Replace 'FN.' shorthand with explicit examples 'F1.', 'F2.' to prevent
LLMs from generating literal 'FN.' labels that the parser rejects.
Identified by cubic.
- Strengthen Prometheus plan template with FORMAT constraint
- Add task label format check to Oracle phase-2 (N/6 → N/7)
- Add format checks to self-review checklist
- New plan-format-validator hook: compares raw checkbox count
against getPlanProgress() after plan writes, warns agent when
labels are malformed (0/0 or partial skip scenarios)
The multimodal-looker prompt described what to do but never told the model which tools are available. Smaller VL models (e.g. Qwen3-VL-8B) would try to call non-existent tools and enter an infinite loop emitting:
Model tried to call unavailable tool 'invalid'. Available tools: call_omo_agent, read.
Add a single sentence at the top of the prompt that explicitly enumerates the only allowed tools ('read' and 'call_omo_agent') and forbids calling any other tool. This matches the runtime allowlist enforced by createAgentToolAllowlist(["read"]).
Regression test asserts the prompt contains the available-tools enumeration so future prompt rewrites don't regress.
Skills with an `agent` frontmatter field are intended for a specific
agent. Previously they still appeared in:
- every agent's system prompt (via `buildAvailableSkills`)
- the `skill` tool's `<available_items>` description visible to all agents
This wasted tokens and could mislead agents into attempting calls that
would be rejected at execution time.
Changes:
- `buildAvailableSkills`: new optional `agentName` parameter; when
provided, skills whose `definition.agent` does not match are excluded
- `builtin-agents.ts`: pass per-agent name to `buildAvailableSkills`
for sisyphus, hephaestus, and atlas, so each agent's prompt only
lists the skills it is allowed to use
- `createSkillTool` (`tools.ts`): exclude agent-restricted skills from
both the eager and lazy description builds, keeping the shared tool
description free of skills the current agent cannot access
Execution-time enforcement (throwing on mismatch) is unchanged; this
change adds the earlier, description-level visibility gate.
Tests: new `available-skills.test.ts` (5 cases) + 3 new cases in
`tools.test.ts` covering the description-filter and execute paths.
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
Tests that mock an AgentFactory were using an `as AgentFactory` cast
and a separate mutation of `mockFactory.mode` to satisfy the type.
Replace with Object.assign so the factory type is constructed correctly
without casts. Also type the empty discoveredSkills fixture so its
element type is inferred from the function signature instead of
collapsing to never[].
The task() tool schema declares run_in_background as a REQUIRED parameter, but three prompt-template files contained category-based task() examples that omitted it. When agents copied these patterns into actual tool calls, the runtime validator threw 'run_in_background parameter is REQUIRED', breaking delegation.
Add run_in_background=false to the four offending category-based examples: 2 in src/agents/dynamic-agent-category-skills-guide.ts (Delegation Pattern + Category Domain Matching CORRECT/WRONG pair), 1 in src/agents/hephaestus/gpt-5-3-codex.ts (frontend example), and 1 in src/agents/sisyphus/gemini.ts (Example 4 DELEGATE). Subagent-type explore/librarian examples already correctly use run_in_background=true. atlas/shared-prompt.ts examples were already correct.
Verification: bun test src/agents/ — 373 pass, 0 fail. bun run typecheck — clean. Grep confirms no remaining category task() examples lack the parameter.
Inserts blocking Oracle verification todos (plan-1b / plan-2b /
plan-6b in the canonical, plan-1b / plan-2b / plan-5b in the gpt and
gemini variants) between each major Prometheus phase. Each gate is a
single task(subagent_type=oracle) invocation that must return
VERDICT: GO; NO-GO is a directive to fix the cited issues and rerun on
the same Oracle session, not a license to skip.
Adds a new 'Oracle Verification (Phase Gates)' section to
plan-generation.ts with the concrete invocation prompts for each gate:
phase 1 verifies interview completeness, phase 2 verifies the generated
plan, phase 3 verifies plan readiness for execution before /start-work
handoff.
Also adds a plan-generation.test.ts smoke suite (9 cases) that pins
the new todo ids, the section name, the GO/NO-GO format, the
'fix the cited issues' fallback, and the relative ordering.
Drops 'Maximum 3 retries' / 'document and move on' across every Atlas
variant (default, opus-4-7, gpt, kimi, gemini). New text forbids the
'false positive' excuse explicitly and instructs Atlas to keep iterating
on the same task_id, attaching a diagnosis plan, until verification
passes — and to spawn a different-angle subagent only when the original
loops.
Adds a shared <boulder_completion_response> section composed by
shared-prompt.ts. When the hook injects the BOULDER COMPLETE nudge,
Atlas now knows to print TOTAL ELAPSED + per-task elapsed times in the
exact summary shape, confirm boulder.json state, and only mark
pass-final-wave after the Final Wave reviewers approve.
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).
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
- 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>
Compress 6 discovery sections into one Discovery & Retrieval section, remove three-way echo of "drive the artifact" between Goal, Manual QA Gate, and Forbidden stops (now once + cross-ref), split persona prose into Tone vs Autonomy & Collaboration per OpenAI's GPT-5.5 prompting guide, trim Output formatting from 13 prior-restating bullets to 5 environment-specific ones, and defer the Oracle wait policy to the dynamic oracleSection so it lives in a single source of truth.
Add buildFrontendGuidanceSection: conditional injection that only emits the anti-slop frontend block when `visual-engineering` category is unavailable, since the category-skills guide already forces visual work to delegate when the category exists.
Static prose drops 277 -> ~175 lines (-37%); every Hephaestus identity item preserved (Manual QA Gate surface map, Intent table + commitment line, three-attempt failure protocol, Hard invariants, no defensive code, AGENTS.md and dirty worktree handling, four dynamic placeholders).
- Add Manual QA Gate as a non-negotiable surface-tool mapping
(interactive_bash for TUI, playwright for browser, curl for HTTP,
driver script for library) to Hephaestus, Sisyphus-Junior, and
Sisyphus' direct-execution mode.
- Restore the dig-deeper trio (tool persistence / dig deeper /
dependency checks) as orthogonal paragraphs so each carries its
own cognitive trigger instead of a fused single statement.
- Harden investigate-before-acting from a soft phrase to a dedicated
block: never speculate about unread code, re-read on every
task hand-off, the worktree may have changed.
- Add 'Parallelize aggressively' as its own block. Reads, searches,
diagnostics, and background sub-agents all batch into a single
response by default.
- Add 'No defensive code, no speculative legacy' to discourage
speculative backward-compatibility branches and unrequested
defensive validation.
- Absorb review-intent and frontend anti-slop coverage so the
prompt stays self-sufficient when the omo agent prompt replaces
the provider prompt.
- Replace literal apply_patch instructions with GPT_APPLY_PATCH_GUIDANCE
(use edit/write tools) so the prompt no longer contradicts the
apply_patch deny that the agent permission applies on GPT models.
- Sisyphus-Junior gains a Review tasks block and a default-behavior
fallback for when the runtime category context is missing or sparse.
- Sisyphus gains an explicit Hard invariants block listing type-
suppression bans, destructive-git bans, and Oracle-completion gating.
- Restore dynamic injections that round out the orchestrator/worker
context: category+skills delegation guide, delegation table,
Oracle dynamic guidance, key triggers, non-Claude planner
reminder. Hephaestus regains optional category delegation while
keeping direct execution as the default.
- Drop em dashes; search guidance points at rg directly throughout.
YeonGyu-Kim