🧩 v0.10.0 epic: 3rd-party resource integration (discover -> vet -> install -> hot-load)

[ZaxShen]

v0.10.0 theme: 3rd-party resource power

Make TMB dramatically more capable by letting bro discover, vet, install, and use external resources — Claude Code skills, toolkits/MCP servers, and plugins — on demand, instead of being limited to what ships in the box.

Motivating example

User: "build a dashboard website frontend." bro recognizes it lacks a strong frontend capability, searches ("2026 best Claude Code skills for dashboard website frontend"), finds e.g. a shadcn skill, vets it, asks the user's permission to install locally, installs it, then loads it (or routes the proper agent to load it) — ideally without a session restart.

The two keys

1. Search — bro autonomously finds relevant 3rd-party resources when a task would benefit.
2. Smooth use — once installed, TMB uses it cleanly in the same session. If a restart is genuinely unavoidable, bro tells the user to run claude --resume (TMB state lives in trajectory-db checkpoints, so resume is safe).

Sub-issues (the pipeline: discover -> vet -> install -> hot-load)

• v0.10.0: 3rd-party discovery/search — detect capability gap, web-search, rank candidates #657 — Discovery/search engine: detect capability gap, web-search, rank candidates
• v0.10.0: trust & vetting gate — assess source reputation + security surface before proposing #658 — Trust & vetting gate: source reputation + security surface before proposing
• v0.10.0: permission-gated local install — propose vetted resource, install on explicit approval #659 — Permission-gated local install: propose, get explicit approval, install into local env
• v0.10.0: hot-load / in-session integration — use installed resource without restart, else claude --resume #660 — Hot-load / in-session integration: use without restart, else claude --resume

Constraints carried from prior doctrine

• Installs are human-approved (never silent); align with the existing permission model.
• Plugin installs follow the marketplace-install path (no seeding/copying) — see benchmarks standing rule.
• Discovery via web search must respect any web-deny posture; surface sources to the user.

⚠️ Determinism mandate (applies to EVERY sub-issue)

Every new hook, MCP tool, and skill added under v0.10.0 must 100% obey docs/prompt-engineering/DETERMINISM.md. Concretely:

• Anything that could fail because the LLM forgot, misordered, or misunderstood it goes in a deterministic layer (atomic composite MCP tool in one DB transaction, PreToolUse block hook, PostToolUse side-effect hook), not in prose.
• Skills/agent prose are reserved for irreducibly judgment-bound steps (e.g. "is this candidate trustworthy"). The mechanical pipeline (search → record → vet → install → activate) must be deterministic tools/hooks, not a checklist the model is asked to follow.
• Apply the boundary test and authoring checklist from DETERMINISM.md to every procedural line before merge. Multi-step sequences collapse to one atomic MCP call; "don't proceed until X" becomes a PreToolUse hook; "after A also do B" becomes a PostToolUse hook.

[ZaxShen]

Superseded by the completed cheatcode lifecycle cluster (epic #823, all sub-issues merged to dev). The discover→vet→install→hot-load pipeline is built and now faithful: search/vet (cheatcode_search/vet), permission-gated install (cheatcode_approve + cheatcode_install via the marketplace path #815/#829), materialization into the consuming agent (#822/#841), honest uninstall (#821), status/health (#819/#820), and an L6 that asserts the materialization on disk (#778). Hot-load model = cold-start-between-steps with trajectory-db rehydration (+ cheatcode_activate + --resume fallback), per the L6 design. Closing per Human direction.