Sub-agent sessions created via the task tool or call_omo_agent tool were
being created without a model. The resolved model was only passed as a
promptAsync body override, which opencode core ignores, causing fallback
to the client's system default model (often a reasoning model that hangs).
OpenCode's session.create API supports model: { id, providerID, variant }
at creation time. This fix ensures the resolved category/agent model is
passed during session creation across all paths:
- createSyncSession() (delegate task / category-based)
- createOrGetSession() (call_omo_agent / direct agent calls)
- BackgroundManager.startTask() (background tasks)
This guarantees each sub-agent session is created with the correct model
regardless of whether promptAsync honors its model override.
Atlas was delegating tasks one-by-one because the workflow framed parallel
as a conditional ("if tasks can run in parallel..."), letting models default
to the safer sequential path. The new shared ATLAS_PARALLEL_BY_DEFAULT block
flips the default: parallel is mandatory; sequential requires a NAMED
blocking dependency (input dependency or file conflict).
Adds two new prompt variants — kimi (K2.6 thinking-mode calibration:
commitment framing + concrete budgets) and opus-4-7 (counters 4.7's lower
default subagent count and literal-following bias). Recalibrates default
(Claude 4.6 family), gpt (GPT-5.5 outcome-first / decision rules over
absolutes), and gemini (preserves TOOL_CALL_MANDATE; replaces stale
session_id with task_id). All five variants share the parallel mandate
positioned BEFORE the workflow so "mandate above" references resolve.
Strengthens the orchestrator-direct-edit reminder hooks
(ORCHESTRATOR_DELEGATION_REQUIRED + DIRECT_WORK_REMINDER) with the central
challenge "Do you ACTUALLY need to be the one doing this?" — replacing the
previous bullet-heavy framing.
Tests now parametrized over all 5 variants. Adds prompt-routing.test
covering GPT/Gemini/Kimi/Opus 4.7/default routing and edge cases, plus a
session_id rejection test (every variant must use task_id for retries).
Replaces fsyncSync(tempFileDescriptor) with tolerantFsyncSync, allowing
EPERM/EACCES/ENOTSUP/EINVAL during fsync while still propagating real
errors. Adds an optional deps.fsyncSync injection point used solely by
the new EPERM tolerance regression tests.
Without this fix, plugin startup itself can fail on synced folders
because writeFileAtomically is used by config migrations and posthog
activity state — the same EPERM-on-fsync failure pattern reported for
team_create.
Replaces direct fileHandle.sync() calls in acquireLock and atomicWrite
with tolerantFsync. Users on iCloud Drive / OneDrive / Desktop sync
folders were hitting 'EPERM: operation not permitted, fsync' during
team_create, which propagated up and aborted the entire team_create
flow even though the actual write+rename had succeeded.
Reported on Discord (omo 4.0.0, opencode desktop 1.14.41, project on
synced Desktop). atomicity is preserved by the temp-file rename; only
the durability hint is now best-effort on filesystems that disallow
fsync.
Adds isToleratedFsyncError, tolerantFsync (async, FileHandle), and
tolerantFsyncSync (sync, fd) helpers that swallow filesystem-limitation
errors during fsync (EPERM, EACCES, ENOTSUP, EINVAL) while still
propagating real errors (EIO, ENOSPC, EBADF, etc.). Synced folders
like iCloud Drive, OneDrive, and antivirus-locked files reject fsync
with EPERM even though the underlying write+rename succeeded; for the
runtime data this codebase persists, losing the durability hint is
acceptable in exchange for not blocking the operation entirely.
The helper is intentionally not barrel-exported (consumers import the
file directly), matching the existing convention for write-file-atomically.
Commit 2dfa6336f changed Metis primary model from claude-opus-4-7 max
to claude-sonnet-4-6 in src/shared/model-requirements.ts but only
updated the unit test in model-requirements.test.ts. The CLI installer
end-to-end snapshot test (src/cli/model-fallback.test.ts) generates
configs from those same fallback chains via generateModelConfig and
captures them as snapshots, so 17 snapshot assertions broke.
Snapshots regenerated with `bun test src/cli/model-fallback.test.ts -u`.
Verified 86/86 pass across model-requirements + model-fallback +
openai-only-model-catalog test files.
The change is a no-op for the web import that this PR is primarily
about — including it here to unblock CI on dev (where the breakage
landed) rather than spinning up a separate hotfix PR.
The previous commit (2dfa6336f) used 'git add -A' which swept up files that
prior commits had intentionally deleted but were still present untracked in
the local workspace. This commit re-deletes them to match the upstream intent.
Files removed (each was previously deleted in the cited commit):
- drafts/gpt-5-5/{README,deep,hephaestus,oracle,sisyphus-junior,sisyphus}.md
(deleted in 40ded422c chore(drafts): remove stale gpt-5.5 prompt drafts)
- src/plugin-dispose.{ts,test.ts}
(deleted in e2f5c0d36 refactor(plugin): remove orphaned createPluginDispose)
- src/features/tmux-subagent/{cleanup,session-created-handler,session-deleted-handler}.ts
(deleted in 7a7926f22 chore(tmux-subagent): remove dead event-handler modules)
- src/tools/delegate-task/{model-string-parser,resolve-call-id,resolve-call-id.test}.ts
(model-string-parser deleted in db056346d; resolve-call-id was scratch)
- src/__debug-test.test.ts (debug scratch never intended for git)
Typecheck + model-requirements tests still pass. The legitimate metis +
AGENTS.md edits from the previous commit remain on dev.
Source code change:
- src/shared/model-requirements.ts: prepend claude-sonnet-4-6 to metis fallback
chain so Sonnet becomes the default. Opus 4.7 max remains as the immediate
fallback for callers who want extra reasoning.
- src/shared/model-requirements.test.ts: update assertion to expect Sonnet
primary + Opus secondary.
AGENTS.md accuracy fixes (verified against source):
- Agent modes: Sisyphus/Hephaestus are 'primary' (not 'all'); Sisyphus-Junior
is 'subagent' (not 'all'). Confirmed via 'const MODE: AgentMode = ...' in
each agent file. Also clarified Prometheus has no agentSources factory and
is built via buildPrometheusAgentConfig.
- Sisyphus fallback chain: corrected order to kimi-k2.6 → k2p5 → kimi-k2.5
→ gpt-5.5 medium → glm-5 → big-pickle (was missing kimi-k2.5).
- Librarian/Explore: added missing minimax-m2.7 step between -highspeed and
claude-haiku-4-5.
- Metis chain: removed fictitious gemini-3.1-pro entry.
- Sisyphus-Junior chain: spelled out the actual fallback (was 'user-configurable').
- Temperatures: Sisyphus/Hephaestus do not set explicit temperature (model
default); Sisyphus-Junior is 0.1 via SISYPHUS_JUNIOR_DEFAULTS.
- Quick category default: gpt-5.4-mini (not gpt-5.4-mini-fast).
Team-mode corrections:
- Eligibility registry has 3 verdicts: eligible (sisyphus, atlas, sisyphus-junior),
conditional (hephaestus — needs D-36 teammate permission), hard-reject
(oracle, librarian, explore, multimodal-looker, metis, momus, prometheus).
- Schema has 11 fields, not 4: added max_messages_per_run, max_wall_clock_minutes,
max_member_turns, base_dir, message_payload_max_bytes, recipient_unread_max_bytes,
mailbox_poll_interval_ms.
- Hooks: 'team-session-events' is 4 sub-handlers in src/plugin/event.ts
(team-idle-wake-hint, team-lead-orphan-handler, team-member-error-handler,
team-member-status-handler), not a single Continuation-tier hook.
- Tier counts now show base + team-mode: ToolGuard 14/15, Transform 5/7.
- Total: 52 base hooks, 59 with team-mode.
Doc cascade for the Metis change:
- docs/guide/orchestration.md, agent-model-matching.md, installation.md
- docs/reference/configuration.md, features.md
OpenCode Desktop 1.14.x filters agents by `mode` field. The `mode` was
only present on the factory function as a static property, but not copied
to the generated AgentConfig object.
This fix copies `source.mode` to `base.mode` in `buildAgent()` when:
- source is a factory function (has `mode` static property)
- the generated config doesn't already define `mode`
Test: agent-builder.test.ts (4 pass)
Closes: #3835
Related: #3762, #3812, #3794, #3475, #3474, #3829, #3831, #3824, #3826, #3721, #3806, #3188
The caller-facing description told orchestrators WHAT `deep` is but
never specified goal cardinality, so they bundled multiple goals into
one call (e.g. "fix X + merge Y + deploy Z"). The deep agent's own
prompt was already configured to refuse such bundles, but only after
they arrived. The rule now lives on the caller side: ONE goal + ONE
deliverable per call; multiple goals must fan out as parallel `deep`
calls.
Uses new Function() to prevent the static ESM loader from seeing
the bun: protocol import at parse time. In Electron/Node.js, the
import() call is evaluated at runtime and gracefully returns null
when bun:sqlite is unavailable.
Fixes#3829, likely fixes#3762
- Remove plugin_loaded telemetry from index.ts (was 46.5% of all events, ~2.83M/month)
- Add enableLocalEvaluation: false to prevent feature flag polling/decide calls
- Add strictLocalEvaluation: true to prevent server fallback requests
- Add disableRemoteConfig: true to prevent remote config network requests
- Remove 'plugin_loaded' from PostHogActivityReason type
- Update tests: remove stale mocks, add SDK options verification test
- enableExceptionAutocapture: false already present (kept)
Estimated billing reduction: ~$960+/month from feature flag requests,
plus ~2.83M fewer events/month from plugin_loaded removal.
- DEEP_CATEGORY_PROMPT_APPEND_GPT_5_3_CODEX preserving all DEEP knowledge (atomic task treatment, root cause bias, ambition scaled, completion bar, status sparse) plus codex-specific additions: bias to action, parallel-batch exploration, code implementation discipline, worktree safety, plan closure, pragmatic final-message format\n- resolveDeepCategoryPromptAppend now routes codex first, then 5.5, else legacy (preserves existing 5.5 and 5.4 test invariants)\n- Tests: assert codex no longer falls back to legacy, content tests mirroring the 5.5 pattern (style markers, knowledge preservation, parallel-batch framing, materially different from siblings)
Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
- Add isGpt5_2Model type guard\n- ORACLE_GPT_5_2_PROMPT consolidating all knowledge from Claude default, GPT-5.4 generic, and GPT-5.5 variants (XML-tagged blocks, concrete verbosity clamps, long-context re-grounding, anti-narration tool rules, high-risk self-check)\n- MOMUS_GPT_5_2_PROMPT preserving blocker-finder philosophy with new tool_usage_rules block for parallel reference verification\n- Momus GPT-5.2 reasoningEffort set to xhigh per evaluation rigor needs
Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
Agent names like \hephaestus\ cause 'Agent not found' errors because
sanitizeSubagentType() was only called in subagent-resolver.ts but not
in the direct manager.launch() path or background-executor.ts.
- manager.ts: strip leading/trailing backslash/quote chars from input.agent
before validation so \hephaestus\ → hephaestus
- background-executor.ts: call sanitizeSubagentType(args.subagent_type)
instead of passing raw value to manager.launch()
- agent-display-names.ts: reuse sanitizeSubagentType in stripAgentListSortPrefix
- Add unit tests for all three fix points
Fixes: sessions dying with 'Agent not found: \hephaestus\'
x-x-gpu