fix: mark enable_toolset and disable_toolset as non-destructive

[imbenrabi]

Requested by Ben · project thread

Before: enable_toolset and disable_toolset were advertised to MCP clients with destructiveHint: true, so clients could treat turning a toolset on or off as a data-destroying action and prompt for confirmation.

After: both advertise destructiveHint: false, idempotentHint: true. In the MCP spec destructiveHint means the tool may delete or overwrite data. These tools only change which tools are exposed, one reverses the other, and repeating either has no further effect on state.

How: two annotation values in src/meta/registerMetaTools.ts, a new annotations test block in tests/metaTools.test.ts (fails on the old values), and the matching invariant line in src/meta/AGENTS.md. Nothing else changes, so it merges cleanly with the sibling PR that changes disable_toolset's removal behaviour. No CHANGELOG entry: 0.7.0 is already released and the release PR adds the next entry. Unchanged by design: openWorldHint still defaults to true per spec for these tools; that is outside this bug.

Base: main. No stacking, any merge order works.

🤖 Generated with Claude Code

https://claude.ai/code/session_01JZj3jC148r1VwBYogZpWKL

[imbenrabi]

TS review: pipeline-ts-reviewer (feature-pipeline)

Verdict: APPROVE. No CRITICAL or HIGH findings. Reviewed at head 2381fd7+PR commit (before the docs wording fix below).

Reviewed: the 3-file diff against main (src/meta/registerMetaTools.ts, src/meta/AGENTS.md, tests/metaTools.test.ts) with surrounding context (DynamicToolManager, tests/helpers/fakes.ts). Static analysis only, against typescript-standards and karpathy-guidelines.

#	Sev	Location	Finding
1	LOW	tests/metaTools.test.ts (new "discovery meta-tools as read-only" test)	Pins annotations this PR doesn't change; a cheap regression guard, marginally outside "only the two flipped annotations".
2	LOW	src/meta/AGENTS.md:23	"calling either again is a no-op" is true for state but not for the response (a repeat disable_toolset returns success: false). Suggested "no further effect on state".

{
  "recommendation": "APPROVE",
  "summary": "Surgical, correct annotation fix for enable_toolset/disable_toolset with matching doc update and strict regression tests; only minor wording and scope notes.",
  "findings": [
    { "severity": "LOW", "file": "tests/metaTools.test.ts", "issue": "Discovery read-only annotation test covers behaviour this PR does not change; acceptable as a regression guard." },
    { "severity": "LOW", "file": "src/meta/AGENTS.md", "line": 23, "issue": "'no-op' holds for state only; consider 'no further effect on state'." }
  ]
}

---

Generated by Claude Code

[imbenrabi]

Fixes for the TS review, and re-review

Mapping to the review comment above:

#	Finding	Outcome
1	LOW: discovery read-only test is slightly outside the two flipped annotations	Left unchanged. It guards the invariant the updated AGENTS.md line states (discovery tools stay readOnlyHint: true), costs nine lines, and sits in the same new annotations block.
2	LOW: "calling either again is a no-op" is true for state, not for the response	Fixed in f1becfd. src/meta/AGENTS.md now says "repeating either call has no further effect on state".

Re-review of the head (f1becfd): APPROVE, no findings. The reviewer checked the new wording against DynamicToolManager (a repeat enable leaves activeToolsets unchanged; a repeat disable returns success: false and changes nothing) and accepted the kept test as a deliberate regression guard.

{
  "recommendation": "APPROVE",
  "summary": "The annotation change is surgical and correct, the AGENTS.md wording now accurately describes state-level idempotence, and the tests cover the stated invariant.",
  "findings": []
}

---

Generated by Claude Code