Problem: Agents frequently omit both 'category' and 'subagent_type' parameters
when calling the task() tool, causing validation failures. The JSON Schema
marks both as optional, and LLMs follow schema structure over description text.
Solution (Option A): Add aggressive visual warnings and failure-mode examples
to the tool description:
- ⚠️ CRITICAL warning header
- COMMON MISTAKE example showing what will FAIL
- CORRECT examples for both category and subagent_type usage
- Clear explanation that ONE must be provided
Tests: All 153 existing tests pass (no behavior change, only prompt improvement)
When browserProvider is not set, agent-browser skill should NOT resolve.
Test assertions were inverted — expected 'Skills not found' but asserted the opposite.
- Add normalizeFallbackModels helper to centralize string/array normalization (P3)
- Export RuntimeFallbackConfig and FallbackModels types from config/index.ts
- Fix agent detection regex to use word boundaries for sessionID matching
- Improve tests to verify actual fallback switching logic (not just log paths)
- Add SessionCategoryRegistry cleanup in executeSyncTask on completion/error (P2)
- All 24 runtime-fallback tests pass, 115 delegate-task tests pass
Implement full fallback_models support across all integration points:
1. Model Resolution Pipeline (src/shared/model-resolution-pipeline.ts)
- Add userFallbackModels to ModelResolutionRequest
- Process user fallback_models before hardcoded fallback chain
- Support both connected provider and availability checking modes
2. Agent Utils (src/agents/utils.ts)
- Update applyModelResolution to accept userFallbackModels
- Inject fallback_models for all builtin agents (sisyphus, oracle, etc.)
- Support both single string and array formats
3. Model Resolver (src/shared/model-resolver.ts)
- Add userFallbackModels to ExtendedModelResolutionInput type
- Pass through to resolveModelPipeline
4. Delegate Task Executor (src/tools/delegate-task/executor.ts)
- Extract category fallback_models configuration
- Pass to model resolution pipeline
- Register session category for runtime-fallback hook
5. Session Category Registry (src/shared/session-category-registry.ts)
- New module: maps sessionID -> category
- Used by runtime-fallback to lookup category fallback_models
- Auto-cleanup support
6. Runtime Fallback Hook (src/hooks/runtime-fallback/index.ts)
- Check SessionCategoryRegistry first for category fallback_models
- Fallback to agent-level configuration
- Import and use SessionCategoryRegistry
Test Results:
- runtime-fallback: 24/24 tests passing
- model-resolver: 46/46 tests passing
Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-opencode)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
transformModelForProvider only handled github-copilot provider, leaving
google provider models untransformed. This caused ProviderModelNotFoundError
when google/gemini-3-flash was sent to the API (correct ID is
gemini-3-flash-preview).
Changes:
- Add google provider to transformModelForProvider with idempotent regex
negative lookahead to prevent double -preview suffix
- Fix category-default path in model-resolution-pipeline when
availableModels is empty but connected provider exists
- Fix getFirstFallbackModel first-run path that constructed raw model IDs
without transformation
- Fix github-copilot provider gemini transforms to also use idempotent
regex (was vulnerable to double-transform)
- Extract transformModelForProvider to shared module (single source of
truth, imported by cli and shared layers)
- Add 20 new test cases: unit tests for both providers, runtime
integration tests for category-default and fallback-chain paths,
double-transform prevention for both providers
Category-based delegation (task(category='quick')) was broken because
SISYPHUS_JUNIOR_AGENT sent 'sisyphus-junior' to session.prompt but
config.agent keys are now display names ('Sisyphus-Junior').
- Use getAgentDisplayName() for SISYPHUS_JUNIOR_AGENT constant
- Replace hardcoded 'sisyphus-junior' strings in tools.ts with constant
- Update background-output local constants to use display names
- Add remapCommandAgentFields() to translate command agent fields
- Add raw-key fallback in tool-config-handler agentByKey()
The messages() mock in 'session_id with background=false' test did not
filter by session ID, causing resolveParentContext's SDK calls for
parent-session to increment messagesCallCount. This inflated
anchorMessageCount to 4 (matching total messages), so the poll loop
could never detect new messages and always hit MAX_POLL_TIME_MS.
Fix: filter messages() mock by path.id so only target session
(ses_continue_test) increments the counter. Restore MAX_POLL_TIME_MS
from 8000 back to 2000.
- runner.test.ts: waitForEventProcessorShutdown timeout 50ms → 500ms
(50ms was consistently too tight for CI runners)
- tools.test.ts: MAX_POLL_TIME_MS 2000ms → 8000ms
(polling timed out at ~2009ms on CI due to resource contention)
The polling loop in executeUnstableAgentTask only checked session status
and message stability, never checking if the background task itself had
been interrupted. This caused the tool call to hang until MAX_POLL_TIME_MS
(10 minutes) when a task was interrupted by prompt errors.
Add manager.getTask() check at each poll iteration to break immediately
on terminal statuses (interrupt, error, cancelled), returning a clear
failure message instead of hanging.
Pass parentTools from session-tools-store through the background task
lifecycle (launch → task → notify) so that when notifyParentSession
sends promptAsync, the original tool restrictions (e.g., question: false)
are preserved. This prevents the Question tool from re-enabling after
call_omo_agent background tasks complete.
Call setSessionTools(sessionID, tools) before every prompt dispatch so
the tools object is captured and available for later retrieval when
background tasks complete.
LLMs sometimes pass load_skills as a serialized JSON string instead
of an array. Add defensive JSON.parse before validation to handle
this gracefully.
Fixes#1701
Community-reported-by: @omarmciver
When a custom agent's last assistant message contains only tool calls (no text/reasoning parts), the sync result fetcher returned empty content. Walk assistant messages newest-first to find the first one with actual text content.
Replace PreToolUse hook-based question tool blocking with the existing
tools parameter approach (tools: { question: false }) which physically
removes the tool from the LLM's toolset before inference.
The hook was redundant because every session.prompt() call already passes
question: false via the tools parameter. OpenCode converts this to a
PermissionNext deny rule and deletes the tool from the toolset, preventing
the LLM from even seeing it. The hook only fired after the LLM already
called the tool, wasting tokens.
Changes:
- Remove subagent-question-blocker hook invocation from PreToolUse chain
- Remove hook registration from create-session-hooks.ts
- Delete src/hooks/subagent-question-blocker/ directory (dead code)
- Remove hook from HookNameSchema and barrel export
- Fix sync-executor.ts missing question: false in tools parameter
- Add regression tests for both the removal and the tools parameter
When delegate-task resumes a session via session_id, the response
task_metadata now includes a subagent field identifying which agent
was running in the resumed session. This allows the parent agent to
know what type of subagent it is continuing.
- sync-continuation: uses resumeAgent extracted from session messages
- background-continuation: uses task.agent from BackgroundTask object
- Gracefully omits subagent when agent info is unavailable
Allow individual categories to be disabled via `disable: true` in
config. Introduce shared `mergeCategories()` utility to centralize
category merging and disabled filtering across all 7 consumption sites.
Child session creation was injecting permission: { question: 'deny' } which
conflicted with OpenCode's child session permission handling, causing subagent
sessions to hang with 0 messages after creation (zombie state).
Remove the permission override from all session creators (BackgroundManager,
sync-session-creator, call-omo-agent) and rely on prompt-level tool restrictions
(tools.question=false) to maintain the intended policy.
Closes#1711
YeonGyu-Kim