pi-model-sync

An extension for omp (oh my pi) and upstream pi: rewrites the model used by subsequent subagents based on the spawning session's current model. Two delegation paths are supported:

• omp native subagents (task / eval agent() / workpool workers).
• The subagent tool of pi-subagent (@mjakl/pi-subagent), under both omp and upstream pi (pi-subagent itself requires Pi ≥ 0.87.1 or a compatible host).

After the main model changes, newly spawned/delegated subagents switch to the model mapped by the matching profile; running subagents never switch mid-flight. Nested spawns (a subagent spawning further subagents) are intercepted too: the lookup key is the spawning session's current model.

Profiles live in a two-layer JSON config and can be grouped into named presets, with exactly one active preset per layer.

Installation

No build, no dependencies — both hosts run the TypeScript sources directly.

omp

# from the omp marketplace
omp plugin marketplace add SenCha930511/omp-marketplace
omp plugin install pi-model-sync@sencha930511
# or from git (pin a tag for a fixed version)
omp plugin install git:github.com/SenCha930511/pi-model-sync@v0.4.2
# or from npm (the same package feeds the Pi gallery)
omp plugin install npm:@sencha930511/pi-model-sync

upstream pi

# from npm (listed in the Pi package gallery at pi.dev/packages)
pi install npm:@sencha930511/pi-model-sync
# or straight from git
pi install git:github.com/SenCha930511/pi-model-sync@v0.4.2

One-off trial without installing: omp --extension /path/to/pi-model-sync / pi --extension /path/to/pi-model-sync (-e is repeatable; pi -e npm:@sencha930511/pi-model-sync works too).

Local / offline

1. Copy the whole directory to ~/.omp/agent/extensions/pi-model-sync/ (omp) or ~/.pi/agent/extensions/pi-model-sync/ (upstream pi); each package.json load key points at ./index.ts.

2. omp alternative: add to ~/.omp/agent/config.yml:

   extensions:
     - /path/to/pi-model-sync

Via your AI agent

This repo ships an llms.txt index (llmstxt.org convention), so coding agents can discover the project and install it directly. Ask yours:

Install pi-model-sync for my omp (or pi) setup.

The agent should run the host-matching command from the sections above, then verify with /model-sync list.

Config files

Two layers, same JSON format. A file is either a named-preset document or the older legacy profile list; both live at the same paths:

• Global: ~/.omp/agent/pi-model-sync.json
  • PI_MODEL_SYNC_CONFIG overrides the path (relative paths resolve against the session cwd).
  • When PI_CODING_AGENT_DIR is set: <its value>/pi-model-sync.json (lower precedence than PI_MODEL_SYNC_CONFIG).
  • Dual-host landing: when none of the above applies, if the ~/.omp file does not exist and ~/.pi/agent/pi-model-sync.json does, the .pi path is used; when neither exists, the write target stays the ~/.omp path.
• Project: <repo>/.omp/pi-model-sync.json (current directory only, no upward search)
  • Isomorphic fallback: when the .omp file does not exist and <repo>/.pi/pi-model-sync.json does, the .pi path is used.

An existing project file wholly replaces the global file (no per-key merge; a broken-JSON project file counts as "exists with no profiles"). Each layer carries its own presets and activePreset, so the two layers select presets independently.

Named presets

{
  "activePreset": "daily",
  "presets": {
    "daily": [
      {
        "main": "openai/gpt-6.1-sol",
        "subagents": {
          "*": "openai/gpt-6-luna",
          "scout": "openai/gpt-6-mini"
        }
      },
      {
        "main": "anthropic/claude-opus-*",
        "subagents": "anthropic/claude-sonnet-5"
      }
    ],
    "deep-work": [
      {
        "main": "openai/gpt-6.1-sol",
        "subagents": {
          "explore": "openai/gpt-6-mini"
        }
      }
    ]
  }
}

• presets: maps preset names to ordered Profile[] arrays. Names are single tokens (no whitespace), matched exactly and case-sensitively.
• activePreset: names the one preset whose profiles are in effect for this layer. When omitted, no preset is active and nothing routes.
• In deep-work only explore is bound; an agent with neither its own key nor a "*" default is not intercepted, so the host's default routing applies to it.
• Profile fields (main, subagents) and selector syntax are described below.

Legacy format

{
  "profiles": [
    {
      "main": "openai/gpt-6.1-sol",
      "subagents": { "*": "openai/gpt-6-luna", "scout": "openai/gpt-6-mini" }
    }
  ]
}

A legacy file behaves as a single virtual preset named default that is always active. Ordinary add/set/remove writes keep the legacy syntax; the first successful preset create or preset copy converts the file (see "Preset conversion and downgrade").

Profile fields

• main: the main model to match (rules below).
• subagents:
  • a string → every agent under this profile uses that model;
  • an object → keys are agent names (e.g. task, scout); "*" is the profile default. An agent with neither an exact key nor a "*" default is left alone.
• Selectors pass through verbatim to host resolution: provider/id, bare id, role aliases (e.g. @smol), and :level suffixes (e.g. openai/gpt-6-luna:high) are handed to the host resolver as-is. Under upstream pi, see "pi-subagent routing rules" for validation semantics.

Active preset selection

activePreset selects the whole profile list of its layer — profiles are never merged across presets or layers. Selection fails closed and never falls back:

• activePreset omitted → no profiles are active; /model-sync list and /model-sync show print No active preset is selected; no profiles are active.
• activePreset names a missing preset → no profiles are active; the warning is Active preset "<name>" does not exist; no profiles are active.
• A malformed preset document (a non-object JSON root, mixing profiles with presets/activePreset, activePreset without presets, a preset value that is not an array, or a blank/whitespace/duplicate preset name) is a config error: nothing routes, and write commands refuse to touch the file until it is repaired or deleted.

Routing on the spawn path is silent when it fails closed; /model-sync list and /model-sync show report the active preset and the warning.

Preset conversion and downgrade

• add, set, and remove edit only the active preset of the target layer; there is no --preset flag. Run /model-sync preset use <name> first to edit another preset.
• preset create creates an empty preset without switching to it; preset copy deep-copies a preset into a new name without switching and never overwrites an existing name; preset use is a no-op when the name is already active; preset remove refuses the active preset.
• On a legacy file the first successful preset create <name> or preset copy default <name> converts it: the existing profiles move to presets.default, activePreset becomes default, and the top-level profiles key is removed. (preset create default is refused because that name is the virtual legacy preset.)
• A missing/empty config is written in legacy form ({ "profiles": [] }) until a preset command converts it.
• Downgrade limit: files using presets/activePreset need an extension version with preset support. Older versions only read profiles — they see a preset file as empty (no routing) — and their next write rewrites the document as legacy, dropping presets and activePreset. Keep a legacy copy if you switch between versions.

Matching rules

• main without * → exact equality against provider/id or bare id (case-insensitive).
• main with * → explicit glob wildcard (* crosses /), tested against both the provider/id and bare id forms.
• No implicit substring matching; the first match in file order wins.

pi-subagent routing rules

The extension intercepts pi-subagent's subagent tool input {calls: [{agent, prompt, model?, ...}]} at tool_call time: when a call has no model, the spawning session's current model is matched against profile main, the call's agent name picks a selector (exact key first, then "*"; key matching is case-sensitive — align pi-subagent agent names with profile keys manually), and call.model is injected. When neither the exact key nor "*" exists, the call is not intercepted and pi-subagent's default routing applies. All other fields (prompt, thinking, cwd, session, ...) are preserved.

• An explicit per-call model always wins and is never overwritten. pi-subagent's own precedence is call model → agent frontmatter model → parent model; injection sits at the call-model layer, so an agent frontmatter model is overridden by profile routing — by design: the profile is the single source of "which model to use now".
• Named-session continuations get injected too.
• Shape guard: only a tool named subagent whose input carries a calls array is intercepted; other same-named tools with a different shape (e.g. nicobailon/pi-subagents) are not affected.
• Under omp, selectors are fully validated via ctx.models.resolve(); under upstream pi they are tier-checked:
  • @alias (an omp-only concept, unresolvable upstream) → unresolvable, no injection;
  • a :level suffix (off/minimal/low/medium/high/xhigh/max) → stripped before validating the base;
  • a glob containing * → passed through for the child process to resolve (a failure surfaces as an explicit error on that delegation);
  • provider/id (split at the first /) → looked up via modelRegistry.find; a bare id → looked up in modelRegistry.getAvailable()/getAll(); no hit, no injection.
• An unresolvable selector is not injected; pi-subagent's default routing (frontmatter → parent model) applies, with a one-time TUI warning (once per selector).
• An injected selector must be resolvable by the child process itself: pi-subagent spawns an independent child (upstream pi or omp), whose model registration sources (e.g. ~/.pi/agent/models.json) must know that selector.

/model-sync command

All command help, output, notifications, and config error messages are in English.

Commands reject extra arguments with a usage hint. Writes target the global file by default; any subcommand with --project writes the project file. add, set, and remove edit only the active preset of the target layer and refuse to write when no preset is active; preset create and preset copy leave the active profiles unchanged; preset list always reports the active layer, so --project does not change its output. You can also edit the JSON directly (changes are detected via an mtime+size cache and take effect on save).

Typing /model-sync or /model-sync preset opens argument completion with descriptions for each subcommand. preset use, preset remove, and the source argument of preset copy also suggest existing preset names from the target config layer. Preset names are global by default; include --project before the name to complete project-layer names.

Updating

No self-update machinery — the hosts' package managers own updates:

Notes

• Running subagents never switch mid-flight; switching the main model only affects later spawns.
• Nested spawns resolve against the spawning session's current model (a child session only knows its own model — the only clean API). This happens only when the host allows nesting depth ≥ 2; the cap is governed by task.maxRecursionDepth (omp doc default 2, negative = unlimited); at the cap, the subagent's task tool is removed so it cannot spawn further.
• When a profile selector is unresolvable, that spawn or delegation is not intercepted — default routing applies — and a warning shows once per selector in the TUI.
• For omp-native spawns, when routing takes effect the task card / Agent Hub shows the routing note: pi-model-sync: <agent> → <model> (main: <main>). pi-subagent delegations show the injected model in the call rendering instead — no extra note, by design.
• Startup is fully silent: no session-start notification.

Contributing

See CONTRIBUTING.md. Review expectations and conventions live in AGENTS.md.

License

MIT © SenCha930511.