chore(release): 0.8.0 by imbenrabi · Pull Request #34 · code-rabi/toolception

imbenrabi · 2026-10-10T18:06:51Z

Requested by Ben · project thread

Before: main is at 0.7.0, and five finished changes (#29, #30, #31, #32, #33) are open as separate PRs with no release entry.

After: version 0.8.0 with one changelog entry covering all five, plus one small addition: PermissionAwareFastifyTransport now rejects a client whose startup toolsets failed to load, the follow-up #33 and #30 asked for. Docs match the new behaviour. Nothing is tagged or published.

Merge order (read first)

Base is main, but none of #29-#33 has merged yet, so this draft branch contains all five merged in (one merge conflict, in CHANGELOG.md, resolved by writing the 0.8.0 entry). Merge them first, in any order; this PR's diff then shrinks to the release commits and merges cleanly on top. Do not rebase or force-push the PR branches.

PR	Change	State
#29	`enable_toolset`/`disable_toolset` advertise `destructiveHint: false`	open
#30	permission-aware transport waits for the orchestrator before `connect()`	open
#31	`disable_toolset` removes the toolset's tools	open
#32	opt-in `http.allowedOrigins` Origin check on `/mcp`	open
#33	failed startup toolset registration fails `initialize` and STATIC startup (Breaking, still a draft)	open

This release ships #33 as written (Strict semantics).

What the release adds beyond the five PRs

package.json / package-lock.json 0.7.0 to 0.8.0, and CHANGELOG.md [0.8.0] with Breaking, Added, Changed and Fixed sections and migration notes.
PermissionAwareFastifyTransport: isReady() false gives 500 / JSON-RPC -32603 and drops the cached bundle, as FastifyTransport does. Two new e2e tests fail without it.
Docs: README (startup failure behaviour), root AGENTS.md invariant 4 (loader errors stay silent, registration failures no longer do) and src/permissions/AGENTS.md. allowedOrigins, destructive hints and disable-removes-tools were already documented by their PRs.

Verification

npm run typecheck, npm run test:run (393 tests) and npm run build pass.
npm publish --dry-run: toolception@0.8.0, 64 files, 107 kB; only dist/, package.json, README.md, CHANGELOG.md and LICENSE; no src/, tests or env files.
npm run verify:toolception in the FMP server (main, 2.7.0) against the packed tarball: 19/21. Both failures are expected and are server work that must merge first:
1. ALL_TOOLS does not start: the calendar and earnings toolsets both list the earnings-transcript module, so getLatestEarningsTranscripts collides (src/constants/toolSets.ts). The server needs a catalog fix, being done in a separate project thread.
2. The dynamic smoke test should disable toolset (state only, tools remain) asserts the old disable behaviour. Fixed by server #113. (The same overlap breaks the legacy/all-tools smoke tests and the installation checks.)

Server prerequisites before the FMP server takes 0.8.0

Merge server #113.
Merge the catalog fix that removes the duplicate earnings-transcript module from one of calendar / earnings.
After 0.8.0 is on npm: npm run verify:toolception -- --source 0.8.0 must be green, then the dependency bump PR, with the docs that say disable is "state tracking only" (docs/USAGE.md, src/endpoints/server-card.ts) changed together with it.

Release steps (Ben)

After the five PRs and this one are merged: tag v0.8.0 on main HEAD (release.yml publishes to npm and creates the GitHub Release). The command will be given in the project thread.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Jke6dnq5CNkrDFWbJSHJWr

Both meta-tools were annotated destructiveHint: true, but neither deletes anything: they change which tools are exposed and can be reversed by the other. Set destructiveHint: false, keep idempotentHint: true, assert the annotations in tests and update the meta intent node. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JZj3jC148r1VwBYogZpWKL

…lient PermissionAwareFastifyTransport connected the per-client server before the bundle's orchestrator finished preloading, so tools registered after connect and the client got 'Method not found' on tools/list. Await isReady() first, as FastifyTransport already does. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Cv2mT1Y95UFmTFB8mZZhcL

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JZj3jC148r1VwBYogZpWKL

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AKosQp8PUoZijAkZ7RWHj4

Disabling a toolset only updated state, so clients could still call its tools. DynamicToolManager now keeps the RegisteredTool handles returned by registerTool and calls remove() on disable, frees the names in ToolRegistry so the toolset can be enabled again, and rolls back a partially failed enable the same way. Intent layer, README and the meta-tool description no longer claim the SDK can't remove tools. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GKeDY2JCSBC6vG56EEKziV

enableToolsets() reports failures in its result instead of throwing, so ServerOrchestrator.isReady() was true even when a startup toolset failed to enable, and FastifyTransport ignored the result anyway. A client could connect with only part of its tools. initializeToolsets() now turns a failed result into initError, and the transport answers initialize with a JSON-RPC -32603 error (HTTP 500) and drops the bundle so the next initialize retries. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UJfs8sWB5EK8dXXjDU7Mv3

- Require the request Origin header to be a bare origin - Type allowedOrigins as readonly string[] - Export OriginGuard type instead of ReturnType<typeof ...> - Cover PermissionAwareFastifyTransport invalid-entry construction Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AKosQp8PUoZijAkZ7RWHj4

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GKeDY2JCSBC6vG56EEKziV

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AKosQp8PUoZijAkZ7RWHj4

Reviewer findings: exposure-policy refusals and duplicate startup toolset names were made fatal by the previous commit. Only failures that happen while registering a toolset's tools (carried as a code on the enable result) now fail readiness. Changelog files the behaviour change as Breaking. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UJfs8sWB5EK8dXXjDU7Mv3

…g entry Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UJfs8sWB5EK8dXXjDU7Mv3

…to claude/project-thread-4ocdp6

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jke6dnq5CNkrDFWbJSHJWr

…ets failed to load Mirrors FastifyTransport: a not-ready bundle gets a 500 / -32603 on initialize and is dropped from the cache. Docs updated for 0.8.0. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jke6dnq5CNkrDFWbJSHJWr

imbenrabi · 2026-10-10T18:09:08Z

TS review: `pipeline-ts-reviewer` (feature-pipeline) — round 1

Verdict: REQUEST_CHANGES. The permission-aware isReady() guard is correct as code and mirrors FastifyTransport, and the version bump is consistent. But the 0.8.0 changelog and docs claim a behaviour change for createPermissionBasedMcpServer that the code cannot deliver.

Reviewed: the full diff against main (head a2857cf) plus ServerOrchestrator.ts, DynamicToolManager.ts, createPermissionBasedMcpServer.ts, permissions.utils.ts, PermissionAwareFastifyTransport.ts, createMcpServer.ts, the src/index.ts exports, ModuleResolver, ClientResourceCache, package.json and package-lock.json. Static analysis only.

#	Sev	Location	Finding
1	HIGH	`CHANGELOG.md` [0.8.0] Breaking, first bullet	Says `createPermissionBasedMcpServer` clients with failing startup toolsets now get `-32603` / HTTP 500. False: the permission orchestrator is built with `toolsets: []`, `createPermissionAwareBundle` enables toolsets itself, partial failure still connects (`failedToolsets`), total failure is 403, so `isReady()` never goes false there.
2	MEDIUM	`CHANGELOG.md` [0.8.0] Fixed (#30)	Describes `PermissionAwareFastifyTransport`, which is not exported from `src/index.ts`, racing on background preload. Unreachable via the public factory, which awaits `enableToolsets` before returning.
3	MEDIUM	`src/permissions/AGENTS.md` flow; root `AGENTS.md` invariant 4	Imply the `isReady` / 500 path covers permission clients, and contradict "partial success continues with warning".
4	MEDIUM	`tests/e2e/permissionBased.e2e.test.ts` (new describes)	Build the transport directly with hand-made bundles, so they pass while the public path is unchanged and hide finding 1. Label as transport-level, and pin the real behaviour through `createPermissionBasedMcpServer`.
5	LOW	`README.md` startup note	"activates with no tools" overstates it; other modules in the toolset still register ("fewer tools").

{
  "recommendation": "REQUEST_CHANGES",
  "summary": "Version bump and most changelog/doc claims match the code, but the Breaking entry (and related docs/tests) claim createPermissionBasedMcpServer clients now get a 500/-32603 on failed startup toolsets, which cannot happen because that factory enables toolsets itself, tolerates partial failure, and leaves orchestrator.isReady() always true.",
  "findings": [
    { "severity": "HIGH", "file": "CHANGELOG.md", "line": null, "issue": "Breaking entry claims permission-based clients get -32603/500 on failed startup toolsets; false for the public factory." },
    { "severity": "MEDIUM", "file": "CHANGELOG.md", "line": null, "issue": "Fixed entry for #30 describes an unexported class racing on preload; unreachable via createPermissionBasedMcpServer." },
    { "severity": "MEDIUM", "file": "src/permissions/AGENTS.md", "line": null, "issue": "Permission Flow and root invariant 4 imply the isReady/500 path covers permission clients and contradict partial-success behaviour." },
    { "severity": "MEDIUM", "file": "tests/e2e/permissionBased.e2e.test.ts", "line": null, "issue": "New e2e tests bypass the public factory and mask the changelog inaccuracy." },
    { "severity": "LOW", "file": "README.md", "line": null, "issue": "'activates with no tools' vs 'fewer tools' elsewhere." }
  ]
}

Generated by Claude Code

createPermissionBasedMcpServer enables toolsets in its bundle factory and continues on partial failure, so the isReady() rejection does not apply. Pin that with a public-factory test and label the transport-level tests. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jke6dnq5CNkrDFWbJSHJWr

imbenrabi · 2026-10-10T18:09:52Z

Fixes for the round 1 TS review

All five findings were correct (I confirmed the permission flow in permissions.utils.ts: toolsets are enabled in the bundle factory, partial failure continues, all-failed throws and becomes a 403). Fixed in 0349957.

#	Finding	Fix
1 (HIGH)	Breaking entry claimed permission-based clients get 500 / `-32603`	Entry now covers STATIC `createMcpServer` only, and says `createPermissionBasedMcpServer` is unaffected (partial failure still connects with `failedToolsets`, all failed is still 403).
2 (MEDIUM)	#30 Fixed entry described an unreachable race	Reworded as hardening of the transport for hand-built `createPermissionAwareBundle` bundles; says the public factory's behaviour is unchanged.
3 (MEDIUM)	`src/permissions/AGENTS.md` flow and root invariant 4 implied the guard covers permission clients	Both now say only a hand-built bundle can be not ready, and that `createPermissionBasedMcpServer` continues on partial failure. Root invariant 4 names it as the exception.
4 (MEDIUM)	e2e tests bypassed the public factory	The two transport tests are labelled "transport level". New e2e case through `createPermissionBasedMcpServer` pins the real behaviour: partial failure connects with the good toolset's tools, all failed is refused. 395 tests pass.
5 (LOW)	README said "no tools"	Now "fewer tools", matching the other docs.

The guard itself stays: it is a small mirror of FastifyTransport for the exported-in-spirit custom-bundle path, and the review found the code correct.

Generated by Claude Code

…ening Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jke6dnq5CNkrDFWbJSHJWr

imbenrabi · 2026-10-10T18:11:42Z

TS review: `pipeline-ts-reviewer` — round 2 (head `0349957`)

Verdict: COMMENT (no CRITICAL/HIGH). All five round-1 findings were fixed correctly and truthfully, every [0.8.0] changelog and doc claim checked matches the code, and the fix commit introduced nothing new. Three LOW notes:

#	Sev	Location	Finding	Fix (`1c58bb8`)
1	LOW	`CHANGELOG.md` Fixed (#30)	"hand-built `createPermissionAwareBundle` callback" implies a path consumers have; the class and helper are not exported.	Reworded as internal hardening: the class is not public API and `createPermissionBasedMcpServer` is unchanged.
2	LOW	`permissionBased.e2e.test.ts` "all failed"	Bare `rejects.toThrow()` passes on any error.	Now asserts `/Access denied/` (the 403 body).
3	LOW	`permissionBased.e2e.test.ts` partial failure	Did not assert the failed toolset's tool is absent.	Now asserts `dup.ping` is not listed, which also pins the #31 rollback on the public path.

{"recommendation":"COMMENT","summary":"All five round-1 findings fixed correctly and truthfully; only LOW polish remains.","findings":[{"severity":"LOW","file":"CHANGELOG.md","line":null,"issue":"#30 entry implies a user-reachable path; reword as internal hardening."},{"severity":"LOW","file":"tests/e2e/permissionBased.e2e.test.ts","line":null,"issue":"bare rejects.toThrow() in the all-failed test."},{"severity":"LOW","file":"tests/e2e/permissionBased.e2e.test.ts","line":null,"issue":"partial-failure test does not assert the failed toolset's tool is absent."}]}

All three are addressed in 1c58bb8 (typecheck clean, 395 tests pass). Round 3 follows to confirm APPROVE.

Generated by Claude Code

imbenrabi · 2026-10-10T18:12:33Z

TS review: `pipeline-ts-reviewer` — round 3 (head `1c58bb8`)

Verdict: APPROVE. All three round-2 LOW notes are fixed, the [0.8.0] changelog and touched docs match the code, and nothing new was found at any severity. CI is green on this head.

{"recommendation":"APPROVE","summary":"All three round-2 notes are fixed (CHANGELOG #30 reads as internal hardening, the all-failed test asserts /Access denied/, the partial-failure test asserts dup.ping is absent), the 0.8.0 CHANGELOG and touched docs match the code, and nothing new was introduced.","findings":[]}

Still a draft on purpose: #29-#33 are unmerged and are included in this branch. Merge them first, then mark this ready.

Generated by Claude Code

claude added 16 commits October 10, 2026 16:51

docs: say repeat toggles have no further effect on state

f1becfd

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JZj3jC148r1VwBYogZpWKL

feat(http): validate Origin on /mcp endpoints via http.allowedOrigins

cfa334c

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AKosQp8PUoZijAkZ7RWHj4

review: tighten removal types and tests

0b4a411

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GKeDY2JCSBC6vG56EEKziV

docs(http): fix Origin guard TSDoc and lower-case non-http hosts

d73df0d

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AKosQp8PUoZijAkZ7RWHj4

docs: clarify the registration-failure code and the Breaking changelo…

cbdd1a3

…g entry Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UJfs8sWB5EK8dXXjDU7Mv3

Merge remote-tracking branch 'origin/claude/project-thread-7cs1pg' in…

3e303f7

…to claude/project-thread-4ocdp6

Merge remote-tracking branch 'origin/claude/project-thread-7rn4lq' in…

3ba9cb1

…to claude/project-thread-4ocdp6

Merge remote-tracking branch 'origin/claude/project-thread-i2z3lz' in…

fe5e062

…to claude/project-thread-4ocdp6

chore(release): 0.8.0 changelog and version

a8ed7c1

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jke6dnq5CNkrDFWbJSHJWr

imbenrabi self-assigned this Oct 10, 2026

review: tighten permission-factory assertions, call #30 internal hard…

1c58bb8

…ening Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jke6dnq5CNkrDFWbJSHJWr

Navigation Menu

Repository navigation

chore(release): 0.8.0 - #34

chore(release): 0.8.0#34
imbenrabi wants to merge 18 commits into
mainfrom
claude/project-thread-4ocdp6

imbenrabi commented Oct 10, 2026 •

edited

Loading

Uh oh!

imbenrabi commented Oct 10, 2026

Uh oh!

imbenrabi commented Oct 10, 2026

Uh oh!

imbenrabi commented Oct 10, 2026

Uh oh!

imbenrabi commented Oct 10, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

imbenrabi commented Oct 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Merge order (read first)

What the release adds beyond the five PRs

Verification

Server prerequisites before the FMP server takes 0.8.0

Release steps (Ben)

Uh oh!

imbenrabi commented Oct 10, 2026

TS review: pipeline-ts-reviewer (feature-pipeline) — round 1

Uh oh!

imbenrabi commented Oct 10, 2026

Fixes for the round 1 TS review

Uh oh!

imbenrabi commented Oct 10, 2026

TS review: pipeline-ts-reviewer — round 2 (head 0349957)

Uh oh!

imbenrabi commented Oct 10, 2026

TS review: pipeline-ts-reviewer — round 3 (head 1c58bb8)

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

imbenrabi commented Oct 10, 2026 •

edited

Loading

TS review: `pipeline-ts-reviewer` (feature-pipeline) — round 1

TS review: `pipeline-ts-reviewer` — round 2 (head `0349957`)

TS review: `pipeline-ts-reviewer` — round 3 (head `1c58bb8`)