3.2 KiB
B2: per-profile plugin-enablement overlays — shipped + measured
Task: B2 in system-prompt-profiles-tasks.md · 2026-07-08 · Depends on C3 findings (c3-spike-findings.md)
Pre-check result: enabledPlugins DEEP-MERGES
The open item from C3 is resolved empirically (transcript skill_listing inspection, two
headless haiku runs): a --settings <overlay.json> file's enabledPlugins is merged
key-by-key into ~/.claude/settings.json — only the plugins explicitly set false
disappear (baseline 64 skills → 61, exactly the 3 skills of the 2 disabled plugins);
unlisted plugins and all hooks are unaffected. Diff-only overlays are safe.
Shipped
~/.claude/profiles/ (see its README.md):
cc-os-design.json— disables rails-ui-component, api-wrapper, invoice-ninja, creative-team, dev-teamclient-dev.json— disables creative-team, invoice-ninja, arch-refactorbrainstorm.json— disables all 9 coding/client plugins (adds ruby-lsp, codex, git-context, arch-refactor)
Usage: claude --settings ~/.claude/profiles/<name>.json. Never combine with a
--setting-sources list that omits user (silent plugin/hook drop — see vault note
claude-code-setting-sources-silent-plugin-drop).
Measured savings (headless haiku, empty dir, 2026-07-08)
| Run | Context tokens | Delta vs baseline |
|---|---|---|
| baseline (no overlay) | 29,045 | — |
| cc-os-design (5 plugins off) | 29,013 | −32 |
| brainstorm (9 plugins off) | 28,413 | −632 |
Honest verdict: token savings are marginal.
Follow-on: YAGNI tool deny (2026-07-08, same session)
Transcript sweep of all sessions since 2026-06-01 (tool_use counts across
~/.claude/projects/**/*.jsonl): heavy use of Bash (7.7k) / Read (4.2k) / Edit (2.1k) /
Agent (1.3k) / Write (1.2k); zero-or-one lifetime calls to Workflow, Artifact,
SendUserFile, ReportFindings — all of which ship full schemas up front (not deferred).
Measured (headless haiku, empty dir): disallowing Workflow alone −5,982 tokens
(its schema ≈ the whole system-prompt text); all four −6,509. A permissions.deny
block in a settings file removes the schemas identically to --disallowedTools
(22,532 vs 22,536).
Applied globally to ~/.claude/settings.json:
"permissions": {"deny": ["Workflow", "Artifact", "SendUserFile", "ReportFindings"]} —
~6.5k saved every session, no flags needed. Revert by deleting the entry (needed if
"ultracode"-style Workflow orchestration is ever wanted). This delivers the
"--tools restriction is the real lever" prediction from the assessment; A1/A2's
remaining value is now just the interactive-Fable confirmation. The assessment's "~4.8k user stack"
is dominated by hooks/memsearch injection and per-plugin fixed listings, not by the
skill descriptions the overlays remove — ToolSearch deferral already made plugin
skills cheap. The remaining value of profiles is focus (no irrelevant slash
commands/agents offered in a brainstorm session), not context cost. This further
deprioritizes Wave 3 (aliases: trivial, do when friction is felt) and Group D
(dropped), and shifts remaining token-diet attention to the memsearch injection
(~1.5–3.9k/session, needs upstream fix — see b3-memsearch-injection-review.md).