codeburn optimize --apply: execute config-class fixes with undo

[iamtoruk]

codeburn optimize --apply: execute config-class fixes with undo

Part of the acting-layer epic. Depends on the action journal issue; do not start before it lands.

What to build

codeburn optimize --apply turns the config-class findings that optimize already prints into applied, journaled, undoable changes. Behavioral findings (session openers, prompts) are explicitly not appliable here; they belong to the guard issue.

Step 1: stable finding ids

WasteFinding (src/optimize.ts) has no id today. Add id: FindingId to the type and to every detector, and include it in the JSON report shape (OptimizeJsonReport.findings[]). Ids are kebab-case and stable across runs:

read-edit-ratio, mcp-low-coverage, mcp-project-scope, low-worth-sessions,
context-heavy-sessions, cost-outliers, redundant-rereads, unused-skills,
unused-agents, unused-commands, build-folder-reads, bash-output-cap,
claude-md-too-long, warmup-heavy, retry-heavy-capabilities

(Match the actual detector list in the file; the names above cover the ones visible in current output, extend if the code has more.)

Step 2: appliable plans

Add src/act/plans.ts with planFor(finding): ActionPlan | null. A plan describes concrete file mutations for the journal framework (see src/act/types.ts). Appliable finding ids and their exact semantics:

mcp-low-coverage and mcp-project-scope (kind: mcp-remove / mcp-project-scope)

• Locate each server by reading, in order: project .mcp.json (mcpServers key), project .claude/settings.json and .claude/settings.local.json (mcpServers), user ~/.claude.json (top-level mcpServers and per-project projects[<path>].mcpServers). Handle only servers found in these files; if a server comes from somewhere else (plugin, managed config), mark it not-appliable with a note in the plan preview.
• Remove: delete the server entry from the file(s) it appears in.
• Project-scope: remove the entry from the global location and add it verbatim to <project>/.mcp.json for each project the finding lists as a keeper (create the file with { "mcpServers": {} } scaffold if missing).
• JSON edits must preserve the rest of the document; parse, mutate, JSON.stringify(obj, null, 2) plus trailing newline. If a target file fails to parse, skip that server and surface the parse error in the summary rather than aborting the whole apply.

unused-skills, unused-agents, unused-commands (kind: archive-*)

• Move each directory or file to a sibling .archived/ directory (~/.claude/skills/.archived/<name> etc., matching the paths the detector reports). Create .archived/ on demand. Name collision in the archive: suffix -2, -3.

claude-md-rule destinations (kind: claude-md-rule)

• For findings whose fix is { type: 'paste', destination: 'claude-md' }: append to the project CLAUDE.md (or ~/.claude/CLAUDE.md when the finding is global) wrapped in markers:

<!-- codeburn:begin <finding-id> -->
<rule text>
<!-- codeburn:end <finding-id> -->

• Idempotent: if a block with the same finding-id already exists, replace its contents instead of appending a duplicate. Undo removes the whole block (the journal backup already guarantees byte-exact restore; the marker discipline is for humans and idempotency).

bash-output-cap (kind: shell-config)

• Append the suggested export to the user's shell rc (detect from $SHELL: .zshrc or .bashrc) inside the same marker scheme using # comments.

Everything else returns null (not appliable) and is shown greyed out with "manual" in the picker.

Step 3: UX

codeburn optimize --apply            # plan + interactive confirm
codeburn optimize --apply --yes      # apply all appliable without prompting
codeburn optimize --apply --dry-run  # print the plan and exit 0, change nothing
codeburn optimize --apply --only mcp-low-coverage,unused-skills

Flow: run the normal optimize scan, list appliable findings numbered with estimated savings and the exact files each plan touches, prompt Apply all / pick numbers / quit (plain readline; do not pull in a new TUI dependency for this). After applying, print a summary with each action's journal id and the line Undo anytime: codeburn act undo <id>.

--apply combined with --json is an error (exit 2) to keep the JSON contract stable.

Tests (tests/optimize-apply.test.ts)

Fixture home dir with crafted ~/.claude.json, project .mcp.json, skills tree, CLAUDE.md:

• mcp remove deletes exactly the named server from the right file, other keys untouched, undo restores byte-identical
• project-scope moves the entry global -> project .mcp.json, creates the file when missing
• server defined in an unparseable file is reported and skipped, remaining plans still apply
• archive moves and undo round-trip, collision gets -2 suffix
• CLAUDE.md marker block: fresh append, second apply replaces in place (no duplicate), undo removes the block
• --dry-run leaves the fixture tree byte-identical (hash the tree before/after)
• every finding in a full optimize run has a stable non-empty id (regression guard for step 1)

Acceptance

• On a machine with real findings, codeburn optimize --apply --dry-run prints a correct plan and --apply + codeburn act undo --last round-trips cleanly.
• optimize --json now carries finding ids; no other JSON shape changes.
• Zero new dependencies, all existing tests green.

Part of epic #602.