Rewrite Atlas GPT verification from a checklist to a 4-phase protocol:
Phase 1 (Read Code First), Phase 2 (Automated Checks), Phase 3 (Hands-On QA),
Phase 4 (Gate Decision). Hands-on QA is now mandatory for user-facing changes,
not 'if applicable'. Hook message reinforces subagent distrust and requires
actually running deliverables before proceeding to next task.
Oracle, Librarian, Explore, Momus, and Metis could modify files via
apply_patch despite being read-only agents. Also fixed duplicate task
entries in Librarian and Explore restriction lists.
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
- Restore Assumptions Check and When to Challenge the User from Sisyphus intent gate
- Add proactive explore/librarian firing to CORRECT behavior list
- Strengthen parallel execution with GPT-5.2 tool_usage_rules (parallelize ALL independent calls)
- Embed reporting into each Execution Loop step (Tell user pattern)
- Strengthen Progress Updates with plain-language and WHY-not-just-WHAT guidance
- Add post-edit reporting to Output Contract and After Implementation
- Fix Output Contract preamble conflict (skip empty preambles, but DO report actions)
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').
Reduce plan-template from 541 to 335 lines by removing redundant verbose
examples while recovering 3 lost context items: tool-type mapping table in
QA Policy, scenario specificity requirements (selectors/data/assertions/
timing/negative) in TODO template, and structured output format hints for
each Final Verification agent.
Strengthen TODO template to make QA scenarios non-optional with explicit
rejection warning. Add Final Verification Wave with 4 parallel review
agents: oracle (plan compliance audit), unspecified-high (code quality),
unspecified-high (real manual QA), deep (scope fidelity check) — each
with detailed verification steps and structured output format.
Add Maximum Parallelism Principle as a top-level constraint and replace
small-scale plan template examples (6 tasks, 3 waves) with production-scale
examples (24 tasks, 4 waves, max 7 concurrent) to steer the model toward
generating fine-grained, dependency-minimized plans by default.
isGptModel only matched openai/ and github-copilot/gpt- prefixes, causing
models like litellm/gpt-5.2 to fall into the Claude code path. This
injected Claude-specific thinking config, which the opencode runtime
translated into a reasoningSummary API parameter — rejected by OpenAI.
Extract model name after provider prefix and match against GPT model
name patterns (gpt-*, o1, o3, o4).
Closes#1788
Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-opencode)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
Port devxoul's PR #821 feature to current codebase structure.
Supports absolute, relative, ~/home paths with percent-encoding.
Gracefully handles malformed URIs and missing files with warnings.
Co-authored-by: devxoul <devxoul@gmail.com>
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.
- VERIFICATION_REMINDER: add Step 2 manual code review (non-negotiable)
- Require Read of EVERY changed file line by line
- Cross-check subagent claims vs actual code
- Verify logic correctness, completeness, edge cases, patterns
- Add Step 5: direct boulder state check via Read plan file
- Count remaining tasks directly, no cached state
- BOULDER_CONTINUATION_PROMPT: add first rule to read plan file immediately
- verification-reminders.ts: restructure steps 5-8 for boulder/todo checks
- Atlas default.ts (Claude): enhance 3.4 QA with A/B/C/D sections
- A: Automated verification
- B: Manual code review (non-negotiable)
- C: Hands-on QA (if applicable)
- D: Check boulder state directly
- Atlas gpt.ts (GPT-5.2): apply same QA enhancements with GPT-optimized structure
- verification_rules: update both Claude and GPT versions with manual review requirements
Addresses issue where Atlas would skip manual code inspection after delegation,
leading to rubber-stamping of broken or incomplete work.
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.
When task_system is enabled, the prompt said 'task tool: BLOCKED' which
LLMs interpreted as blocking task_create/task_update/task_list/task_get
too. Now the constraints section explicitly separates 'task (agent
delegation tool): BLOCKED' from 'task_create, task_update, ...: ALLOWED'
so Junior no longer refuses to use task management tools.
sisyphus-junior prompt always used todo-based discipline text regardless of
experimental.task_system setting because the useTaskSystem flag was never
forwarded from agent-config-handler to createSisyphusJuniorAgentWithOverrides.
- Update main AGENTS.md with current file sizes
- Update complexity hotspot line counts
- Update agent count from 11 to 32 files
- Update CLI utility count to 70
- Update test file count from 100+ to 163
🤖 Generated with assistance of OhMyOpenCode
- Change PROMETHEUS_PERMISSION bash from 'allow' to 'deny' to prevent unrestricted bash execution
- Prometheus is a read-only planner and should not execute bash commands
- The prometheus-md-only hook provides additional blocking as backup
Revert .default([]) on load_skills schema back to required, restore the runtime error for missing load_skills, and add explicit load_skills=[] to all task() examples in agent prompts that were missing it.
Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-opencode)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
