The agent-cli binary was a thin wrapper over the agent library with
nothing else depending on it. Moving it into the agent crate as a
`pub mod cli` with a `[[bin]]` entry reduces workspace complexity
and follows the attractor crate pattern.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The agent's Turn::Assistant decomposed responses into text, tool_calls,
and reasoning fields, discarding ContentPart::Other items. This lost
OpenAI reasoning items needed for Responses API round-tripping.
Add provider_parts field to Turn::Assistant to carry opaque content
parts through the history, and emit them first in convert_to_messages
so they precede function_call items as the API requires.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Introduce `ToolApprovalFn` callback in `SessionConfig` to gate tool
execution by permission level. Create `agent-cli` crate as a thin CLI
binary wrapping `Session` with provider/model resolution, permission
model (read-only/read-write/full), interactive approval prompts,
real-time event rendering, debug middleware, and SIGINT handling.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The Responses API requires that reasoning items (type: "reasoning",
id: "rs_xxx") are included alongside their associated function_call
items when replaying conversation history. Without them:
"function_call was provided without its required 'reasoning' item"
Store reasoning output items as ContentPart::Other and replay them
as raw input items in translate_input. Handles both non-streaming
(parse_output) and streaming (handle_output_item_done) paths.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The Responses API returns two distinct IDs on function_call items:
- `id` (item-level, starts with `fc_`)
- `call_id` (call-level, starts with `call_`)
When sending function calls back as input, the `id` field must start
with `fc_`. Previously we used `call_id` for both fields, causing:
"Invalid 'input[1].id': Expected an ID that begins with 'fc'."
Now we preserve the item-level `id` in provider_metadata and use it
for the `id` field, while `call_id` continues to link tool results.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Covers two recent runtime failures that lacked test coverage:
- Anthropic: assert deprecated beta header values are not sent
- Gemini: test function call parsing/translation with and without thoughtSignature
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Gemini 3 models include a thoughtSignature field on function call
parts that must be returned in subsequent conversation turns. Add
provider_metadata to ToolCall to carry this through, and extract/emit
it in both the non-streaming and streaming Gemini code paths.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The `extended-thinking-2025-04-14` and `max-tokens-3-5-sonnet-2025-04-14`
beta headers are no longer accepted by the Anthropic API.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Move the ullm binary from a separate crate into unified-llm as
src/bin/ullm.rs, consolidating CLI dependencies into the library crate.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Stores DOT source in ManagedPipeline and pipes it through `dot -Tsvg`
on request. Returns image/svg+xml on success, 502 if graphviz is
unavailable, 404 if pipeline not found. Resolves spec gap #1.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The SSE stream terminated prematurely when a network chunk contained
only unhandled event types (e.g. response.in_progress, reasoning
output_item.added). dispatch_sse_messages returned an empty vec, which
the unfold interpreted as stream-end. Now continues reading when
dispatch produces no StreamEvents.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add a [[bin]] target to the attractor crate with two subcommands:
- `attractor validate <pipeline.dot>` -- parse and validate only
- `attractor run <pipeline.dot>` -- full pipeline execution with LLM backend
The CLI supports --dry-run, --auto-approve, --resume, --model, --provider,
and two-level verbosity (-v one-line summaries, -vv full event details).
Extracts a shared `default_registry()` function in handler/mod.rs so both
the CLI and server can build a fully-wired HandlerRegistry without
duplicating handler registration boilerplate.
The AgentBackend in cli/backend.rs implements CodergenBackend by creating
a coding-agent-loop Session per node invocation, giving LLM nodes access
to file and shell tools.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Closes the subgraph class derivation test gap by adding a parse() pipeline
test that verifies DOT subgraph labels produce correct CSS-like classes on
contained nodes. Updates gap analysis to remove resolved items, add spec
contradictions section, and narrow remaining gaps to SVG rendering and
retry predicate customization.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Gap 1: Real preamble synthesis per fidelity mode (truncate, compact,
summary:low/medium/high) in new preamble.rs module, replacing placeholder
- Gap 2: Pass thread_id to CodergenBackend.run() so backends can reuse
LLM sessions across nodes sharing the same thread
- Gap 7: Engine cancellation via AtomicBool token checked between nodes,
wired to server cancel endpoint, new Cancelled error variant
- Gap 8: RecordingInterviewer serialization (to_json/from_json, file I/O)
and new ReplayInterviewer for replaying recorded Q&A sessions
- Gap 12: Preset retry policies selectable from DOT via retry_policy attr
(none, standard, aggressive, linear, patient)
- Gap 15: Engine calls Interviewer.inform() at pipeline start, stage
start, and stage complete lifecycle points
37 new tests (450 unit + 49 integration, all passing).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Change server registry_factory to accept Arc<dyn Interviewer> so WaitHumanHandler
shares the same WebInterviewer as the REST API endpoints
- Add full HTTP lifecycle tests: approve-and-complete flow + cancel flow
- Add SSE event stream content parsing test with frame-level verification
- Add sub-pipeline E2E test through the engine with context propagation
- Add manager loop E2E test with SimulatingChildObserver
- Add graph merge E2E test verifying module prefixing and execution ordering
- Add 3 real LLM tests (#[ignore]) using claude-haiku via AnthropicAdapter
- Add dotenvy and http-body-util dev dependencies
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Eliminate ~300 lines of duplicated test infrastructure:
- Add MockExecutionEnvironment::linux() constructor, replacing 4
identical linux_env() helpers across profile test modules
- Extend MockExecutionEnvironment with written_files, captured_timeout,
and apply_read_offset_limit fields to replace 4 specialized mocks
(ReadFileEnv, WriteFileEnv, EditFileEnv, ShellCapturingEnv) in tools.rs
- Add MutableMockExecutionEnvironment for apply_patch tests that need
writes visible to subsequent reads, replacing MockFileEnv in openai.rs
- Merge ParallelTestProfile into TestProfile with configurable
parallel_tool_calls and context_window fields
- Replace ProviderTestProfile with TestProfile in provider_profile.rs
tests, updating TestProfile::build_system_prompt to include user
instructions like real profiles
- Extract shared CapturingLlmProvider into test_support.rs, replacing
two inline capturing provider structs in session.rs
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Proposal 7.1: Introduce required_str() helper in tools.rs to replace ~15
repetitions of the args.get("param").and_then(|v| v.as_str()).ok_or_else(...)
pattern across tools.rs and subagent.rs. Provides consistent error messages.
- Proposal 4.2: Simplify SubAgentManager::spawn success path by using ? on
process_input() result directly, eliminating the redundant success variable
(always true on the Ok path) and the if-let-Err-return pattern.
- Proposal 8.2: Mark SubAgent::depth() and SubAgentManager::get() as
#[cfg(test)] since they are only used in tests. Remove SubAgent::id()
entirely as it was unused even in tests.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- 2.2: Introduce BaseProfile struct with shared id/model/registry fields;
AnthropicProfile, GeminiProfile, OpenAiProfile now delegate to it
- 5.4: ProviderProfile::id() and ::model() return &str instead of owned String,
eliminating unnecessary heap allocations on every call
- 6.1: Rename EnvContext fields: date -> current_date, model_name -> model
for consistency with trait method names
- 8.1: Mark build_env_context_block (no-context variant) as #[cfg(test)]
since it is only used in one test
- 9.1: Convert SubAgentManager methods (spawn, send_input, wait, close) from
Result<T, String> to Result<T, AgentError>; tool executors convert at boundary
via .map_err(|e| e.to_string())
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Remove History::new() (redundant with #[derive(Default)]), replace all callers with History::default()
- Replace match with let-else in extract_signatures_from_assistant for clearer happy path
- Add doc comments to Turn::System and Turn::Steering explaining their LLM role mapping
- Verified: Arc import in openai.rs is used in production code, mutex .expect() messages are consistent
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Replace loop/match/break with clearer let-else pattern for followup
queue processing in process_input (proposal 4.1)
- Simplify validate_tool_args empty schema check by separating null
and empty-object checks into distinct if-blocks (proposal 4.3)
- Cache system prompt once per input cycle in run_single_input and pass
it to build_request and check_context_usage/estimate_token_count,
avoiding redundant string construction on every tool round (5.2, 5.3)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add SubPipelineHandler (handler/sub_pipeline.rs) that inline-executes a parsed
sub-graph within the same engine, reading DOT source from node attributes and
propagating context diffs back to the parent pipeline.
Add GraphMergeTransform (transform.rs) that merges nodes and edges from secondary
graphs into a primary graph with namespace-prefixed IDs to avoid collisions.
Add WebInterviewer (interviewer/web.rs) backed by oneshot channels for async
question/answer flow, and HTTP server (server.rs) with 8 axum endpoints behind
a "server" feature flag for pipeline management and human-in-the-loop via web.
Fix tempfile dev-dependency usage in server production code by using std::env::temp_dir.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Create comprehensive READMEs for unified-llm, unified-llm-cli, and
coding-agent-loop crates, and expand the attractor README from a
one-liner into full documentation covering key concepts, API usage,
and code examples.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Adds the attractor crate implementing all 11 sections of the attractor spec:
- DOT parser (lexer, grammar, semantic analysis) for strict DOT subset
- Pipeline execution engine with edge selection, goal gates, retry logic,
failure routing, checkpoint save/resume, and loop_restart
- 9 node handlers: start, exit, codergen, wait_human, conditional, parallel
(concurrent with join/error policies), fan_in (with LLM eval), tool, manager_loop
- State management: PipelineContext, Outcome, Artifact store, fidelity resolution
- Human-in-the-loop: Interviewer trait with auto_approve, callback, queue,
recording, and console implementations, plus timeout enforcement
- Validation: 14 built-in lint rules with custom rule registration API
- Model stylesheet with universal/shape/class/ID selectors and specificity
- Transforms: variable expansion, stylesheet application, preamble; plus
PipelineBuilder with register_transform and prepare_pipeline
- Condition expression language with =, !=, bare-key truthiness, && combinator
- Event system with all 16 event types emitted by engine and handlers
- Tool call hooks (pre/post) for CodergenHandler
- Run directory with manifest.json and per-node status.json
370 tests (354 unit + 16 integration) covering all spec sections.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Extract shared test infrastructure (MockExecutionEnvironment, TestProfile,
MockLlmProvider) replacing 11 duplicate mock implementations across tests
- Deduplicate tool execution logic between sequential and parallel paths
- Narrow ProviderProfile trait from 14 to 7 required methods via
ProfileCapabilities struct and default implementations
- Replace stringly-typed HashMap event data with typed EventData enum
- Extract shared assemble_system_prompt helper and register_subagent_tools
default method, eliminating copy-paste across all 3 profiles
- Replace fragile shell-based glob with glob crate, fix rg detection
- Add delete_file to ExecutionEnvironment, wire git context into env block
- Remove dead code (AgentError::Io, count_turns, trivial derived-trait tests)
- Use match-based lookups in truncation instead of per-call HashMap allocation
Net reduction: -1,401 lines across 20 files. All 180 tests pass.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add knowledge_cutoff() method to ProviderProfile trait so session can
populate EnvContext from the profile instead of leaving it empty
- Set subagent default max_turns to 50 per spec (was using session
factory default which could be 0/unlimited)
- Add corrected spec compliance review after manual verification found
the initial 5-agent review was largely false positives
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Fix all gaps identified by spec review against docs/specs/coding-agent-loop-spec.md:
- Replace placeholder system prompts with substantial provider-aligned prompts
for OpenAI (codex-rs style), Anthropic (Claude Code style), and Gemini
(gemini-cli style) covering identity, tool usage, and coding best practices
- Fix ExecutionEnvironment trait: add offset/limit to read_file, depth to
list_directory, path to glob, remove separate args from exec_command
- LocalEnv: use /bin/bash -c, spawn process groups with setsid, SIGTERM to
-pid, use ripgrep with grep fallback, add env var safelist
- Add environment context block with <environment> XML tags including git
branch, date, model, and knowledge cutoff fields
- Capture git context snapshot (branch, status, recent commits) on init
- Add user_instructions to SessionConfig, appended as final prompt layer
- Emit AssistantTextStart before LLM calls, SESSION_END on abort path
- Fix truncation messages to match spec wording exactly
- Implement provider_options() for all profiles (reasoning, beta headers,
safety settings)
- Add Gemini-specific tools: read_many_files, list_dir, web_search, web_fetch
- Wire spawn_agent max_turns parameter
- Fall back to working_dir for project doc discovery outside git repos
All 188 tests pass.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Close spec compliance gaps:
- Add user_instructions to SessionConfig and ProviderProfile trait (spec Section 6.1 layer 5)
- Populate EnvContext with git branch, date, model name during Session::initialize()
- Expand all three profile system prompts with identity, tool guidance, coding practices
- Implement provider_options: Anthropic beta headers, OpenAI reasoning effort, Gemini safety settings
- Add register_subagent_tools() to all profiles for spawn_agent/send_input/wait/close_agent
- Add list_dir tool to Gemini profile (spec Section 3.6)
- Add chrono dependency for date formatting
20 new tests (170 → 190), all passing.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Abort now transitions to CLOSED state and returns Err(Aborted)
- Closed sessions no longer emit SessionStart before rejecting input
- Add set_reasoning_effort() for mid-session reasoning effort changes
- Fix spawn_agent truncation limit from 30k to spec-required 20k
- Add JSON Schema validation of tool arguments before execution
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
resolve_model() was passing raw alias strings (e.g. "gpt5") directly to
APIs instead of resolving them to actual model IDs (e.g. "gpt-5.2").
Also rename gpt-5.2-mini to gpt-5-mini, which is the correct OpenAI
model ID.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace all stub tools in profiles with real make_*_tool() factories,
wire up project docs discovery in session, add missing events
(SessionStart, SteeringInjected, Error), enrich environment context
block, fix per-tool truncation modes, improve loop detection to check
all groups, add SIGTERM-before-SIGKILL on timeout, and update subagent
with SubAgentResult struct.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Extend the Middleware trait with process_stream_event() for event-level
observation/transformation of streaming responses. Add
wrap_stream_with_middleware() helper. Re-export set_default_client at the
crate root per spec Section 2.5.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
New crate `unified-llm-cli` with binary `ullm` providing:
- `prompt` command: generate text via streaming/non-streaming, with
system prompts, options (-o temperature=0.5), stdin piping, and
token usage display
- `models list` command: browse catalog models with --provider and
--query filters
- Auto-detects provider from catalog for correct API routing
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Enforce stream_read timeout (30s default) in all 4 providers' streaming code
- Add with_timeout() builder method to all adapter constructors
- Fix ResponseFormatType::JsonObject to serialize as "json" per spec
- Add STEP_FINISH to StreamEventType enum in spec doc
- Add UnsupportedToolChoice error and enforce in all adapters via validate_tool_choice()
- Fix error classification to check status code before message content
- Add stop_sequences support to OpenAI Responses API adapter
- Handle Gemini thought parts (thought: true) in both complete and streaming paths
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- OpenAI adapter: include is_error flag on function_call_output items
- Anthropic adapter: extract retry_after from headers in streaming error path
- Error retryability: unknown errors now default to retryable per spec
- stream_object: add ObjectStreamResult wrapper with object() accessor
- TimeoutConfig: add From<f64> for total-only timeout shorthand
- Message::tool_result: accept serde_json::Value to preserve structured content
- GenerateParams: expose repair_tool_call field wired to execute_all_tools_with_repair
- Both generate() and stream() tool loops use repair-aware execution
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Anthropic adapter estimates reasoning_tokens from thinking block text lengths
- Add tool call validation against JSON schema + repair_tool_call callback
- Call adapter.initialize() on provider registration
- Extract model catalog to catalog.json data file loaded via include_str!
- Add ContentPart::Other variant for unknown/extensible content kinds
- Change StreamEvent::Error field from String to SdkError
- Add ToolContext (tool_call_id, messages, abort_signal) to execute handlers
- Gemini adapter uses gRPC status codes from error bodies for classification
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add Audio/Document content part handling across all providers:
Anthropic supports documents natively, Gemini supports both audio
and documents, OpenAI and OpenAI-compatible produce text fallbacks
- Add stop_when support to streaming tool loops (was only in generate())
- Add retry on initial stream connection (matching generate() behavior)
- Add total and per_step timeout support to streaming tool loops
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
