- New src/agents/sisyphus/kimi-k2-6.ts based on gpt-5-4.ts 8-block architecture
- New src/agents/sisyphus-junior/kimi-k2-6.ts worker variant
- Preserves 4 pillars: intent gate + verbalization, parallel tools, verification
- Adds <re_entry_rule>: suppress re-verbalization for already-resolved turns
- Adds <exploration_budget>: hard stop conditions alongside aggressive parallelism
- Tiered <verification_loop> (V1/V2/V3): V3 keeps full rigor with harsh enforcement
- Adds <token_economy>: verbalization explicitly excluded from trim mandate
- isKimiK2Model in types.ts: matches kimi, k2p5/k2p6 variants (case-insensitive)
- Routing in sisyphus.ts + sisyphus-junior/agent.ts
- Tests: 3 new kimi routing cases in sisyphus-junior/index.test.ts (all pass)
Motivation: K2.x was post-trained with Toggle RL (~25-30% token reduction) and a
GRM scoring appropriate detail + intent inference. Reusing Claude-style prompts
double-taxes the model — external strictness on top of RL-learned strictness causes
over-deliberation on already-resolved requests. The re-entry rule and exploration
budget fix this without weakening verification rigor.
Refs: kimi.com/blog/kimi-k2-6, arxiv 2602.02276 §4.4.2 (Toggle, GRM)
Ground-up rewrite styled after OpenAI Codex's gpt-5.4 prompt
architecture: '# General' -> '## Autonomy and Persistence' -> '## Task
execution' -> '## Validating your work' -> '# Working with the user' ->
'# Tool Guidelines' section hierarchy.
Key differences from the gpt-5-4 variant:
- Prose-first output, bullets only when content is list-shaped
- Contract frames replace threat frames (GPT-5.5 follows instructions
well; NEVER/FORBIDDEN rhetoric adds entropy without compliance gain)
- Explicit opener blacklist for 'Done -', 'Got it', 'Great question'
- '{{ personality }}' slot reserved for future persona substitution
- '{{ taskSystemGuide }}' slot switches todo/task tools per harness cfg
- Codex-compatible clickable file reference format
Sisyphus factory now checks isGpt5_5Model before isGptNativeSisyphusModel,
so gpt-5.5 models route to the new prompt while gpt-5.4, gpt-5.6+, and
other matches stay on the existing gpt-5-4 prompt.
Extract hardcoded GPT apply_patch permission logic into a reusable module
to ensure consistent behavior across all agents. This prevents GPT models
from using the unreliable apply_patch tool while allowing other models.
- Add gpt-apply-patch-guard.ts with GPT_APPLY_PATCH_GUIDANCE and getGptApplyPatchPermission
- Update Hephaestus agent to use centralized permission logic
- Update Sisyphus-Junior agent to use centralized permission logic
- Update all GPT prompt builders to reference shared guidance constant
🤖 Generated with assistance of OhMyOpenCode
Built-in agents (Sisyphus, Hephaestus, Atlas) now include explicit
identity section in their system prompts, matching custom agent behavior.
12 agent identity tests pass, 266 total pass, tsc clean.
Closes#2964
GPT models (5.3-codex, 5.4, etc.) frequently hang when using apply_patch
due to verification loops. This adds:
1. Tool restriction: apply_patch is denied for GPT variants of
Hephaestus, Sisyphus-Junior, and Sisyphus agents
2. Prompt guidance: GPT-specific prompts now explicitly instruct using
edit/write tools instead of apply_patch
3. Removed the 'Always use apply_patch' instruction from
sisyphus-junior/gpt-5-4.ts that contradicted the fix
The deny is model-conditional — Claude variants retain apply_patch
access since it works reliably there.
Tool return values contained CTAs ("Use background_output to check") that
caused models to immediately poll instead of waiting for <system-reminder>
notifications. Unified all 5 tool return messages with explicit "Do NOT call
background_output now" guard, added pre-notification warning to BackgroundOutput
tool description, and strengthened Background Result Collection sections across
all 3 Sisyphus prompt variants (default, gpt-5-4, main) with BLOCKING
anti-pattern enforcement.
🤖 Generated with assistance of [OhMyOpenCode](https://github.com/code-yeongyu/oh-my-opencode)
These agents should only be usable as primary session agents, not as
subagent targets via call_omo_agent/task(). Previously MODE was 'all'
which allowed them to be spawned as subagents, leading to confusing
behavior (e.g. Atlas delegating to Hephaestus as a subagent).
Subagent-callable agents remain: oracle, explore, librarian,
multimodal-looker, metis, momus, sisyphus-junior.
Note: Prometheus is not a BuiltinAgentName and is only invoked via
slash commands, so no change needed there.
Tighten Background Result Collection instructions so the model waits
for explore/librarian results instead of duplicating their work and
delivering premature answers.
- Remove 'Continue working immediately' which models interpreted as
'do ALL work yourself, ignore agents'
- Clarify step 2: only do DIFFERENT independent work while waiting
- Add explicit step 3: end response when no other work remains
- Add 'not for files you already know' to explore section header
Fixes#2124, fixes#1967
Non-Claude models skip planning and under-parallelize. Two new sections
injected only when model is not Claude:
- Plan Agent Dependency: multi-step tasks MUST consult Plan Agent first,
use session_id for follow-ups, ask aggressively when ambiguous
- Deep Parallel Delegation (rewrite): explicit '4 units = 4 agents'
pattern, each with clear GOAL + success criteria, all run_in_background
Sisyphus prompt instructed 'your next action is background_output' which
caused agents to repeatedly poll running tasks instead of ending their
response and waiting for the system notification.
- Replace 'STOP all other output' with 'end your response' (actionable)
- Add system-reminder notification mechanism explanation
- Add explicit 'Do NOT poll' prohibition
- Reduce background_cancel(all=true) mentions from 5x to 1x (Hard Blocks)
- Reduce Oracle collect obligation from 4x to 2x
- Remove motivational fluff ('blind spots', 'normal and expected')
Net: -2 lines, clearer mechanism, eliminates polling loop root cause.
- fix(hooks): skip todo continuation when agent has pending question (#1888)
Add pending-question-detection module that walks messages backwards
to detect unanswered question tool_use, preventing CONTINUATION_PROMPT
injection while awaiting user response.
- fix(config): allow custom agent names in disabled_agents (#1693)
Change disabled_agents schema from BuiltinAgentNameSchema to z.string()
and add filterDisabledAgents helper in agent-config-handler to filter
user, project, and plugin agents with case-insensitive matching.
- fix(agents): change primary agents mode to 'all' (#1891)
Update Sisyphus, Hephaestus, and Atlas agent modes from 'primary'
to 'all' so they are available for @mention routing and task()
delegation in addition to direct chat.
Counter Gemini's tendency to skip Phase 0 intent classification by
injecting a mandatory self-check gate before tool calls. Includes
intent type classification, anti-skip mechanism, and common mistake
table showing wrong vs correct behavior per intent type.
🤖 Generated with assistance of [OhMyOpenCode](https://github.com/code-yeongyu/oh-my-opencode)
Add buildDeepParallelSection() function that injects guidance for non-Claude
models on parallel deep agent delegation:
- Detect when model is non-Claude and 'deep' category is available
- Inject instructions to decompose tasks and delegate to deep agents in parallel
- Give goals, not step-by-step instructions to deep agents
- Update Sisyphus prompt builder to pass model and call new function
This helps GPT-based Sisyphus instances leverage deep agents more effectively
for complex implementation tasks.
🤖 Generated with assistance of OhMyOpenCode
Atlas and Sisyphus prompts instructed agents to use background_cancel(all=true)
before final answers. This destroys uncollected background task results and
contradicts existing NEVER directives in the Sisyphus prompt, causing agents
to lose explore/librarian outputs mid-session.
Replace with individual task cancellation pattern that preserves completed
task results while still cleaning up running disposable tasks.
Convert all markdown tables in Sisyphus and dynamic-agent-prompt-builder
to plain bullet lists for cleaner prompt rendering.
Add explicit Oracle safeguards:
- Hard Block: background_cancel(all=true) when Oracle running
- Hard Block: delivering final answer before collecting Oracle result
- Anti-Pattern: background_cancel(all=true) and skipping Oracle
- Oracle section: NEVER cancel, collect via background_output first
- Background Result Collection: split cancel/wait into separate steps
with explicit NEVER use background_cancel(all=true) instruction
Remove crash-causing name fields from 6 agent configs (sisyphus, hephaestus,
atlas, metis, momus, prometheus). The name field approach breaks opencode
because Agent.get(agent.name) uses name as lookup key.
Add getAgentConfigKey() to agent-display-names.ts for resolving display names
back to lowercase config keys (e.g. 'Atlas (Plan Executor)' -> 'atlas').
Expand prompt structure comment to 4-field format (CONTEXT/GOAL/DOWNSTREAM/REQUEST).
Update all explore/librarian task() examples across Sisyphus, Hephaestus,
Prometheus interview-mode, and both ultrawork variants with richer context
including downstream usage, scope limits, and return format expectations.
Add CONTEXT + GOAL + QUESTION + REQUEST structure to agent delegation examples.
This guides users to provide richer context when invoking explore/librarian agents.
Subagents (explore, librarian, oracle, etc.) now use their own fallback
chain instead of inheriting the UI-selected model. This fixes the issue
where explore agent was incorrectly using Opus instead of Haiku.
- Add AgentMode type and static mode property to AgentFactory
- Each agent declares its own mode via factory.mode = MODE pattern
- createBuiltinAgents() checks source.mode before passing uiSelectedModel
All documentation, agent prompts, and skill descriptions were still
referencing the old 'skills' parameter name for delegate_task, but the
tool implementation requires 'load_skills' (renamed in commit aa2b052).
This caused confusion and errors for users following the docs.
Fixes#1008
Co-authored-by: sisyphus-dev-ai <sisyphus-dev-ai@users.noreply.github.com>
Simplify sisyphus prompt by removing redundant sections and inline generation. Delete generate-sisyphus-prompt.ts as prompts are now managed directly in the agent definition.
Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-opencode)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
- Add 'CORRECT: Background Exploration' example with run_in_background=true
- Update PARALLEL_EXECUTION examples with correct params (subagent_type, run_in_background, skills)
- Ensures agents fire explore/librarian in parallel as background tasks
YeonGyu-Kim