [ ] Refactor
- [ ] Documentation
- [ ] Chore/CI
## Affected areas
- [x] Core (Go)
- [ ] Transports (HTTP)
- [ ] Providers/Integrations
- [ ] Plugins
- [ ] UI (Next.js)
- [ ] Docs
## How to test
Verify that configuration keys with blacklisted models are properly stored and retrieved from the RDB configstore.
```sh
# Core/Transports
go version
go test ./...
```
Test creating configuration entries with `BlacklistedModels` specified and ensure they persist correctly through the RDB layer.
## Screenshots/Recordings
N/A
## Breaking changes
- [ ] Yes
- [x] No
## Related issues
N/A
## Security considerations
None - this change only adds field mapping for existing blacklisted models functionality.
## Checklist
- [ ] I read `docs/contributing/README.md` and followed the guidelines
- [ ] I added/updated tests where appropriate
- [ ] I updated documentation where needed
- [ ] I verified builds succeed (Go and UI)
- [ ] I verified the CI pipeline passes locally if applicable
* feat: add image edit input view on logs (#2321)
## Summary
Adds support for logging image edit and image variation requests by introducing new database columns and UI components to track and display these image manipulation operations alongside existing image generation functionality.
## Changes
- Added `image_edit_input` and `image_variation_input` columns to the logs table with corresponding database migrations
- Extended the Log struct with new fields for storing and parsing image edit/variation input data
- Updated logging plugin to capture image edit and variation request data with large payload threshold handling
- Enhanced UI to display input images and prompts for image edit operations and input images for variation operations
- Added image MIME type detection for proper display of base64-encoded images in the UI
## Type of change
- [x] Feature
- [ ] Bug fix
- [ ] Refactor
- [ ] Documentation
- [ ] Chore/CI
## Affected areas
- [x] Core (Go)
- [ ] Transports (HTTP)
- [ ] Providers/Integrations
- [x] Plugins
- [x] UI (Next.js)
- [ ] Do…
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
4 participants
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.
Changes
\nto newline conversion, allowing Starlark's native string escape processing to handle\n,\t, and other escape sequences correctlyType of change
Affected areas
How to test
Test the enhanced Starlark features with MCP CodeMode:
Breaking changes
The Starlark changes are additive and maintain backward compatibility while enabling more Python-like syntax.
Security considerations
Starlark CodeMode maintains its existing sandboxing with no additional network or filesystem access. The dialect enhancements only affect language features within the existing security boundary.