Advanced agent and model configuration

Command results are masked and secret-detected before full filtered or truncated copies reach disk. Live tool output holds incomplete lines and private-key blocks so split credentials are not emitted; an overlong pending line pauses live output until the final result. Output storage uses owner-only real directories and exclusive files, refusing symlinked output directories even if replaced after an earlier save. Existing disk and retention limits still apply.

The daemon owns these additive settings.set fields. settings.get returns the saved configuration; null patch values clear a field. $schema is an editor/schema reference, never a URL the daemon fetches. Validation uses v2's protocol schemas.

{
  "$schema": "https://example.test/mendcode-config.json",
  "shell": "/bin/zsh",
  "watcher": { "ignore": ["**/generated/**"] },
  "enabled_providers": ["chatgpt", "local"],
  "disabled_models": ["local/retired"],
  "default_agent": "build",
  "plan_exit_agent": "reviewer",
  "subagent_variant": "quick",
  "agent": {
    "build": { "temperature": 0.2, "top_p": 0.9, "color": "accent", "steps": 8 },
    "reviewer": { "description": "Review code", "mode": "primary", "prompt": "Review and verify." }
  },
  "tool_output": { "max_lines": 2000, "max_bytes": 51200 },
  "models": { "title": { "model": "local:fast" }, "summary": { "model": "local:large" } },
  "provider": {
    "local": {
      "npm": "@ai-sdk/openai-compatible",
      "options": { "baseURL": "http://localhost:8080/v1", "timeout": 300000, "chunkTimeout": 60000 },
      "whitelist": ["fast", "large"],
      "models": {
        "fast": {
          "name": "Fast", "family": "local", "release_date": "2026-01-01", "status": "beta",
          "cost": { "input": 1, "output": 3, "cache_read": 0.1 },
          "limit": { "context": 64000, "input": 60000, "output": 4096 },
          "modalities": { "input": ["text", "image"], "output": ["text"] },
          "temperature": true, "reasoning": true, "tool_call": true,
          "headers": { "X-Route": "fast" },
          "options": { "seed": 4 },
          "variants": { "quick": { "effort": "low", "temperature": 0.3 }, "retired": { "disabled": true } }
        },
        "large": { "limit": { "context": 200000, "output": 8192 } }
      }
    }
  }
}

Agent settings merge with trusted layer files and override builtins; nested options merge. Project files retain their trust boundary. An invalid, hidden, disabled or subagent-only default fails session creation. Without an explicit default, the first enabled visible primary is used, starting with build. plan_exit_agent aliases planExitAgent and takes precedence; invalid implementation destinations fall back to a visible primary. Plan mode remains read-only until approval. An explicit task model/variant overrides agent defaults; subagent_variant applies when neither the task nor its agent chooses a variant. Hidden agents remain explicitly callable; they cannot be implementation destinations.

Model denies accept exact catalog IDs (local:fast) or donor IDs (local/fast); an empty provider allowlist permits none. Provider whitelist/blacklist apply to raw model IDs and also bind runtime requests. Provider settings overlay providers.json. Provider generation defaults, model options, selected variant and agent sampling apply in that order. Conversation, authentication and tool-routing fields cannot be overwritten through generation parameters. Disabled variants are omitted.

Model metadata exposes costs (dollars per million tokens), context/input/output limits, modalities, status, capabilities and release information. Output limits cap generated tokens; text-only models omit images with a note, models without tool calls receive no tool specs, and models without temperature support omit sampling fields. Usage without a provider cost uses configured input/output/cache pricing, including context_over_200k. Authentication headers are redacted in provider/model reports. Total timeout defaults to 300000 ms and chunk idle timeout to 60000 ms; false disables either.

An existing <dataDir>/models.dev.json (up to 8 MB, models.dev provider/model shape) supplies local metadata, costs, capabilities and variants. Explicit model entries overlay it. No catalog download is made. Known SDK names for OpenAI, Anthropic and OpenAI-compatible endpoints select v2's native wire adapters. Other SDK packages are reported as unsupported; an explicit supported wire type can be used for a compatible endpoint. Arbitrary SDK execution and live catalog/service validation remain outside this local, dependency-free batch.

Tool output previews use the global line/UTF-8 byte caps. Larger masked outputs are saved to <dataDir>/tool-output/ with read/grep hints: up to 8 MB per excerpt, 64 MB total, one day retention. These disposable files are separate from attachment blobs. Once the context has reached 80 % of the compaction threshold, older tool results are pruned from model history after protecting the latest two user turns and 40k tool tokens; pruning requires over 20k reclaimable tokens and protects skills. Persisted transcript items remain available.

Command output filters

tool_output.filters (or features.output_filters, /tier output_filters on|off) condenses what shell and shell_output return before the model sees it. It follows the tier: on in balanced and full, off in minimalist and with no tier; your own switch wins. The filters live in packages/engine-native/src/output-filters/, one file each, pure and run in this order: ansi (colour codes, \r progress redraws), tests (bun/jest/vitest/pytest/go/cargo passing lines fold to a count), git (status hints, diff index lines, transfer progress), tsc (diagnostics grouped under their file), search (grep/rg hits grouped under their file) and repeats (runs of three or more identical lines, blank runs). A line that looks like a failure (error, fail, warn, ✗, (fail)…) is never folded, and nothing is reordered.

Nothing is lost: the full masked output is saved first (same <dataDir>/tool-output/ store and limits) and the result starts with [filtered tests: 249 → 10 lines, ~2327 tokens saved. Full output: <path>]; read the path with read offset/limit or grep. With no way to save it, or a saving under 60 estimated tokens (the header must pay for itself), the output is returned untouched. A filtered result still over the line/byte caps is cut as before and says the unfiltered copy is at the path above. meta.savedTokens (chars / 4, an estimate, not billed) and meta.filter ride on the completed tool item; /usage > Tools sums them per filter as "Tokens saved". The recorded-run test in packages/engine-native/test/output-filters.test.ts is the budget: a 241-test passing run drops from over 2,000 to under 150 estimated tokens. Other tools (read, grep, edit…) are not filtered. No external binary and no model call is involved.

Frames: bun scripts/render-frames.ts .agents-frames/B42 100x30 180x50 --b42. The harness renders /agent, /model, its summary role selection and the colored-agent conversation through the app's factories. Additional compact/wide sizes are covered by the same harness tests.

Local instructions and skills (B43)

settings.set accepts instructions: ["rules/**/*.md", "/explicit/user/rules.md"] and skills: { paths: ["/explicit/user/skills"], urls: ["https://example.test/skills"] }. ~ paths expand to the home directory; relative paths resolve in each session project. User instruction globs also search project ancestors up to the repository root. AGENTS/CLAUDE/legacy CONTEXT fallback and global instructions remain; real paths deduplicate matches. Each file is bounded to 16,000 characters and the combined block to 64,000.

The project's .mendcode/project.json can declare the same fields. Project declarations must stay inside the project; external locations require an explicit user setting. Remote instruction/skill URL declarations are retained but never fetched. project config-show lists their blocked reasons. Network/trust policy and importing upstream licensed vendor prompt snapshots remain decisions; native family guidance and local custom/SYSTEM prompts work already.

.agents/.claude skills, commands and agents, AGENTS.override.md, instruction fallbacks and root markers, the shell environment policy, proxy, strict config and web domain rules are in docs/resources-config.md.

Skills search extension layers, then explicit configured folders recursively with bounded depth/count; configured names win. Only names/descriptions enter full-mode prompts; the skill tool and /skills load bodies on demand. Disabled skills remain discoverable and cannot be loaded or injected. Changes apply at the next native call without restart.