YeonGyu-Kim
2dfa6336f5
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
89 lines
6.5 KiB
Markdown
89 lines
6.5 KiB
Markdown
# GPT-5.5 System Prompt Drafts
|
|
|
|
This directory contains ground-up rewrites of the Sisyphus, Hephaestus, Oracle, and Deep system prompts, styled after OpenAI Codex's gpt-5.4 prompt architecture and targeted at GPT-5.5.
|
|
|
|
## Files
|
|
|
|
- `sisyphus.md` — Orchestrator. Intent gate, delegation philosophy, parallel execution discipline, verification.
|
|
- `hephaestus.md` — Autonomous deep worker. Persistence, exploration-first, forbidden stops, root-cause bias.
|
|
- `oracle.md` — Read-only strategic advisor. Three-tier response structure, hard verbosity limits, confidence signaling.
|
|
- `deep.md` — Category-spawned deep worker (runs as Sisyphus-Junior under the `deep` category). Goal-oriented autonomous execution.
|
|
|
|
## Design principles applied
|
|
|
|
Each prompt applies the same small set of principles, borrowed and adapted from Codex's gpt-5.4 prompt work:
|
|
|
|
1. **Single identity header with `{{ personality }}` slot.** Separates persona from logic so the same base prompt can ship in default / friendly / pragmatic variants without duplication.
|
|
2. **`# General` → `## Autonomy and Persistence` → `## Task execution` → `## Validating your work` → `# Working with the user` → `# Tool Guidelines` structure.** Lifted directly from Codex's `gpt_5_2_prompt.md` and `gpt-5.2-codex_prompt.md`. Keeps the same section contract for every agent so readers can navigate consistently.
|
|
3. **Prose-first output, bullets only when list-shaped.** GPT-5.5 reads and writes prose naturally; bullet overuse is a GPT-5.3 coping mechanism, not a genuine formatting need.
|
|
4. **Contract frames over threat frames.** Rules are stated as agreements and expectations, not as "NEVER DO X OR YOU WILL FAIL". GPT-5.5's instruction following is strong enough that threats add entropy without improving compliance.
|
|
5. **Opener blacklist is explicit.** "Done —", "Got it", "Great question", "Sure thing", and similar filler are called out by name. These are the most common failure modes across all models.
|
|
6. **File reference formatting is unified.** Clickable markdown links with absolute paths, no `file://` or `https://` for local files, no line ranges.
|
|
7. **Why, not just what.** Each major rule is accompanied by the reasoning. Rules without reasons get ignored when models judge them weakly-grounded; rules with reasons get applied even in novel situations.
|
|
|
|
## Agent-specific shape
|
|
|
|
### Sisyphus
|
|
- Intent classification table (surface form → true intent → routing).
|
|
- Zero-tolerance visual-engineering delegation rule.
|
|
- Six-section delegation prompt contract.
|
|
- Session continuity (`task_id` reuse) as a first-class topic.
|
|
- Oracle consultation as a separate section with clear use/not-use guidance.
|
|
|
|
### Hephaestus
|
|
- Forbidden stops as a named list.
|
|
- Three-attempt failure protocol.
|
|
- Exploration-first as explicit philosophy (5-15 minutes is normal).
|
|
- "Dig deeper" subsection for root-cause bias.
|
|
- Ambition vs precision distinction for greenfield vs existing codebase work.
|
|
- Task-tool restriction stated as an intentional design decision with rationale.
|
|
|
|
### Oracle
|
|
- Three-tier response structure (Essential / Expanded / Edge cases) with hard numerical limits.
|
|
- Effort estimation (Quick / Short / Medium / Large) as a required field.
|
|
- Confidence signaling (high / medium / low) added as a required field — new in v5.5, borrowed from Codex's `review_prompt.md`.
|
|
- Pragmatic minimalism as explicit decision framework.
|
|
- "No commentary channel; every word is the final answer" constraint acknowledged.
|
|
|
|
### Deep
|
|
- Explicitly positioned as Sisyphus-Junior in `deep` mode (category-spawned counterpart to Hephaestus).
|
|
- Extensive exploration expectation stated.
|
|
- Final-answer structure tuned for orchestrator relay: "What changed / Key decisions / Verification / Observations / Blockers".
|
|
- Commentary cadence tuned down (sparse) since the user is not directly on the other side.
|
|
|
|
## Known deviations from Codex
|
|
|
|
These are intentional choices where oh-my-opencode's architecture differs from Codex's:
|
|
|
|
- **`task()` delegation is central** for Sisyphus (it is the orchestrator), entirely absent for Oracle (read-only consultant), research-only for Hephaestus and Deep (they execute directly).
|
|
- **No `update_plan` tool**; the harness uses `task_create` / `task_update` instead. Each prompt references its own tool set.
|
|
- **Sub-agent ecosystem** (explore, librarian, oracle, metis, momus) is specific to this harness and does not exist in Codex. Each prompt explains when and how to use these agents.
|
|
- **Skill loading** is a first-class concept via the `skill` tool. Codex has a simpler skill model.
|
|
- **Commentary / final channels** are named the same way as Codex's output contract, but the actual transport layer is different (OpenCode, not Codex CLI).
|
|
|
|
## Line counts
|
|
|
|
For reference, approximate line counts after this rewrite versus the current production prompts:
|
|
|
|
| Agent | Current (assembled) | Draft | Delta |
|
|
|---|---:|---:|---:|
|
|
| Sisyphus GPT-5.4 | ~500 | ~270 | -46% |
|
|
| Hephaestus GPT-5.4 | ~400 | ~270 | -33% |
|
|
| Oracle GPT | ~120 | ~160 | +33% |
|
|
| Deep category append | ~20 | ~250 (as standalone) | N/A |
|
|
|
|
Oracle grew because v5.5 adds Confidence signaling and explicitly documents follow-up session behavior. Deep grew because the draft is a standalone prompt rather than a category append; in production it would either replace Sisyphus-Junior's GPT-5.5 variant entirely or layer on top of a minimal Sisyphus-Junior base.
|
|
|
|
## What this draft is not
|
|
|
|
- **Not a `.ts` file.** These are markdown drafts. Converting to TypeScript template strings (with `{todoHookNote}`, `{keyTriggers}`, etc. interpolation) is the next step, once the content is validated.
|
|
- **Not a tested prompt.** These have not been run against evals. Before shipping, each prompt should be benchmarked with `skill-creator`'s eval loop against the current production prompts on a representative task set.
|
|
- **Not personality-substituted.** The `{{ personality }}` slot is a placeholder. Default / friendly / pragmatic content still needs to be authored.
|
|
|
|
## Suggested next steps
|
|
|
|
1. **Author personality variants.** Three short paragraphs (default, friendly, pragmatic) that slot into `{{ personality }}` and can be reused across all four prompts.
|
|
2. **Build an eval harness.** Pick 5-10 representative tasks per agent and run current-prod vs draft-v5.5 head-to-head.
|
|
3. **Convert to `.ts` with dynamic composition helpers.** Preserve the existing `buildAgentIdentitySection`, `buildToolSelectionTable`, etc. integration points where they still apply.
|
|
4. **Ship behind a feature flag.** Opt-in for `gpt-5.5` model selection until eval confidence is high.
|