- Update DEFAULT_CATEGORIES to use 'openai/gpt-5.4-high' directly instead of separate model + variant
- Add helper functions (isExplicitHighModel, getExplicitHighBaseModel) to preserve explicit high models during fuzzy matching
- Update category resolver to avoid collapsing explicit high models to base model + variant pair
- Update tests to verify explicit high model handling in both background and sync modes
- Update documentation examples to reflect new configuration
🤖 Generated with OhMyOpenCode assistance
Align runtime defaults, tests, docs, and generated artifacts with the newer GPT-5.4 baseline. Keep think-mode and prompt-routing expectations consistent after the model version bump.
Restructure from 13 scattered XML blocks to 8 dense blocks with 9
named sub-anchors, following OpenAI GPT-5.4 prompting guidance and
Oracle-reviewed context preservation strategy.
Key changes:
- Merge think_first + intent_gate + autonomy into unified <intent>
with domain_guess classification and <ask_gate> sub-anchor
- Add <execution_loop> as central workflow: EXPLORE -> PLAN -> ROUTE ->
EXECUTE_OR_SUPERVISE -> VERIFY -> RETRY -> DONE
- Add mandatory manual QA in <verification_loop> (conditional on
runnable behavior)
- Move <constraints> to position #2 for GPT-5.4 attention pattern
- Add <completeness_contract> as explicit loop exit gate
- Add <output_contract> and <verbosity_controls> per GPT-5.4 guidance
- Add domain_guess (provisional) in intent, finalized in ROUTE after
exploration -- visual domain always routes to visual-engineering
- Preserve all named sub-anchors: ask_gate, tool_persistence,
parallel_tools, tool_method, dependency_checks, verification_loop,
failure_recovery, completeness_contract
- Add skill loading emphasis at intent/route/delegation layers
- Rename EXECUTE to EXECUTE_OR_SUPERVISE to preserve orchestrator
identity with non-execution exits (answer/ask/challenge)
Add MANUAL_QA_MANDATE sections to all three ultrawork prompts (default,
GPT, Gemini). Agents must now define acceptance criteria in TODO/Task items
before implementation, then execute manual QA themselves after completing
work. lsp_diagnostics alone is explicitly called out as insufficient since
it only catches type errors, not functional bugs.
Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-opencode)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
Remove deep parallel delegation section from GPT-5.4 Sisyphus prompt since
it encouraged direct implementation over orchestration. Add zero-tolerance
category domain matching guide to all Sisyphus prompts with visual-engineering
examples. Rewrite visual-engineering category prompt with 4-phase mandatory
workflow (analyze design system, create if missing, build with system, verify)
targeting Gemini's tendency to skip foundational steps.
Add <parallel_tool_calling> and <tool_usage_rules> blocks that GPT-5.4
treats as first-class behavioral contracts. Add parallel-planning question
to <think_first>, strengthen Exploratory route in intent gate, and add
IN PARALLEL annotations to verification loop.
Remove the hardcoded external_directory: "allow" default from
applyToolConfig(). This was silently overriding OpenCode's built-in
default of "ask" and any user-configured external_directory permission.
With this change, external_directory permission is fully controlled by
OpenCode's defaults and user configuration, as intended.
Fixes#1973Fixes#2194
Replace real setTimeout(7000) with fake timer interception in atlas
retry tests (35s -> 227ms). Add missing opencode provider to glm-5
fallback in unspecified-high category.
When atlas injects a boulder continuation via promptAsync() and the
model's response is immediately aborted (MessageAbortedError), OpenCode
fires a burst of session.idle events within milliseconds. Atlas blocks
all of them due to the 5-second cooldown. After the burst, OpenCode
stops generating session.idle events (it's state-change based, not
periodic), leaving the session stuck forever.
Fix: When cooldown blocks an idle event for a boulder session with an
incomplete plan, schedule a one-shot setTimeout (cooldown + 1s) to
re-attempt injection. The timer callback re-checks boulder state, plan
progress, and continuation-stopped flag before injecting. Only one timer
per session is allowed (deduped via pendingRetryTimer field). Timers are
cleaned up on session.deleted and session.compacted events.
Split monolithic hephaestus.ts into directory with model-specific prompt
variants (gpt-5-4.ts, gpt-5-3-codex.ts, gpt.ts) mirroring the
sisyphus-junior pattern. Generic gpt.ts uses pre-codex-tuning prompt as
fallback for non-specific GPT models.
Also adds isGpt5_4Model and isGpt5_3CodexModel helpers to types.ts.
YeonGyu-Kim