- Add language field to PrCommentFields interface in cf-api and cf-webapp
- For JS/TS PRs, extract first paragraph as a visible summary ("TL;DR")
and collapse the full explanation into a <details> block
- Use correct code block language (typescript vs python) for generated
test code in PR body
- Remove Python-specific comments from type definitions
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Coverage analysis in the Claude pr-review job needs these env vars
to run pytest, matching how django-unit-tests and codeflash-aiservice
workflows configure them.
The ty type checker correctly flags that list[str] is not a subtype of list[str | None] due to list invariance. Added explicit cast.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The JSON parsing path returned the LLM's explicit ranking array,
which sometimes contradicted its own per-dimension scores. Use
_scores_to_ranking() to compute the ranking from weighted scores
when available, falling back to the LLM ranking only when scores
are absent.
The ranker LLM was rewarding candidates that cache global variables
into locals as a performance win. Add an explicit rule: this is only
relevant on Python ≤3.10; on 3.11+ LOAD_GLOBAL uses adaptive
specialization and is nearly as fast as LOAD_FAST.
The non-greedy regex in FIRST_CODE_BLOCK_PATTERN stopped at the first
``` occurrence, even inside triple-quoted strings or nested code fence
blocks. This truncated the extracted code and lost test functions when
LLMs embedded function definitions using ```python:filepath syntax.
Switch to greedy matching and require the closing ``` to be alone on
its line so intermediate backticks are skipped.
The get_or_create defaults passed test lists without positional
indexing, so when a higher test_index created the row first its
content landed at index 0 and was overwritten by the lower index
update, losing a test.
Renamed skill files from using colons to dashes (e.g., tessl:add-api-endpoint → tessl-add-api-endpoint) to fix checkout issues on Windows filesystems which don't allow colons in filenames.
Skills will continue to work as the files contain relative paths to .tessl directory and don't reference their own filenames.
The LLM prompt preprocessing now highlights __init__ signatures for
regular classes, not just @dataclass ones, reducing brute-force
constructor guessing and pytest.skip() fallbacks in generated tests.
log_features() appended test results in call-completion order, causing
model attribution swaps when LLM responses arrived out of order. Pass
test_index through and use positional insertion instead of append.
Replace sticky positioning + ResizeObserver height calc with a flex
column layout (h-screen container, flex-1 panel group) that reliably
fills the viewport. Drop useDefaultLayout hook (not SSR-safe) in favor
of manual localStorage persistence inside useEffect.
Replace the fixed 480px chat overlay with a draggable split-pane layout
using react-resizable-panels, and make tool rounds expandable to show
the actual data the agent retrieved (code, errors, LLM call details).
Remove MAX_TOOL_ROUNDS cap so the model decides when to stop calling
tools. Add a safety net that makes a final tool-free API call if the
loop ends without emitting any visible text, fixing empty assistant
bubbles. Clean up redundant comments.
- Move trace data to top of prompt (long-context best practice: data
before instructions improves quality ~30%)
- Wrap sections in XML tags (<trace_data>, <role>, <domain_knowledge>,
<guidelines>, <use_parallel_tool_calls>) for better parseability
- Remove aggressive language (MUST, CRITICAL, HARD REQUIREMENT) that
causes overtriggering on Opus 4.6
- Replace rigid 4-step investigation workflow with general guidelines
to let adaptive thinking handle reasoning strategy
- Remove duplicate content (tool reference section, two checklists)
- Add <use_parallel_tool_calls> block per Anthropic's recommended pattern
- Tone down tool descriptions from directive to descriptive
- Net reduction: 49 fewer lines in system prompt
Thinking blocks from previous tool rounds (10-50KB each) were
accumulating in conversation history, causing Azure AI Foundry to hang
after 4+ rounds. Redact thinking content before each API call while
preserving required block structure. Also adds per-round timeout safety
net and status indicators between rounds.
Eliminate redundant API call by extracting text from the loop's final
response directly instead of making a separate streaming call. Pre-build
candidatesBySource, candidatesById, and testModelMap in indexTraceData()
to replace repeated O(n) linear searches in tool calls and prompt
building. Combine cost/token aggregation into a single pass.
Restructure agent loop to use stream()+finalMessage() for all API calls,
fixing the SDK's non-streaming timeout error with max_tokens 32k. Add
parallel tool execution, tool activity bubbles in the frontend, and
restructure the system prompt for better investigation behavior.
Guide the chat agent to use the new tools proactively: a DEBUGGING TOOLS
section with structured guidance for get_llm_call_detail and codebase
browsing, a 4-step workflow (OBSERVE → INVESTIGATE → LOCATE → RECOMMEND),
and a RESPONSE CHECKLIST at the end of the prompt requiring the agent to
cite real file paths before responding.
Give the observability chat agent four new tools: get_llm_call_detail
(full prompt/response for any LLM call), read_file, search_code, and
list_directory for navigating the codeflash-internal and codeflash CLI
repos. This lets the agent trace problems end-to-end from trace data
through actual prompts to pipeline source code.
- Add id to IndexedTraceData.llmCalls so the agent can reference calls
- Make resolveToolCall async (Prisma + fs + child_process)
- Make processToolUseResponse async to match
- Bump MAX_TOOL_ROUNDS from 5 to 15 for multi-step code browsing
- Add CODEFLASH_INTERNAL_REPO_PATH / CODEFLASH_CLI_REPO_PATH env vars
- Path traversal protection, file size caps, search result limits
