v0.9.14: plugin suggestion matcher nags every turn on topic keywords — gate to once-per-session and skip domains a loaded skill covers

[Hmbown]

Why this matters

Every user turn, the engine appends a <recommended_plugins> fragment matched on conversation keywords. A session that merely talks about a topic gets the same plugin nag repeatedly, and the matcher cannot tell that a locally loaded skill already covers the domain. Users dismiss or ignore the prompt and it comes back next turn — it reads as spam, and each injection is wasted context.

Current behavior

• crates/tui/src/core/engine.rs:3711 — user_content_blocks calls recommended_plugins_user_fragment(&text, …) on every user turn.
• crates/tui/src/plugins/recommend.rs:232 — match_plugin_for_draft keyword/name matches score 700–900 while PROACTIVE_MIN_SCORE is 700, so a topical keyword always clears the proactive bar.
• Dismissals are honored (settings.dismissed_plugin_suggestions), but there is no once-per-session cap and no loaded-skill awareness.

Repro (live, this session, 4 hits)

One real session, four different injections, each tracking the topic of the user's message rather than a capability gap:

1. takeover/handoff talk → suggested handoff (cw2:handoff) while the local cw-handoff skill was already loaded
2. continuation mentioning research → research (cw2:research)
3. goal-continuation text → remember (anthropic:remember)
4. a pasted summary mentioning release → release (cw2:release)

Desired behavior

1. A plugin id is injected at most once per Engine lifetime (evaluate once, consume — the grokbuild pattern: refs/grokbuild …/app_view.rs ssh_wrap_tip_evaluated one-shot tips; kimicode's capability catalog is pull-based with no proactive keyword injection).
2. Suppress a suggestion when a loaded local skill already covers the plugin name/domain (skill handoff loaded → never suggest plugin handoff).
3. Keep honoring dismissals; keep the fragment append-only on the user turn (never the pinned system prefix).

Acceptance criteria

• Same plugin id never appears in two fragments in one Engine session (unit test over user_content_blocks or the fragment fn with a session-state set)
• A loaded skill with a matching name suppresses the plugin suggestion (unit test)
• Dismissed plugins still never return (existing behavior kept, test still passes)
• cargo test -p codewhale-tui --lib --locked green for the touched modules

Related

• agent tool description claims the write-contention refusal names release; it does not, and the liveness filter says it could not #6272 (adjacent remedy-text accuracy in the same suggestion surface)