…er-agent opt-out Resolves two related issues users have hit on the chat.system.transform hook, both rooted in injecting Magic Context guidance into agents that don't benefit from it: Issue #52 — small_model / title-gen / summary / compaction get the full `## Magic Context` block injected. These hidden agents have a fixed single-shot job, no tools, and no `ctx_reduce`. Now skip ALL injection (guidance, project docs, user profile, key files, sticky date) when the system prompt opens with the distinctive heads from OpenCode's `title.txt`, `summary.txt`, or `compaction.txt`. Hash is also NOT updated for these calls so they don't trigger spurious cache busts on later main-agent turns. Issue #53 + #42 — users want to opt specific agents (e.g. council orchestrator, read-only QA agents) out of injection because the guidance tells the LLM to use tools that the agent has been denied. Adds a new top-level config: system_prompt_injection: { enabled: true, // global escape hatch skip_signatures: [""] // per-agent opt-out } When `enabled=false`, no injection happens for any agent. When any signature substring matches the system prompt, that specific call is opted out — same no-injection-no-hash-update behavior as the internal-agent path. Empty strings in skip_signatures are filtered out defensively (would otherwise match all prompts and silently disable injection globally). For omo-slim users on issue #42, opting council out is now a one-line config: skip_signatures: ["You are the Council agent — a multi-LLM"] Tests: +12 new (5 internal-agent + 6 issue-#53 + 1 schema). Full suite 1061/1061 pass. Typecheck + lint + build clean across all 3 packages.

System prompt injection into small_model #52

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions