fix(agents): omit model: inherit so non-Claude tools don't see a bad model id

[AsimRaza10]

Summary

Fixes #167 (ProviderModelNotFoundError when running /understand on opencode + deepseek and similar setups).

model: inherit is a Claude Code-specific keyword meaning "use the parent session's model." Other tools that read the same agent frontmatter (opencode, codex, etc.) don't understand it — they treat inherit as a literal model id and forward it to whatever provider the user has configured, which rejects it.

Reproduction (from the issue):

I am using opencode with deepseek v4

│ Project-Scanner Task — Phase 1 SCAN project
ProviderModelNotFoundError

Maintainer already acknowledged: "Oops, seems deepseek thought inherit is a model name…."

Approach

Omit the model: field from all 9 agent frontmatters. With the field absent, each platform falls back to its own configured default:

• Claude Code → user's default subagent model
• opencode / codex / Gemini CLI / Antigravity / etc. → their globally configured model

This is the smallest change that unblocks non-Claude users without requiring install-time file rewriting (which would break the current symlink-based install flow in install.sh).

Trade-off worth flagging

Claude Code users running an Opus session previously had subagents auto-inherit Opus. After this change, subagents follow Claude Code's default-subagent-model setting instead — a single global config now controls all 9 agents. That's a behavior change worth surfacing in release notes; for users who want the old behavior, configuring the default subagent model globally restores it.

If you'd rather preserve model: inherit for Claude Code and strip it only for other platforms, that's a bigger change (install.sh would need to copy + transform agent files rather than symlink), and I'm happy to do that as a follow-up if you'd prefer.

Files changed

File	Change
understand-anything-plugin/agents/architecture-analyzer.md	drop model: inherit
understand-anything-plugin/agents/article-analyzer.md	drop model: inherit
understand-anything-plugin/agents/assemble-reviewer.md	drop model: inherit
understand-anything-plugin/agents/domain-analyzer.md	drop model: inherit
understand-anything-plugin/agents/file-analyzer.md	drop model: inherit
understand-anything-plugin/agents/graph-reviewer.md	drop model: inherit
understand-anything-plugin/agents/knowledge-graph-guide.md	drop model: inherit
understand-anything-plugin/agents/project-scanner.md	drop model: inherit
understand-anything-plugin/agents/tour-builder.md	drop model: inherit
CLAUDE.md	update the agent-pipeline note to reflect the new convention and reference the issue

Test plan

• All 9 agent frontmatters still parse (verified visually — frontmatter delimiters intact, no orphan keys)
• No source-side test or schema references model: inherit (grep returns clean)
• Maintainer to verify against Claude Code, opencode + Anthropic, and opencode + deepseek
• Maintainer to decide whether the global-default trade-off is acceptable, or whether to go with the per-platform install transformation instead

Closes #167

[AsimRaza10]

@Lum1104 friendly ping when you get a chance — small agent config fix so Cursor/opencode don't choke on model: inherit as an invalid model id. Low-risk diff, happy to adjust scope or rework if you'd prefer a different approach. Thanks!

[Lum1104]

Hi, I was wondering how to do the update if we let install.sh to remove this line. Since git pull won't work after that.

[AsimRaza10]

Good catch — you're right, that's exactly why I went with the source-side approach rather than install-time rewriting. If install.sh mutated agent files in $REPO_DIR/understand-anything-plugin/agents/ directly, the next --update would hit git pull --ff-only against a dirty working tree and refuse (or conflict if upstream touched the same lines).

If we ever wanted to bring back per-platform transformation, the only way to keep --update working would be to stop linking the source files and instead render them to a separate dir, e.g.:

1. After git pull, run a render step that copies agents/*.md → $REPO_DIR/.rendered/<platform>/agents/*.md, applying per-platform transforms (strip model: inherit for non-Claude, keep it for Claude Code).
2. Symlinks point at .rendered/<platform>/ instead of the source tree.
3. .rendered/ is gitignored, so git pull stays clean.

That's more moving parts than the current symlink design buys us, though, and it adds a re-render step every update. The merged approach trades the Claude-Code Opus auto-inheritance behavior for one global subagent-model setting, which I think is the better deal for a multi-platform plugin — but happy to do the rendering pass as a follow-up if you'd rather restore the old Claude Code behavior without breaking other tools.