model_not_supported errors from providers (e.g. OpenAI returning
{"error": {"code": "model_not_supported"}}) were not recognized as
retryable. Subagents would silently fail with no response, hanging the
parent session.
Fix:
- Add "model_not_supported", "model not supported", "model is not
supported" to RETRYABLE_MESSAGE_PATTERNS in model-error-classifier.ts
- Add regex patterns to RETRYABLE_ERROR_PATTERNS in
runtime-fallback/constants.ts to match "model ... is ... not ...
supported" with flexible spacing
- Add regression test covering all three variations
Now model_not_supported errors trigger the normal fallback chain instead
of silent failure.
Fire-and-forget session.abort() calls during subagent completion left
dangling promises that raced with parent session teardown. In Bun on
WSL2/Linux, this triggered a StringImplShape assertion (SIGABRT) as
WebKit GC collected string data still referenced by the inflight request.
Fix: await session.abort() in all four completion/error paths:
- startTask promptAsync error handler (launch path)
- resume promptAsync error handler (resume path)
- cancelTask (explicit cancel path)
- tryCompleteTask (normal completion path)
Also marks the two .catch() error callbacks as async so the await is
valid.
Test: update session.deleted cascade test to flush two microtask rounds
since cancelTask now awaits abort before cleanupPendingByParent.
The installer wrapper and postinstall script still hardcoded the old
oh-my-opencode-* platform package family. When users installed the renamed
npm package oh-my-openagent (for example via npx oh-my-openagent install on
WSL2/Linux), the main package installed correctly but the wrapper looked for
nonexistent optional binaries like oh-my-opencode-linux-x64.
Fix:
- derive packageBaseName from package.json at runtime in wrapper/postinstall
- thread packageBaseName through platform package candidate resolution
- keep legacy oh-my-opencode default as fallback
- add regression test for renamed package family
skill-context already filtered browser-related skills using the configured
browser provider, but the skill tool rebuilt discovery without forwarding
browserProvider. That caused skills like agent-browser to be prompt-visible
while skill() could still fail to resolve them unless browser_automation_engine.provider
was explicitly threaded through both paths.
Fix:
- pass skillContext.browserProvider from tool-registry into createSkillTool
- extend SkillLoadOptions with browserProvider
- forward browserProvider to getAllSkills()
- add regression tests for execution and description visibility
The skill tool previously only merged disk-discovered skills and preloaded
options.skills, so skills registered natively via ctx.skills (for example
from config.skills.paths or other plugins) were prompt-visible but not
discoverable by skill().
Fix:
- pass ctx.skills from tool-registry into createSkillTool
- extend SkillLoadOptions with nativeSkills accessor
- merge nativeSkills.all() results into getSkills()
- add test covering native PluginInput skill discovery
task_system now defaults to true instead of false. Users on opencode-go
or other plans were unable to invoke oracle or use subagent delegation
because experimental.task_system defaulted off.
The task system is the primary mechanism for oracle, explore, and other
subagent invocations — it should not be behind an experimental flag.
Users can still explicitly set experimental.task_system: false to opt out.
resolveSkillPathReferences: add looksLikeFilePath() guard that requires
a file extension or trailing slash before resolving @scope/package
references. npm packages like @mycom/my_mcp_tools@beta were incorrectly
being rewritten to absolute paths in skill templates.
2 new tests.
TTL (pruneStaleTasksAndNotifications) now resets on last activity:
- Uses task.progress.lastUpdate as TTL anchor for running tasks
(was always using startedAt, causing 30-min hard deadline)
- Added taskTtlMs config option for user-adjustable TTL
- Error message shows actual TTL duration, not hardcoded '30 minutes'
- 3 new tests for the new behavior
subagent-resolver: when falling back to matchedAgent.model for custom
subagents, verify the model is actually available via fuzzyMatchModel
before setting it as categoryModel. Prevents delegate_task from using
an unavailable model when the user's custom agent config references
a model they don't have access to.
Test updated to include the target model in available models mock.
- completedTaskSummaries now includes status and error info
- notifyParentSession: noReply=false for failed tasks so parent reacts
- Batch notification distinguishes successful vs failed/cancelled tasks
- notification-template updated to show task errors
- task-poller: session-gone tests (85 new lines)
- CI: add Bun shim to PATH for legacy plugin migration tests
When a subagent session disappears from the status registry (process
crashed), the main agent was waiting the full stale timeout before
acting. Fix:
- Add sessionGoneTimeoutMs config option (default 60s, vs 30min normal)
- task-poller: use shorter timeout when session is gone from status
- manager: verify session existence when gone, fail crashed tasks
immediately with descriptive error
- Add legacy-plugin-toast hook for #2823 migration warnings
- Update schema with new config option
- logLegacyPluginStartupWarning now emits console.warn (visible to user,
not just log file) when oh-my-opencode is detected in opencode.json
- Auto-migrates opencode.json plugin entry from oh-my-opencode to
oh-my-openagent (with backup)
- plugin-config.ts: add console.warn when loading legacy config filename
- test: 10 tests covering migration, console output, edge cases
Two issues fixed:
1. process-cleanup.ts used fire-and-forget void Promise for shutdown
handlers — now properly collects and awaits all cleanup promises
via Promise.allSettled, with dedup guard to prevent double cleanup
2. TmuxSessionManager was never registered for process cleanup —
now registered in create-managers.ts via registerManagerForCleanup
Also fixed setTimeout().unref() which could let the process exit
before cleanup completes.
Background subagents (explore/librarian) failed with auth errors because
resolveModelForDelegateTask() always picked the first fallback entry when
availableModels was empty — often an unauthenticated provider like xai
or opencode-go.
Fix: when connectedProvidersCache is populated, iterate fallback chain
and pick the first entry whose provider is in the connected set.
Legacy behavior preserved when cache is null (not yet populated).
- model-selection.ts: use readConnectedProvidersCache to filter fallback chain
- test: 4 new tests for connected-provider-aware resolution
lastCompactionTime was only set on successful compaction. When compaction
failed (rate limit, timeout, etc.), no cooldown was recorded, causing
immediate retries in an infinite loop.
Fix: set lastCompactionTime before the try block so both success and
failure respect the cooldown window.
- test: add failed-compaction cooldown enforcement test
- test: fix timeout retry test to advance past cooldown
The server-health module used module-level state for inProcessServerRunning,
which doesn't survive when Bun loads separate module instances in the same
process. Fix: use globalThis with Symbol.for key so the flag is truly
process-global.
- server-health.ts: replace module-level boolean with globalThis[Symbol.for()]
- export markServerRunningInProcess from tmux-utils barrel
- test: verify flag skips HTTP fetch, verify globalThis persistence
When .sisyphus/ is gitignored, task state written during worktree execution
is lost when the worktree is removed. Fix:
- add worktree-sync.ts: syncSisyphusStateFromWorktree() copies .sisyphus/
contents from worktree to main repo directory
- update start-work.ts template: documents the sync step as CRITICAL when
worktree_path is set in boulder.json
- update work-with-pr/SKILL.md: adds explicit sync step before worktree removal
- export from boulder-state index
- test: 5 scenarios covering no-.sisyphus, nested dirs, overwrite stale state
- fix(switcher): use lastIndexOf for multi-slash model IDs (e.g. aws/anthropic/claude-sonnet-4)
- fix(model-resolution): same lastIndexOf fix in doctor parseProviderModel
- fix(call-omo-agent): resolve model from agent config and forward to both
background and sync executors via DelegatedModelConfig
- fix(subagent-resolver): inherit category model/variant when agent uses
category reference without explicit model override
- test: add model override forwarding tests for call-omo-agent
- test: add multi-slash model ID test for switcher
- fix(context-limit): check modelContextLimitsCache for all Anthropic
models, not just GA-model set; user config/cache wins over 200K default
(fixes#2836)
- fix(agent-key-remapper): preserve config key aliases alongside display
names so `opencode run --agent sisyphus` resolves correctly
(fixes#2858)
- fix(tool-config): respect host permission.skill=deny by disabling
skill/skill_mcp tools when host denies them (fixes#2873)
- test: update context-limit and agent-key-remapper tests to match new
behavior
Models frequently hallucinate a 'directory' parameter alongside filePath,
causing hard failures. Instead of rejecting, accept directory as an alias
for filePath and gracefully handle when both are provided (prefer filePath).
This prevents the 'filePath and directory are mutually exclusive' error
that users hit when models pass both parameters.
Fixes model confusion with lsp_diagnostics tool parameters.
