feat(#106): Planner agent + TaskSpec + auto-inference — Phase B backend#121
Merged
Abernaughty merged 6 commits intomainfrom Apr 5, 2026
Merged
feat(#106): Planner agent + TaskSpec + auto-inference — Phase B backend#121Abernaughty merged 6 commits intomainfrom
Abernaughty merged 6 commits intomainfrom
Conversation
Phase B Step 1-3: Planner agent module with: - TaskSpec Pydantic model for structured task specifications - PlannerSession for managing conversational planning sessions - ChecklistStatus for tracking readiness checklist completion - infer_workspace_stack() for auto-detecting languages/frameworks - Planner LLM integration (PLANNER_MODEL env var, zero tools) - Comprehensive unit tests for all planner logic
- EventType.PLANNER_MESSAGE for streaming planner responses via SSE - PlannerStartRequest, PlannerMessageRequest for API endpoints - PlannerSessionResponse, PlannerSubmitResponse for responses - PlannerChecklistItemResponse, PlannerChecklistResponse for readiness - PlannerTaskSpecResponse for structured task spec display
Three new endpoints for the Planner conversational flow:
- POST /tasks/plan — start a session with workspace auto-inference
- POST /tasks/plan/{id}/message — send message to Planner agent
- POST /tasks/plan/{id}/submit — confirm and create task from TaskSpec
Includes workspace validation, protected workspace PIN checks,
PLANNER_MESSAGE SSE events, and submission guardrails (required
fields must be satisfied, workspace never overridden by LLM).
📝 WalkthroughWalkthroughAdds a Planner agent for conversational task-spec refinement: workspace manifest inference, checklist validation, TTL session management, LLM-driven structured updates, an in-memory session store, SSE events, three API endpoints (create/message/submit), and tests covering inference, checklist, JSON handling, and sessions. Changes
Sequence Diagram(s)sequenceDiagram
actor Client
participant API as API Server
participant Planner as Planner Agent
participant LLM as LLM Service
participant Store as Session Store
participant EventBus as Event Bus
Client->>API: POST /tasks/plan (workspace)
API->>Planner: infer_workspace_stack(workspace)
Planner-->>API: {languages, frameworks}
API->>Planner: create_planner_session(...)
Planner->>Store: store session
Planner-->>API: PlannerSessionResponse
Client->>API: POST /tasks/plan/{id}/message
API->>Store: get session
API->>Planner: send_planner_message(session, message)
Planner->>LLM: call model with spec + checklist
LLM-->>Planner: text + JSON block
Planner->>Planner: extract JSON, apply updates, rebuild checklist
Planner-->>API: PlannerResponse (ready, warnings)
API->>EventBus: publish PLANNER_MESSAGE
EventBus-->>Client: SSE event
Client->>API: POST /tasks/plan/{id}/submit
API->>Store: get & validate session
API->>Planner: verify required_satisfied
API->>API: state_manager.create_task + task_runner.submit
API->>Store: mark session submitted
API-->>Client: PlannerSubmitResponse
Estimated code review effort🎯 4 (Complex) | ⏱️ ~45 minutes Possibly related issues
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 |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 5
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@dev-suite/src/agents/planner.py`:
- Around line 271-315: The heuristics currently use substring matches and weak
file presence checks which cause false positives (e.g., matching "preact" as
React via _PACKAGE_JSON_FRAMEWORK_MAP and treating any build.gradle as Kotlin);
update the inference logic to use exact package-name matching for package.json
deps (compare dependency keys exactly against keys in
_PACKAGE_JSON_FRAMEWORK_MAP, allow scoped-prefix exceptions explicitly like
"react-" if intended), change pyproject checks to require whole-token matches
against _PYPROJECT_FRAMEWORK_MAP, and tighten Kotlin detection so build.gradle
alone does not cause "Kotlin" to be appended—only add Kotlin when a
build.gradle.kts exists or when build.gradle contains explicit Kotlin DSL/plugin
coordinates or the Kotlin JVM/Android plugin identifiers; keep references to the
same variables (languages, frameworks, workspace) and ensure build_checklist()
receives only strongly-evidenced entries.
- Around line 549-564: The _apply_spec_updates function currently writes any
parsed JSON value into TaskSpec fields, which allows wrong types (e.g.,
"languages": "Python") to corrupt the spec; add runtime validation before
mutating: for each field in updates (inside _apply_spec_updates) check the
expected TaskSpec field type/shape (use TaskSpec.model_fields or
TaskSpec.__fields__ to get expected types), coerce or reject mismatched types
(e.g., require list for languages, dict/structured types for
acceptance_criteria), log or skip invalid updates, and only setattr when the
value matches the expected schema; reference TaskSpec, _apply_spec_updates,
build_checklist(), and TaskSpec.to_description() to ensure incoming values will
be compatible with downstream consumers.
In `@dev-suite/src/api/main.py`:
- Around line 450-456: Before calling state_manager.create_task/
task_runner.submit with session.task_spec, re-validate the workspace policy for
session.task_spec.workspace (i.e., run the same allow-list and PIN checks that
were used at session open) or verify a carried PlannerSession auth flag is still
valid; if the re-check fails, abort and return an error instead of creating the
task. Locate the usage around session.task_spec.to_description(),
state_manager.create_task and task_runner.submit and either invoke the existing
workspace policy check function (the one used when opening a session) or consult
and invalidate a stored PlannerSession auth state before proceeding.
In `@dev-suite/tests/test_planner.py`:
- Around line 307-312: The test test_ignore_unknown_fields is currently vacuous
because the assertion uses "or True"; update the assertion in that test to
actually verify unknown fields are not set: remove the "or True" and assert that
hasattr(spec, "unknown_field") is False (or use "assert not hasattr(spec,
'unknown_field')"), ensuring _apply_spec_updates and TaskSpec behavior is
exercised; locate the test function test_ignore_unknown_fields and replace the
final assertion accordingly so it fails if unknown fields are applied.
- Around line 9-28: The import block includes unused names (Path,
ChecklistPriority, PlannerSession) and is not sorted, causing Ruff I001/F401;
remove the unused imports (drop Path, ChecklistPriority, PlannerSession) from
the import list and reorder the remaining imports into a properly sorted group
(standard library first like json, time, pathlib if kept, then third-party like
pytest, then local module imports) or run isort/ruff --fix to normalize
ordering; ensure the test references still match the retained symbols such as
AsyncMock, patch, pytest, and the planner symbols that are actually used (e.g.,
TaskSpec, _apply_spec_updates, build_checklist, create_planner_session, etc.).
🪄 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: defaults
Review profile: CHILL
Plan: Pro
Run ID: b5a3b371-c820-4886-ae50-b35b6fedb694
📒 Files selected for processing (5)
dev-suite/src/agents/planner.pydev-suite/src/api/events.pydev-suite/src/api/main.pydev-suite/src/api/models.pydev-suite/tests/test_planner.py
1. test_planner.py: Remove unused imports (Path, ChecklistPriority, PlannerSession) fixing ruff F401/I001 CI failure 2. test_planner.py: Fix vacuous assertion (remove `or True`) 3. test_planner.py: Add tests for type coercion and preact/gradle edge cases 4. planner.py: Use exact dep matching in infer_workspace_stack — prevents "preact" → React false positive 5. planner.py: Tighten Kotlin detection — only from .kts, not plain build.gradle 6. planner.py: Add type validation in _apply_spec_updates — coerce string→list, reject non-string/non-list values 7. planner.py: Move `import re` to top-level (was lazy-imported twice)
Addresses CodeRabbit Major: submit_planner_session now re-checks ws_mgr.is_allowed() and ws_mgr.is_protected() before creating the task. Prevents a session opened before workspace removal/protection from submitting into that directory.
There was a problem hiding this comment.
♻️ Duplicate comments (1)
dev-suite/tests/test_planner.py (1)
15-25:⚠️ Potential issue | 🟡 MinorFix import sorting to resolve CI failure.
The ruff linter is failing with
I001because imports within thesrc.agents.plannerblock are not sorted according to ruff's isort rules (typically case-insensitive alphabetical order).🧹 Proposed fix
from src.agents.planner import ( - PlannerSessionStore, - TaskSpec, _apply_spec_updates, _extract_task_spec_updates, _strip_json_block, build_checklist, create_planner_session, infer_workspace_stack, + PlannerSessionStore, send_planner_message, + TaskSpec, )Alternatively, run
ruff check --fixorruff formatto auto-fix.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@dev-suite/tests/test_planner.py` around lines 15 - 25, The import block from src.agents.planner is not sorted which triggers ruff I001; reorder the imported symbols (PlannerSessionStore, TaskSpec, _apply_spec_updates, _extract_task_spec_updates, _strip_json_block, build_checklist, create_planner_session, infer_workspace_stack, send_planner_message) into ruff/isort-compliant alphabetical order (case-insensitive) or run `ruff check --fix`/`ruff format` to auto-fix the imports so the test file passes CI.
🧹 Nitpick comments (2)
dev-suite/src/agents/planner.py (2)
797-841: Consider thread safety if scaling beyond single-worker deployment.
PlannerSessionStoreuses a plain dict which is not thread-safe for concurrent modifications. This is fine for:
- Single-worker asyncio (default uvicorn mode)
- Multiple workers (each has isolated memory)
If you later need shared session state across workers or use threaded concurrency, consider using
threading.Lockor migrating to Redis/external store.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@dev-suite/src/agents/planner.py` around lines 797 - 841, PlannerSessionStore is not thread-safe because it mutates _sessions without synchronization; add a threading.Lock (e.g. self._lock = threading.Lock()) and acquire/release it (use context manager) around all accesses and mutations of self._sessions in create, get, remove, count, and _cleanup_expired to prevent concurrent races, or alternatively replace the in-memory dict with a process-shared store (e.g. Redis) and update create/get/remove/count/_cleanup_expired to use that backend if cross-worker sharing is required.
373-385: Semantic note onauto_inferredflag.The
auto_inferredflag is set toTruewheneverlanguagesorframeworkshave values, even if the user manually provided them. This is technically imprecise—it only indicates "has values" rather than "values came from auto-inference."For the MVP this is acceptable since it's primarily for UI hints, but consider tracking the actual inference source if future features need to distinguish user-provided vs auto-detected values.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@dev-suite/src/agents/planner.py` around lines 373 - 385, The auto_inferred flag for the ChecklistItem entries for "languages" and "frameworks" incorrectly treats any non-empty value as auto-inferred; update the code that builds those items (the ChecklistItem construction for fields "languages" and "frameworks" in planner.py) to read a true inference indicator from the TaskSpec instead of using bool(task_spec.languages/frameworks). Add or use explicit flags on TaskSpec (e.g., task_spec.inferred_languages and task_spec.inferred_frameworks or similar properties), set auto_inferred=task_spec.inferred_languages for the languages item and auto_inferred=task_spec.inferred_frameworks for the frameworks item, and fall back to False if those inference flags are not present to preserve current behavior.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Duplicate comments:
In `@dev-suite/tests/test_planner.py`:
- Around line 15-25: The import block from src.agents.planner is not sorted
which triggers ruff I001; reorder the imported symbols (PlannerSessionStore,
TaskSpec, _apply_spec_updates, _extract_task_spec_updates, _strip_json_block,
build_checklist, create_planner_session, infer_workspace_stack,
send_planner_message) into ruff/isort-compliant alphabetical order
(case-insensitive) or run `ruff check --fix`/`ruff format` to auto-fix the
imports so the test file passes CI.
---
Nitpick comments:
In `@dev-suite/src/agents/planner.py`:
- Around line 797-841: PlannerSessionStore is not thread-safe because it mutates
_sessions without synchronization; add a threading.Lock (e.g. self._lock =
threading.Lock()) and acquire/release it (use context manager) around all
accesses and mutations of self._sessions in create, get, remove, count, and
_cleanup_expired to prevent concurrent races, or alternatively replace the
in-memory dict with a process-shared store (e.g. Redis) and update
create/get/remove/count/_cleanup_expired to use that backend if cross-worker
sharing is required.
- Around line 373-385: The auto_inferred flag for the ChecklistItem entries for
"languages" and "frameworks" incorrectly treats any non-empty value as
auto-inferred; update the code that builds those items (the ChecklistItem
construction for fields "languages" and "frameworks" in planner.py) to read a
true inference indicator from the TaskSpec instead of using
bool(task_spec.languages/frameworks). Add or use explicit flags on TaskSpec
(e.g., task_spec.inferred_languages and task_spec.inferred_frameworks or similar
properties), set auto_inferred=task_spec.inferred_languages for the languages
item and auto_inferred=task_spec.inferred_frameworks for the frameworks item,
and fall back to False if those inference flags are not present to preserve
current behavior.
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro
Run ID: f08a9345-e23e-460c-8f51-4274f8863c65
📒 Files selected for processing (2)
dev-suite/src/agents/planner.pydev-suite/tests/test_planner.py
There was a problem hiding this comment.
Actionable comments posted: 2
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@dev-suite/src/api/main.py`:
- Around line 376-381: The exception message returned to the client leaks
internal details; keep the existing logger.error with exc_info=True for full
diagnostics but replace the client-facing _error call so it sends a generic
message (e.g., "Planner failed" or "Internal planner error") and the same 502
status without including the exception object; update the except block around
send_planner_message to call _error with a sanitized string while still logging
the full exception via logger.error and then return.
- Around line 462-464: The planner submission path currently calls
state_manager.create_task(...) and task_runner.submit(task_id, description,
workspace=workspace) but omits the publish_pr flag that the regular POST /tasks
path forwards; update the planner flow to accept and propagate publish_pr (e.g.
add publish_pr to PlannerStartRequest or the planner handler's params) and pass
it into task_runner.submit(task_id, description, workspace=workspace,
publish_pr=publish_pr) so planner-created tasks honor the same PR publishing
option as the regular endpoint.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
Comment on lines
+376
to
+381
| try: | ||
| planner_resp = await send_planner_message(session, body.message) | ||
| except Exception as e: | ||
| logger.error("Planner LLM call failed: %s", e, exc_info=True) | ||
| _error(f"Planner failed: {e}", 502) | ||
| return |
There was a problem hiding this comment.
Sanitize exception details before returning to client.
The error message f"Planner failed: {e}" may leak internal details (stack traces, file paths, or configuration info) from the underlying LLM call or JSON extraction. Consider a generic message for the client and log the full exception separately.
🛡️ Suggested fix
try:
planner_resp = await send_planner_message(session, body.message)
except Exception as e:
logger.error("Planner LLM call failed: %s", e, exc_info=True)
- _error(f"Planner failed: {e}", 502)
+ _error("Planner service temporarily unavailable. Please try again.", 502)
return📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
Suggested change
| try: | |
| planner_resp = await send_planner_message(session, body.message) | |
| except Exception as e: | |
| logger.error("Planner LLM call failed: %s", e, exc_info=True) | |
| _error(f"Planner failed: {e}", 502) | |
| return | |
| try: | |
| planner_resp = await send_planner_message(session, body.message) | |
| except Exception as e: | |
| logger.error("Planner LLM call failed: %s", e, exc_info=True) | |
| _error("Planner service temporarily unavailable. Please try again.", 502) | |
| return |
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@dev-suite/src/api/main.py` around lines 376 - 381, The exception message
returned to the client leaks internal details; keep the existing logger.error
with exc_info=True for full diagnostics but replace the client-facing _error
call so it sends a generic message (e.g., "Planner failed" or "Internal planner
error") and the same 502 status without including the exception object; update
the except block around send_planner_message to call _error with a sanitized
string while still logging the full exception via logger.error and then return.
Comment on lines
+462
to
+464
| # Create task via existing flow | ||
| task_id = await state_manager.create_task(description, workspace=workspace) | ||
| task_runner.submit(task_id, description, workspace=workspace) |
There was a problem hiding this comment.
Missing publish_pr parameter for planner-submitted tasks.
The regular POST /tasks endpoint at line 201 passes publish_pr=body.publish_pr to task_runner.submit(), but the planner submit path omits it. Planner-created tasks will always use the default PR behavior with no user control.
Consider adding publish_pr to the submit request or PlannerStartRequest, or document that planner tasks always use the default.
🔧 Suggested fix if adding to submit endpoint
+from .models import PlannerSubmitRequest # Add new request model
`@app.post`("/tasks/plan/{session_id}/submit", response_model=ApiResponse)
async def submit_planner_session(
session_id: str,
+ body: PlannerSubmitRequest | None = None,
_auth: str | None = Depends(require_auth),
):
# ... existing validation ...
# Create task via existing flow
task_id = await state_manager.create_task(description, workspace=workspace)
- task_runner.submit(task_id, description, workspace=workspace)
+ task_runner.submit(
+ task_id,
+ description,
+ workspace=workspace,
+ publish_pr=body.publish_pr if body else None,
+ )🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@dev-suite/src/api/main.py` around lines 462 - 464, The planner submission
path currently calls state_manager.create_task(...) and
task_runner.submit(task_id, description, workspace=workspace) but omits the
publish_pr flag that the regular POST /tasks path forwards; update the planner
flow to accept and propagate publish_pr (e.g. add publish_pr to
PlannerStartRequest or the planner handler's params) and pass it into
task_runner.submit(task_id, description, workspace=workspace,
publish_pr=publish_pr) so planner-created tasks honor the same PR publishing
option as the regular endpoint.
Single blank line between import block and first code, not two.
There was a problem hiding this comment.
Actionable comments posted: 1
🧹 Nitpick comments (1)
dev-suite/tests/test_planner.py (1)
386-389: UseSESSION_TTL_SECONDSin expiry tests instead of hard-coded3600.These assertions are coupled to an implicit 30-minute policy but encode 1 hour directly. If TTL changes, tests can become misleading. Prefer deriving from the source constant.
Proposed diff
from src.agents.planner import ( PlannerSessionStore, + SESSION_TTL_SECONDS, TaskSpec, @@ def test_session_expiry(self): session = create_planner_session(workspace="/proj") - session.last_activity = time.time() - 3600 # 1 hour ago + session.last_activity = time.time() - (SESSION_TTL_SECONDS + 1) assert session.is_expired @@ def test_expired_session_removed(self): store = PlannerSessionStore() session = create_planner_session(workspace="/proj") - session.last_activity = time.time() - 3600 + session.last_activity = time.time() - (SESSION_TTL_SECONDS + 1) store.create(session) assert store.get(session.session_id) is NoneAlso applies to: 412-417
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@dev-suite/tests/test_planner.py` around lines 386 - 389, Update the expiry tests to compute the stale timestamp using the module constant SESSION_TTL_SECONDS instead of the hard-coded 3600; specifically, in test_session_expiry (and the similar test around lines for session inactivity) set session.last_activity = time.time() - (SESSION_TTL_SECONDS + 1) (or - SESSION_TTL_SECONDS to test boundary) so the assertion session.is_expired derives from SESSION_TTL_SECONDS, and import or reference the constant from the planner/session module where create_planner_session and session.is_expired are defined.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@dev-suite/tests/test_planner.py`:
- Around line 391-396: The test_session_touch uses time.sleep which causes
flakiness; instead, set session.last_activity to a deterministic older timestamp
and call session.touch(), then assert session.last_activity is greater than the
previous value. Modify test_session_touch to create the session via
create_planner_session(), assign session.last_activity to a past datetime (or
timestamp) value, call session.touch(), and assert the updated
session.last_activity > previous value (referencing the test function name
test_session_touch and methods create_planner_session, session.touch, and
session.last_activity).
---
Nitpick comments:
In `@dev-suite/tests/test_planner.py`:
- Around line 386-389: Update the expiry tests to compute the stale timestamp
using the module constant SESSION_TTL_SECONDS instead of the hard-coded 3600;
specifically, in test_session_expiry (and the similar test around lines for
session inactivity) set session.last_activity = time.time() -
(SESSION_TTL_SECONDS + 1) (or - SESSION_TTL_SECONDS to test boundary) so the
assertion session.is_expired derives from SESSION_TTL_SECONDS, and import or
reference the constant from the planner/session module where
create_planner_session and session.is_expired are defined.
🪄 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: defaults
Review profile: CHILL
Plan: Pro
Run ID: 68c5fb54-8cf2-4e77-afce-254a0cc3424b
📒 Files selected for processing (1)
dev-suite/tests/test_planner.py
Comment on lines
+391
to
+396
| def test_session_touch(self): | ||
| session = create_planner_session(workspace="/proj") | ||
| old_time = session.last_activity | ||
| time.sleep(0.01) | ||
| session.touch() | ||
| assert session.last_activity > old_time |
There was a problem hiding this comment.
Avoid time.sleep() in test_session_touch to reduce CI flakiness.
Sleeping in unit tests is slow and timing-sensitive. You can set last_activity to a known older value and assert touch() updates it.
Proposed diff
def test_session_touch(self):
session = create_planner_session(workspace="/proj")
- old_time = session.last_activity
- time.sleep(0.01)
+ old_time = session.last_activity
+ session.last_activity = old_time - 10
session.touch()
- assert session.last_activity > old_time
+ assert session.last_activity > old_time📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
Suggested change
| def test_session_touch(self): | |
| session = create_planner_session(workspace="/proj") | |
| old_time = session.last_activity | |
| time.sleep(0.01) | |
| session.touch() | |
| assert session.last_activity > old_time | |
| def test_session_touch(self): | |
| session = create_planner_session(workspace="/proj") | |
| old_time = session.last_activity | |
| session.last_activity = old_time - 10 | |
| session.touch() | |
| assert session.last_activity > old_time |
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@dev-suite/tests/test_planner.py` around lines 391 - 396, The
test_session_touch uses time.sleep which causes flakiness; instead, set
session.last_activity to a deterministic older timestamp and call
session.touch(), then assert session.last_activity is greater than the previous
value. Modify test_session_touch to create the session via
create_planner_session(), assign session.last_activity to a past datetime (or
timestamp) value, call session.touch(), and assert the updated
session.last_activity > previous value (referencing the test function name
test_session_touch and methods create_planner_session, session.touch, and
session.last_activity).
Closed
29 tasks
Abernaughty
added a commit
that referenced
this pull request
Apr 5, 2026
Wire the dashboard ChatView to the Planner backend (PR #121). Conversational task planning flow: workspace select → describe objective → Planner validates checklist → submit to Architect. New files: - stores/planner.svelte.ts — session state, message history, checklist - routes/api/planner/+server.ts — proxy POST /tasks/plan - routes/api/planner/[id]/message/+server.ts — proxy message endpoint - routes/api/planner/[id]/submit/+server.ts — proxy submit endpoint Modified files: - types/api.ts — Planner types + planner_message SSE event - sse.ts — dispatch planner_message events to plannerStore - stores/index.ts — export + teardown plannerStore - components/views/ChatView.svelte — full rewrite with Planner flow Closes #106
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Implements Issue #106 Phase B backend: the Planner agent, TaskSpec schema, workspace auto-inference, readiness checklist validation, and three new API endpoints for conversational task planning.
Closes the backend portion of #106 Phase B. Dashboard (ChatView) updates will follow in a subsequent PR.
What's New
Planner Agent (
src/agents/planner.py)infer_workspace_stack()— auto-detects languages and frameworks from project files (package.json, pyproject.toml, Cargo.toml, go.mod, etc.)send_planner_message()— core Planner interaction: sends user message to LLM, extracts TaskSpec updates from response JSON block, updates checklistPlannerSessionStore— in-memory session store with lazy TTL cleanupPLANNER_MODELenv var — configurable model (default:gemini-2.0-flash-lite), supports Gemini, Anthropic, and OpenAI-compatible modelsAPI Endpoints (
src/api/main.py)POST /tasks/plan— Start a new Planner session with workspace validation + auto-inferencePOST /tasks/plan/{session_id}/message— Send a message to the Planner agentPOST /tasks/plan/{session_id}/submit— Confirm and create task from TaskSpec (validates required checklist items)SSE Events (
src/api/events.py)PLANNER_MESSAGEevent type — streams Planner responses to the dashboard in real-timeAPI Models (
src/api/models.py)PlannerStartRequest,PlannerMessageRequest— request bodiesPlannerSessionResponse,PlannerSubmitResponse— response schemasPlannerChecklistItemResponse,PlannerChecklistResponse— readiness checklist displayPlannerTaskSpecResponse— structured task spec for dashboard renderingTests (
tests/test_planner.py)Security Considerations
_apply_spec_updates()explicitly skipsworkspacefield from LLM outputDesign Decisions
PLANNER_MODELenv var, defaultgemini-2.0-flash-liteWhat's Next (Session 2)
planner_messageevents to dashboard storeTesting
cd dev-suite uv run --group dev pytest tests/test_planner.py -vSummary by CodeRabbit
New Features
Tests