Skip to content

Composio powers 1000+ toolkits, tool search, context management, authentication, and a sandboxed workbench to help you build AI agents that turn intent into action.

docs: add Dependabot remediation plans

jkomynoe313ee4161fa
diff --git a/advisor-plans/001-refresh-actions-stale.md b/advisor-plans/001-refresh-actions-stale.md
new file mode 100644
index 000000000..dd609be9f
--- /dev/null
+++ b/advisor-plans/001-refresh-actions-stale.md
@@ -0,0 +1,128 @@
+# Plan 001: Refresh `actions/stale` without changing repository policy
+
+> **Executor instructions**: Follow each step and run every verification. Stop
+> on any condition listed below; do not improvise. When complete, update this
+> plan's row in `advisor-plans/README.md`.
+>
+> **Drift check (run first)**:
+> `git diff --stat 03291bfaa..HEAD -- .github/workflows/stale.yml .github/scripts/check-action-repos.sh`
+> If either file changed, compare the live workflow and checker with the state
+> below before proceeding.
+
+## Status
+
+- **Priority**: P2
+- **Effort**: S
+- **Risk**: LOW
+- **Depends on**: none
+- **Category**: maintenance
+- **Planned at**: commit `03291bfaa`, 2026-08-03
+- **Source PR**: [GitHub #4001](https://github.com/ComposioHQ/composio/pull/4001) ([Glen review](https://app.tryglen.com/review/ComposioHQ/composio/4001))
+
+## Why this matters
+
+The v11 action retains the configured inputs and Node 24 runtime while updating
+its bundled dependencies. The current Dependabot head is green but predates the
+current `next`, and scheduled workflow behavior is not exercised by PR CI.
+
+## Current state
+
+- `.github/workflows/stale.yml:18` pins `actions/stale` v10.4.0 at
+ `1e223db275d687790206a7acac4d1a11bd6fe629`.
+- The proposed v11.0.0 revision is
+ `4391f3da665fdf50b6810c1a66712fb9ba21aa93`.
+- Lines 19-46 configure issue and PR thresholds, labels, messages,
+ `ascending`, `remove-stale-when-updated`, and `debug-only: false`.
+- `.github/scripts/check-action-repos.sh` verifies that pinned action commits
+ and action metadata exist.
+- Manual dispatch is unsafe as a test because the workflow has write
+ permissions and `debug-only` is false.
+
+## Commands
+
+| Purpose | Command | Expected result |
+| ------------- | -------------------------------------------- | --------------------------------------- |
+| Workflow lint | `actionlint .github/workflows/stale.yml` | exit 0, no findings |
+| Action pins | `bash .github/scripts/check-action-repos.sh` | exit 0; v11 revision and metadata found |
+| Patch hygiene | `git diff --check` | exit 0 |
+
+## Scope
+
+**In scope**:
+
+- `.github/workflows/stale.yml`
+
+**Out of scope**:
+
+- Thresholds, labels, messages, permissions, cron schedule, and `debug-only`.
+- Any manual or API dispatch of the workflow.
+- Other workflow action pins.
+
+## Git workflow
+
+- Branch from the latest `origin/next`.
+- Commit: `chore(deps): bump actions/stale to v11`
+- Do not push, dispatch a workflow, or open a PR unless instructed.
+
+## Steps
+
+### Step 1: Refresh and validate the immutable revision
+
+Confirm that v11.0.0 still resolves to the proposed commit and that its
+`action.yml` exposes every input used in `.github/workflows/stale.yml`:
+
+```bash
+gh api repos/actions/stale/git/ref/tags/v11.0.0 --jq '.object.sha'
+stale_metadata="$(mktemp)"
+gh api 'repos/actions/stale/contents/action.yml?ref=v11.0.0' \
+ --jq '.content' | base64 --decode > "$stale_metadata"
+rg -n 'repo-token|days-before-stale|days-before-close|ascending|stale-issue-label|exempt-issue-labels|stale-issue-message|close-issue-message|stale-pr-label|exempt-pr-labels|stale-pr-message|close-pr-message|remove-stale-when-updated|debug-only' "$stale_metadata"
+```
+
+The first command must print `4391f3da665fdf50b6810c1a66712fb9ba21aa93`.
+The `rg` command must print all 14 configured input names from the temporary
+metadata file. If the tag moved or an input disappeared, stop.
+
+### Step 2: Change only the pinned action line
+
+Replace line 18 with:
+
+```yaml
+uses: actions/stale@4391f3da665fdf50b6810c1a66712fb9ba21aa93 # v11.0.0
+```
+
+**Verify**:
+
+```bash
+git diff -- .github/workflows/stale.yml
+```
+
+Expected: one deletion and one addition on the `uses:` line only.
+
+### Step 3: Run static checks and obtain a fresh CI result
+
+Run all commands in the Commands table. After a PR is opened by an authorized
+operator, require all repository checks to finish on the rebased head. Observe
+the next scheduled run after merge; do not manually dispatch it.
+
+## Test plan
+
+No new test is warranted for a one-line immutable action pin. Static metadata
+validation, workflow lint, refreshed PR checks, and the next scheduled run are
+the regression gates.
+
+## Done criteria
+
+- [ ] The workflow pins the verified v11.0.0 commit.
+- [ ] No stale-policy input or permission changed.
+- [ ] `actionlint` and `check-action-repos.sh` exit 0.
+- [ ] Refreshed PR checks pass on the current `next` base.
+- [ ] The workflow was not manually dispatched.
+- [ ] Only `.github/workflows/stale.yml` changed.
+
+## STOP conditions
+
+- The v11.0.0 tag no longer resolves to the recorded commit.
+- Any configured input is absent or renamed in v11 metadata.
+- The action no longer uses a runner supported by GitHub-hosted runners.
+- Validation requires changing stale policy or permissions.
diff --git a/advisor-plans/002-refresh-development-tooling.md b/advisor-plans/002-refresh-development-tooling.md
new file mode 100644
index 000000000..659bbe4d9
--- /dev/null
+++ b/advisor-plans/002-refresh-development-tooling.md
@@ -0,0 +1,141 @@
+# Plan 002: Refresh the seven-package development dependency group
+
+> **Executor instructions**: Follow each step and run every verification. Stop
+> on any condition listed below; do not improvise. When complete, update this
+> plan's row in `advisor-plans/README.md`.
+>
+> **Drift check (run first)**:
+> `git diff --stat 03291bfaa..HEAD -- package.json pnpm-lock.yaml ts/examples/mastra/package.json ts/examples/mcp/package.json ts/examples/tool-router/package.json ts/examples/vercel/package.json ts/packages/cli-keyring/package.json ts/packages/cli-local-tools/package.json ts/packages/cli/package.json ts/packages/experimental/package.json ts/packages/json-schema-to-zod/package.json`
+> If an in-scope manifest changed, recalculate the target versions from the live
+> manifest and registry before proceeding.
+
+## Status
+
+- **Priority**: P2
+- **Effort**: S
+- **Risk**: LOW
+- **Depends on**: none
+- **Category**: maintenance
+- **Planned at**: commit `03291bfaa`, 2026-08-03
+- **Source PR**: [GitHub #4003](https://github.com/ComposioHQ/composio/pull/4003) ([Glen review](https://app.tryglen.com/review/ComposioHQ/composio/4003))
+
+## Why this matters
+
+The Dependabot head already passed audit, build, typecheck, unit, examples,
+CLI scratch, Node, Deno, and Cloudflare checks. Rebuilding it from current
+`next` clears seven development updates without publishing runtime dependency
+changes; the remaining risk is stale lockfile resolution.
+
+## Current state
+
+The source PR proposed these updates:
+
+| Package | Proposed change | Role |
+| --------------------------------- | ---------------- | ---------------------------------------------------------- |
+| `@types/node` | 26.1.1 to 26.1.2 | root and seven additional first-party TypeScript manifests |
+| `oxlint` | 1.75.0 to 1.76.0 | root linter |
+| `oxlint-plugin-eslint` | 1.75.0 to 1.76.0 | must match `oxlint` |
+| `turbo` | 2.10.6 to 2.10.7 | workspace task runner |
+| `tar` | 7.5.21 to 7.5.22 | CLI ACP archive extraction build path |
+| `@earendil-works/pi-coding-agent` | 0.82.0 to 0.82.1 | experimental adapter dev dependency |
+| `eve` | 0.27.3 to 0.27.8 | experimental adapter dev dependency |
+
+Targets may have advanced. Select the newest version that satisfies the same
+major line and the root `minimumReleaseAge: 4320` policy. Keep `oxlint` and
+`oxlint-plugin-eslint` on the identical version. Do not edit vendor manifests.
+
+## Commands
+
+| Purpose | Command | Expected result |
+| --------------------- | ------------------------------------------------------------------------------------------------------------------- | ---------------------------------- |
+| Resolve | `mise exec -- pnpm install --lockfile-only` | exit 0; lockfile generated by pnpm |
+| Reproducibility | `CI=true mise exec -- pnpm install --frozen-lockfile` | exit 0, no lockfile drift |
+| Lint | `mise exec -- pnpm lint:packages` | exit 0 |
+| Build | `mise exec -- pnpm build:packages` | exit 0 |
+| Types | `mise exec -- pnpm typecheck` | exit 0 |
+| Experimental adapters | `mise exec -- pnpm --filter @composio/experimental test && mise exec -- pnpm --filter @composio/experimental build` | all pass |
+| CLI binary | `mise exec -- pnpm build:binary` | exit 0 |
+| Tests | `mise exec -- pnpm test` | all pass |
+| CLI E2E | `mise exec -- pnpm test:e2e:cli` | all pass |
+
+## Scope
+
+**In scope**:
+
+- `package.json`
+- `pnpm-lock.yaml` through pnpm only
+- `ts/examples/{mastra,mcp,tool-router,vercel}/package.json`
+- `ts/packages/{cli-keyring,cli-local-tools,cli,experimental,json-schema-to-zod}/package.json`
+
+**Out of scope**:
+
+- `docs/package.json` and `ts/vendor/**`.
+- Production dependencies or source changes.
+- Any weakening of `minimumReleaseAge`.
+- A Changeset; all changes are development-only and CLI packages are excluded.
+
+## Git workflow
+
+- Branch from the latest `origin/next`.
+- Commit: `chore(deps-dev): refresh development tooling`
+- Do not push or open a PR unless instructed.
+
+## Steps
+
+### Step 1: Re-resolve eligible versions
+
+Query registry version and publication data for all seven packages. Preserve
+their current major lines, choose policy-eligible targets, and update only the
+listed manifests. Change every first-party occurrence touched by #4003, and
+keep the two oxlint package versions equal.
+
+**Verify**:
+
+```bash
+rg -n '"(@types/node|oxlint|oxlint-plugin-eslint|turbo|tar|@earendil-works/pi-coding-agent|eve)"' \
+ package.json ts/examples/{mastra,mcp,tool-router,vercel}/package.json \
+ ts/packages/{cli-keyring,cli-local-tools,cli,experimental,json-schema-to-zod}/package.json
+```
+
+Expected: every intended occurrence shows the chosen version; Pi and Eve peer
+ranges remain unchanged.
+
+### Step 2: Generate and prove the lockfile
+
+Run the Resolve and Reproducibility commands. Then run:
+
+```bash
+git diff --name-only
+```
+
+Expected: only the in-scope manifests and `pnpm-lock.yaml` are listed.
+
+### Step 3: Exercise tooling and adapters
+
+Run the remaining Commands table in order. The CLI binary and CLI E2E gates
+exercise the `tar` path; the experimental tests exercise the Pi and Eve
+adapters. Obtain a fresh complete CI run after an authorized operator opens or
+refreshes the PR.
+
+## Test plan
+
+No new tests are required because behavior is unchanged and dedicated Pi, Eve,
+CLI binary, and runtime suites already cover the affected contracts.
+
+## Done criteria
+
+- [ ] All seven package families use current policy-eligible versions.
+- [ ] `oxlint` and its ESLint plugin have identical versions.
+- [ ] Frozen install, lint, build, typecheck, tests, binary build, and CLI E2E pass.
+- [ ] `pnpm-lock.yaml` was generated by pnpm from current `next`.
+- [ ] No Changeset or source file was added.
+- [ ] Only in-scope files changed.
+
+## STOP conditions
+
+- A target version is younger than the 72-hour policy or requires weakening it.
+- Registry metadata requires a major update not named by this plan.
+- Oxlint and its plugin cannot resolve at the same version.
+- A test failure requires production source changes rather than a compatible
+ dependency selection.
+- The lockfile update changes unrelated direct dependencies.
diff --git a/advisor-plans/003-align-effect-cli-family.md b/advisor-plans/003-align-effect-cli-family.md
new file mode 100644
index 000000000..03fd4b159
--- /dev/null
+++ b/advisor-plans/003-align-effect-cli-family.md
@@ -0,0 +1,158 @@
+# Plan 003: Align the CLI on one maintained Effect release family
+
+> **Executor instructions**: Follow each step and run every verification. Stop
+> on any condition listed below; do not improvise. When complete, update this
+> plan's row in `advisor-plans/README.md`.
+>
+> **Drift check (run first)**:
+> `git diff --stat 03291bfaa..HEAD -- pnpm-workspace.yaml pnpm-lock.yaml ts/packages/cli/package.json .github/dependabot.yml`
+> If any in-scope dependency changed, re-check the current peer matrix before
+> editing.
+
+## Status
+
+- **Priority**: P1
+- **Effort**: M
+- **Risk**: MED
+- **Depends on**: none
+- **Category**: migration
+- **Planned at**: commit `03291bfaa`, 2026-08-03
+- **Source PR**: [GitHub #4004](https://github.com/ComposioHQ/composio/pull/4004) ([Glen review](https://app.tryglen.com/review/ComposioHQ/composio/4004))
+
+## Why this matters
+
+`@effect/[email protected]` is an older package generation despite its higher
+semver major. It removes the `@effect/printer-ansi/Ansi` export used by the
+current CLI, causing 56 test suites and the CLI binary E2E build to fail. The
+safe update is the maintained, coordinated 0.51 family.
+
+## Current state
+
+- `pnpm-workspace.yaml` catalogs `@effect/cli ^0.76.0`,
+ `@effect/platform ^0.97.0`, and `effect ^3.22.0`.
+- `ts/packages/cli/package.json` uses `@effect/printer` and
+ `@effect/printer-ansi ^0.50.0` plus related Effect packages.
+- Live npm peer metadata coordinates these targets:
+
+| Package | Target line |
+| ----------------------------------------- | ----------- |
+| `@effect/cli` | `^0.77.0` |
+| `@effect/printer`, `@effect/printer-ansi` | `^0.51.0` |
+| `@effect/platform` | `^0.97.1` |
+| `effect` | `^3.22.1` |
+| `@effect/cluster` | `^0.60.2` |
+| `@effect/experimental` | `^0.61.1` |
+| `@effect/platform-bun` | `^0.91.2` |
+| `@effect/platform-node-shared` | `^0.61.1` |
+| `@effect/rpc` | `^0.76.2` |
+| `@effect/sql` | `^0.52.1` |
+| `@effect/workflow` | `^0.19.1` |
+
+Keep `@effect/typeclass` on the compatible 0.41 line. Re-query peer metadata at
+execution time and use the newest coordinated, 72-hour-eligible patch set.
+
+## Commands
+
+| Purpose | Command | Expected result |
+| ---------------- | ---------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------- |
+| Resolve | `mise exec -- pnpm install --lockfile-only` | exit 0, no peer errors |
+| Reproducibility | `CI=true mise exec -- pnpm install --frozen-lockfile` | exit 0 |
+| Dependency graph | `mise exec -- pnpm --filter @composio/cli list @effect/cli @effect/printer @effect/printer-ansi @effect/platform effect --depth 1` | only the coordinated current family; no printer 1.0.0 |
+| Types | `mise exec -- pnpm --filter @composio/cli typecheck` | exit 0 |
+| Unit tests | `mise exec -- pnpm --filter @composio/cli test` | all pass |
+| Package build | `mise exec -- pnpm --filter @composio/cli build` | exit 0 |
+| Binary build | `mise exec -- pnpm build:binary` | exit 0 |
+| CLI E2E | `mise exec -- pnpm test:e2e:cli` | all pass |
+
+## Scope
+
+**In scope**:
+
+- `pnpm-workspace.yaml`
+- `ts/packages/cli/package.json`
+- `.github/dependabot.yml`
+- `pnpm-lock.yaml` through pnpm only
+
+**Out of scope**:
+
+- `ts/vendor/**` and CLI source files.
+- Patching missing package exports or adding overrides for obsolete packages.
+- Changesets; `@composio/cli` is excluded from Changesets.
+
+## Git workflow
+
+- Branch from the latest `origin/next`.
+- Commit: `chore(cli): align Effect dependencies`
+- Do not push or open a PR unless instructed.
+
+## Steps
+
+### Step 1: Confirm the maintained peer matrix
+
+Query npm metadata for the listed targets. Confirm that the selected
+`@effect/cli` version peers on the selected printer, platform, and Effect lines.
+Confirm that `@effect/[email protected]` still peers on legacy `@effect/io` or
+`@effect/data`. If upstream republished 1.x for the current generation, stop and
+re-plan instead of adding the ignore below.
+
+### Step 2: Update the coordinated dependency set
+
+Update catalog entries in `pnpm-workspace.yaml` and direct CLI development
+dependencies in `ts/packages/cli/package.json` as one unit. Do not update only
+the printer package. Generate the lockfile, prove frozen install, and inspect
+the dependency graph with the Commands table.
+
+**Verify**:
+
+```bash
+rg -n '"@effect/(cluster|experimental|printer|printer-ansi|rpc|sql|typeclass|workflow)"' \
+ ts/packages/cli/package.json
+rg -n "'@effect/(cli|platform|platform-bun|platform-node-shared)'|effect:" pnpm-workspace.yaml
+```
+
+Expected: every entry belongs to the coordinated family and printer 1.0.0 is
+absent from `pnpm-lock.yaml`.
+
+### Step 3: Prevent the invalid major proposal from recurring
+
+Add this ignore to the npm update block in `.github/dependabot.yml`:
+
+```yaml
+ignore:
+ - dependency-name: '@effect/printer-ansi'
+ versions: ['1.0.0']
+```
+
+**Verify**:
+
+```bash
+yq -e '.updates[] | select(.package-ecosystem == "npm") | .ignore[] | select(.dependency-name == "@effect/printer-ansi")' .github/dependabot.yml
+```
+
+Expected: exit 0 and exactly one matching ignore entry.
+
+### Step 4: Exercise the CLI contract
+
+Run typecheck, unit tests, package build, binary build, and CLI E2E from the
+Commands table. A source change is not part of this dependency maintenance
+plan; stop if one appears necessary.
+
+## Test plan
+
+No new test is required. The regression is an import and binary-build failure
+already covered by CLI unit tests and scratch E2E.
+
+## Done criteria
+
+- [ ] The CLI resolves one current, peer-compatible Effect family.
+- [ ] Neither printer 1.0.0 nor legacy `@effect/io`/`@effect/data` enters the lockfile.
+- [ ] Dependabot ignores only `@effect/[email protected]`.
+- [ ] Frozen install, CLI typecheck, tests, builds, and E2E pass.
+- [ ] No vendor, source, or Changeset file changed.
+
+## STOP conditions
+
+- Current peer metadata differs from the documented coordinated matrix.
+- Resolution introduces printer 1.0.0, `@effect/io`, or `@effect/data`.
+- pnpm reports unresolved or incorrectly satisfied Effect peer dependencies.
+- Passing the gates requires CLI source edits, an override, or a vendor change.
diff --git a/advisor-plans/004-rebuild-npm-production-rollup.md b/advisor-plans/004-rebuild-npm-production-rollup.md
new file mode 100644
index 000000000..1e4f268b0
--- /dev/null
+++ b/advisor-plans/004-rebuild-npm-production-rollup.md
@@ -0,0 +1,174 @@
+# Plan 004: Rebuild the npm production rollup on a coherent AI SDK train
+
+> **Executor instructions**: Follow each step and run every verification. Stop
+> on any condition listed below; do not improvise. When complete, update this
+> plan's row in `advisor-plans/README.md`.
+>
+> **Drift check (run first)**:
+> `git diff --stat 03291bfaa..HEAD -- pnpm-workspace.yaml pnpm-lock.yaml ts/e2e-tests/runtimes/cloudflare/cf-workers-tool-router-ai/package.json ts/e2e-tests/runtimes/node/claude-agent-sdk/package.json ts/examples/anthropic/package.json ts/examples/langchain/package.json ts/examples/mcp/package.json ts/examples/openai/package.json ts/examples/tool-router/package.json ts/examples/vercel/package.json ts/packages/experimental/package.json ts/packages/providers/anthropic/package.json ts/packages/providers/claude-agent-sdk/package.json ts/packages/providers/openai-agents/package.json`
+> Re-audit any changed manifest before selecting versions.
+
+## Status
+
+- **Priority**: P1
+- **Effort**: M
+- **Risk**: MED
+- **Depends on**: none
+- **Category**: maintenance
+- **Planned at**: commit `03291bfaa`, 2026-08-03
+- **Source PRs**: [GitHub #4002](https://github.com/ComposioHQ/composio/pull/4002) ([Glen review](https://app.tryglen.com/review/ComposioHQ/composio/4002)); [GitHub #4006](https://github.com/ComposioHQ/composio/pull/4006) ([Glen review](https://app.tryglen.com/review/ComposioHQ/composio/4006))
+
+## Why this matters
+
+The grouped production update fails example typechecking because
+`@ai-sdk/mcp` and `@ai-sdk/openai` resolve newer internal provider types than
+the catalog's `ai` release. Updating unrelated packages in the same failing
+commit hides their status. The replacement keeps one coherent AI SDK 7 train,
+retains deliberate AI SDK 6 compatibility coverage, and validates the other
+package families in small commits.
+
+## Current state
+
+PR #4002 proposed these 12 updates; re-query versions because the targets are
+already stale:
+
+| Package | PR target |
+| -------------------------------- | ------------ |
+| `@anthropic-ai/claude-agent-sdk` | 0.3.220 |
+| `@ai-sdk/mcp` | 2.0.18 |
+| `@anthropic-ai/sdk` | 0.115.0 |
+| `langchain` | 1.5.4 |
+| `@openai/agents` | 0.14.0 |
+| `@langchain/anthropic` | 1.5.2 |
+| `typebox` | 1.3.8 |
+| `@ai-sdk/openai` | 4.0.22 |
+| `@cloudflare/workers-types` | 5.20260728.1 |
+| `@mastra/core` | 1.53.0 |
+| `@modelcontextprotocol/sdk` | 1.30.0 |
+| `hono` | 4.12.32 |
+
+At the failed head, `@ai-sdk/[email protected]` and
+`@ai-sdk/[email protected]` use provider 4.0.4/provider-utils 5.0.14 while
+`[email protected]` uses 4.0.3/5.0.12. The resulting error is at
+`ts/examples/vercel/src/tool-router-ai.ts:73` when assigning MCP tools to
+`ToolSet`.
+
+The six explicit `ai: ^6.0.219` entries in the Mastra Zod 3/4 runtimes,
+Anthropic/Mastra examples, and Anthropic/Mastra provider dev dependencies are
+intentional compatibility coverage. Keep them. The normal catalog is the AI
+SDK 7 lane and must advance with `@ai-sdk/mcp` and `@ai-sdk/openai`.
+
+`typebox` is a runtime dependency of published `@composio/experimental`, so it
+requires a patch Changeset.
+
+## Commands
+
+| Purpose | Command | Expected result |
+| --------------- | -------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------- |
+| Resolve | `mise exec -- pnpm install --lockfile-only` | exit 0 |
+| Reproducibility | `CI=true mise exec -- pnpm install --frozen-lockfile` | exit 0 |
+| AI graph | `mise exec -- pnpm --filter vercel-example list ai @ai-sdk/mcp @ai-sdk/openai @ai-sdk/provider @ai-sdk/provider-utils --depth 4` | AI SDK 7 consumer graph uses compatible provider internals |
+| Peer policy | `mise exec -- pnpm check:peer-deps` | exit 0 |
+| Changesets | `mise exec -- pnpm validate:changesets` | exit 0 |
+| Focused example | `mise exec -- pnpm --filter vercel-example typecheck` | exit 0; no `ToolSet` error |
+| AI v6 fixture | `mise exec -- pnpm --filter @e2e-tests/node-vercel-ai-sdk-v6 test:e2e:node` | all pass |
+| AI v7 fixture | `mise exec -- pnpm --filter @e2e-tests/node-vercel-ai-sdk-v7 test:e2e:node` | all pass |
+| All examples | `mise exec -- pnpm typecheck:examples && mise exec -- pnpm test:examples` | all pass |
+| Packages | `mise exec -- pnpm build:packages && mise exec -- pnpm typecheck` | all pass |
+| Tests | `mise exec -- pnpm test` | all pass |
+| Runtime E2E | `mise exec -- pnpm test:e2e:node && mise exec -- pnpm test:e2e:cloudflare && mise exec -- pnpm test:e2e:cli` | all pass in credentialed CI |
+| Audit | `mise exec -- pnpm audit --prod --audit-level=high` | no high/critical production advisory introduced |
+
+## Scope
+
+**In scope**:
+
+- `pnpm-workspace.yaml` and `pnpm-lock.yaml` through pnpm.
+- The 12 manifest files listed in the drift check.
+- One new `.changeset/*.md` for the `typebox` runtime update.
+
+**Out of scope**:
+
+- The six explicit AI SDK 6 compatibility pins named above.
+- `ts/e2e-tests/runtimes/node/vercel-ai-sdk-v6/**` and its AI SDK 7 sibling.
+- OpenAI Node v7; that public migration is Plan 005.
+- Transitive overrides, source casts, or edits that suppress incompatible
+ `ToolSet` types.
+
+## Git workflow
+
+- Branch from the latest `origin/next`.
+- Use conventional commits in this order:
+ 1. `chore(deps): align AI SDK dependencies`
+ 2. `chore(deps): refresh agent framework dependencies`
+ 3. `chore(deps): refresh runtime support dependencies`
+- Do not push or open a PR unless instructed.
+
+## Steps
+
+### Step 1: Select one eligible AI SDK 7 train
+
+Query versions, publication dates, and dependency metadata for `ai`,
+`@ai-sdk/mcp`, and `@ai-sdk/openai`. Choose releases older than 72 hours whose
+exact `@ai-sdk/provider` and `@ai-sdk/provider-utils` requirements are mutually
+compatible. Add `@ai-sdk/mcp` to the root catalog and replace its repeated
+first-party manifest ranges with `catalog:`. Advance the existing `ai` and
+`@ai-sdk/openai` catalog entries together.
+
+Generate the lockfile and run the AI graph, focused example, all-example
+typecheck, Vercel v6 fixture, and Vercel v7 fixture before continuing.
+
+**Verify compatibility pins remain**:
+
+```bash
+rg -n '"ai": "\^6\.0\.219"' \
+ ts/e2e-tests/runtimes/node/mastra-tool-router-zod-v{3,4}/package.json \
+ ts/examples/{anthropic,mastra}/package.json \
+ ts/packages/providers/{anthropic,mastra}/package.json
+```
+
+Expected: six matches.
+
+### Step 2: Update framework integrations in small commits
+
+Update the Anthropic, Claude Agent SDK, LangChain, OpenAI Agents, Mastra, and
+MCP SDK packages to current policy-eligible versions within the majors named by
+#4002. After each framework family, regenerate the lockfile and run its package
+tests/build plus affected example typechecks. Do not batch a failing family
+with the next one.
+
+### Step 3: Update TypeBox and support packages
+
+Update `typebox`, Cloudflare workers types, and Hono. Add a patch Changeset for
+`@composio/experimental` describing the TypeBox runtime dependency refresh.
+Generate the lockfile and run experimental package tests/build plus Cloudflare
+E2E.
+
+### Step 4: Run the complete regression gates
+
+Run every command in the Commands table on a clean frozen install. In CI,
+require the Claude Agent, Cloudflare Tool Router, Mastra Zod 3/4, Node, Deno,
+and CLI scratch jobs that GitHub selects for the changed paths.
+
+## Test plan
+
+Do not add tests if compatible dependency selection makes existing example and
+runtime gates pass. Add a regression test only if a real first-party public
+contract must change; stop first and report that scope expansion.
+
+## Done criteria
+
+- [ ] All 12 named production packages use current, policy-eligible targets.
+- [ ] The AI SDK 7 catalog resolves compatible provider internals.
+- [ ] All six intentional AI SDK 6 manifest pins remain.
+- [ ] A valid patch Changeset covers `@composio/experimental`.
+- [ ] Frozen install, peer checks, examples, builds, types, tests, E2E, and audit pass.
+- [ ] No source cast, transitive override, or unrelated dependency update was added.
+
+## STOP conditions
+
+- No 72-hour-eligible AI SDK 7 release train has compatible internal packages.
+- Fixing the `ToolSet` error requires a cast or first-party public API change.
+- A framework update requires a new major not named by the source PR.
+- The Mastra AI SDK 6 lane changes or its Zod 3/4 runtime checks fail.
+- The TypeBox runtime update cannot pass without changing experimental behavior.
diff --git a/advisor-plans/005-migrate-openai-v7.md b/advisor-plans/005-migrate-openai-v7.md
new file mode 100644
index 000000000..209898829
--- /dev/null
+++ b/advisor-plans/005-migrate-openai-v7.md
@@ -0,0 +1,157 @@
+# Plan 005: Migrate published OpenAI consumers to v7 while preserving v6 support
+
+> **Executor instructions**: Follow each step and run every verification. Stop
+> on any condition listed below; do not improvise. When complete, update this
+> plan's row in `advisor-plans/README.md`.
+>
+> **Drift check (run first)**:
+> `git diff --stat 03291bfaa..HEAD -- pnpm-workspace.yaml pnpm-lock.yaml ts/packages/core/package.json ts/packages/slim/package.json ts/packages/providers/openai/package.json ts/packages/core/src/provider/OpenAIProvider.ts ts/packages/providers/openai/src/OpenAIResponsesProvider.ts ts/e2e-tests/runtimes/node/openai-zod4-compat`
+> Re-audit OpenAI-owned exported types and package metadata if any path changed.
+
+## Status
+
+- **Priority**: P1
+- **Effort**: M
+- **Risk**: HIGH
+- **Depends on**: none
+- **Category**: migration
+- **Planned at**: commit `03291bfaa`, 2026-08-03
+- **Source PR**: [GitHub #4005](https://github.com/ComposioHQ/composio/pull/4005) ([Glen review](https://app.tryglen.com/review/ComposioHQ/composio/4005))
+
+## Why this matters
+
+The current Dependabot diff changes a runtime dependency of published
+`@composio/core` and `@composio/slim`, and changes the published peer major of
+`@composio/openai`. OpenAI v7's documented runtime break is Node 22+, which
+matches this repository's support floor, but exported OpenAI-owned types and
+downstream installation still need dual-major proof. The existing PR ran only
+build and generic checks.
+
+## Current state
+
+- `pnpm-workspace.yaml:52` catalogs `openai: ^6.49.0`.
+- `ts/packages/core/package.json` and `ts/packages/slim/package.json` carry the
+ catalog entry as a runtime dependency.
+- `ts/packages/providers/openai/package.json` uses `catalog:` as its peer range;
+ changing the catalog directly would drop declared v6 compatibility.
+- Core and the dedicated provider export types derived from OpenAI namespaces
+ in `OpenAIProvider.ts` and `OpenAIResponsesProvider.ts`.
+- `ts/e2e-tests/runtimes/node/openai-zod4-compat` pins OpenAI 6, but currently
+ installs only `@composio/core` with `--legacy-peer-deps`; it does not prove the
+ dedicated provider's peer contract and has stale `openai@5` assertion labels.
+- The Node E2E matrix is 22.22.3, 24.17.0, and 25.9.0, all compatible with
+ OpenAI v7's Node 22 floor.
+
+Use the newest OpenAI 7 release permitted by `minimumReleaseAge: 4320`, not the
+stale 7.0.0 target from the bot body.
+
+## Commands
+
+| Purpose | Command | Expected result |
+| ----------------- | ------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------- |
+| Resolve | `mise exec -- pnpm install --lockfile-only` | exit 0 |
+| Reproducibility | `CI=true mise exec -- pnpm install --frozen-lockfile` | exit 0 |
+| Peer policy | `mise exec -- pnpm check:peer-deps` | exit 0 |
+| Changesets | `mise exec -- pnpm validate:changesets` | exit 0 |
+| Core | `mise exec -- pnpm --filter @composio/core typecheck && mise exec -- pnpm --filter @composio/core test` | all pass |
+| OpenAI provider | `mise exec -- pnpm --filter @composio/openai build && mise exec -- pnpm --filter @composio/openai test` | all pass |
+| Examples | `mise exec -- pnpm typecheck:examples && mise exec -- pnpm test:examples` | all pass |
+| Packages | `mise exec -- pnpm build:packages && mise exec -- pnpm typecheck` | all pass |
+| Unit tests | `mise exec -- pnpm test` | all pass |
+| OpenAI v6 fixture | `mise exec -- pnpm --filter @e2e-tests/node-openai-zod4-compat test:e2e:node` | install, typecheck, and runtime assertions pass on Node matrix |
+| OpenAI v7 fixture | `mise exec -- pnpm --filter @e2e-tests/node-openai-v7-zod4-compat test:e2e:node` | install, typecheck, and runtime assertions pass on Node matrix |
+| Cloudflare | `mise exec -- pnpm test:e2e:cloudflare` | all pass in credentialed CI |
+| Audit | `mise exec -- pnpm audit --prod --audit-level=high` | no high/critical production advisory introduced |
+
+## Scope
+
+**In scope**:
+
+- `pnpm-workspace.yaml`, `pnpm-lock.yaml` through pnpm.
+- `ts/packages/providers/openai/package.json` for an explicit dual-major peer.
+- `ts/e2e-tests/runtimes/node/openai-zod4-compat/**` to make the v6 proof real
+ and correct stale labels.
+- A new `ts/e2e-tests/runtimes/node/openai-v7-zod4-compat/**` sibling fixture.
+- One or more `.changeset/*.md` files covering `@composio/core`,
+ `@composio/slim`, and `@composio/openai`.
+
+**Out of scope**:
+
+- Dropping OpenAI 6 support without an explicit product decision.
+- Changing the repository's Node support floor.
+- Live API calls, real keys, generated SDK surfaces, or OpenAI API behavior.
+- Source edits made only to silence incompatible exported types.
+
+## Git workflow
+
+- Branch from the latest `origin/next`.
+- Commit 1: `test(openai): cover v6 and v7 consumers`
+- Commit 2: `chore(deps): migrate OpenAI runtime to v7`
+- Do not push or open a PR unless instructed.
+
+## Steps
+
+### Step 1: Establish consumer fixtures before the catalog bump
+
+Strengthen `openai-zod4-compat` so it packs and installs the published
+`@composio/openai` package against explicit OpenAI 6 and runs a TypeScript
+assignability check for exported tool/response types. Retain its offline
+`wrapTool` runtime assertions. Correct all `openai@5` comments and markers to
+OpenAI 6.
+
+Create `openai-v7-zod4-compat` from the same pattern with explicit OpenAI 7.
+Both fixtures must install packed packages as consumers receive them; do not
+use workspace source imports or `skipLibCheck` to hide peer/type conflicts.
+
+Run both fixtures before changing the catalog. The v7 fixture may fail at this
+point because the provider peer excludes v7; that is the expected baseline.
+
+### Step 2: Update runtime and peer ranges
+
+Update the catalog to the selected OpenAI 7 release. Replace the dedicated
+provider's `catalog:` peer with an explicit `^6.49.0 || ^7.0.0` range so the
+catalog update does not silently drop v6 consumers. Generate the lockfile and
+run core and provider typecheck/build/tests.
+
+Do not alter exported type aliases unless the existing shapes fail under one of
+the supported majors. If they do, stop and report the exact incompatible types.
+
+### Step 3: Add release metadata
+
+Add a patch Changeset for fixed-group packages `@composio/core` and
+`@composio/slim`, describing the runtime dependency refresh. Add a minor
+Changeset for `@composio/openai`, following the prior `@composio/vercel` dual
+major support precedent, because the published peer contract gains v7 support.
+
+### Step 4: Run the full regression matrix
+
+Run every command in the Commands table from a frozen install. Require both
+consumer fixtures on all supported Node majors and the existing Cloudflare
+tests, because Cloudflare examples also consume the OpenAI catalog entry.
+
+## Test plan
+
+- OpenAI 6: packed provider install, exported-type compile check, client
+ construction with a fake key, `wrapTool`, and no network request.
+- OpenAI 7: the same checks against the v7 package.
+- Keep Zod 4 in both fixtures to retain the prior peer-resolution regression.
+- Existing core/provider unit suites remain authoritative for wrapping behavior.
+
+## Done criteria
+
+- [ ] Core and slim resolve a policy-eligible OpenAI 7 runtime.
+- [ ] `@composio/openai` declares and proves OpenAI 6 and 7 peers.
+- [ ] Both packed consumer fixtures pass on Node 22, 24, and 25 without real keys.
+- [ ] Patch changesets cover core/slim; a minor changeset covers the provider.
+- [ ] Frozen install, peer, examples, builds, types, tests, E2E, and audit pass.
+- [ ] The Node support floor and generated code are unchanged.
+
+## STOP conditions
+
+- The selected OpenAI v7 release requires a runtime newer than the supported
+ Node matrix.
+- Exported provider types cannot compile against both OpenAI 6 and 7.
+- Supporting both majors requires weakening types, `skipLibCheck`, or duplicate
+ public APIs.
+- Existing OpenAI 6 behavior fails after the runtime dependency update.
+- The change needs a product decision to drop v6 or change a public type.
diff --git a/advisor-plans/006-migrate-autogen-provider-to-ag2-v1.md b/advisor-plans/006-migrate-autogen-provider-to-ag2-v1.md
new file mode 100644
index 000000000..dfb67a143
--- /dev/null
+++ b/advisor-plans/006-migrate-autogen-provider-to-ag2-v1.md
@@ -0,0 +1,235 @@
+# Plan 006: Migrate the AutoGen provider to AG2 v1
+
+> **Executor instructions**: Follow each step and run every verification. Stop
+> on any condition listed below; do not improvise. When complete, update this
+> plan's row in `advisor-plans/README.md`.
+>
+> **Drift check (run first)**:
+> `git diff --stat 03291bfaa..HEAD -- python/providers/autogen python/tests/test_provider.py python/tests/test_type_inference_autogen.py docs/content/docs/providers/autogen.mdx .github/workflows/py.test.yml`
+> Re-audit AG2's public imports, tool protocol, and installation metadata if any
+> listed path or AG2's latest 1.x release changed.
+
+## Status
+
+- **Priority**: P1
+- **Effort**: M
+- **Risk**: HIGH
+- **Depends on**: none
+- **Category**: migration
+- **Planned at**: commit `03291bfaa`, 2026-08-03
+- **Source PR**: [GitHub #4008](https://github.com/ComposioHQ/composio/pull/4008) ([Glen review](https://app.tryglen.com/review/ComposioHQ/composio/4008))
+
+## Why this matters
+
+AG2 is the maintained forward package. Version 1 replaces the Classic
+`autogen` namespace and two-agent registration API with the `ag2` namespace,
+`Agent`, and native tools. Dependabot correctly identifies the major dependency
+upgrade, but its metadata-only change installs AG2 1.0.1 while the provider
+still imports `autogen`; Python 3.10 therefore fails immediately with
+`ModuleNotFoundError`, and the later matrix jobs are cancelled.
+
+The safe fix is an AG2 v1 provider migration, not a switch to the legacy
+`autogen` distribution. Use the [AG2 package](https://pypi.org/project/ag2/)
+and its [v1 tool API](https://docs.ag2.ai/docs/user-guide/tools/tools/) as the
+upstream contracts.
+
+## Current state
+
+- `python/providers/autogen/pyproject.toml:16` and `setup.py:25` declare
+ `ag2>=0.14,<1.0` plus direct Classic-era `flaml` and `autogen_core`
+ dependencies.
+- `provider.py:6-8` imports `autogen`, `ConversableAgent`, and
+ `autogen_core.tools.FunctionTool`.
+- `provider.py:45-65` registers each tool with separate caller and executor
+ agents through `autogen.agentchat.register_function`.
+- `wrap_tool` already creates a deterministic function name, a schema-derived
+ Python signature, and an execution closure that restores reserved keywords.
+ Preserve those Composio-specific behaviors while changing the framework
+ wrapper.
+- `python/tests/test_provider.py` reaches into the Classic tool's private
+ `_func` attribute, and `test_type_inference_autogen.py` imports the Microsoft
+ `autogen_core` type. Neither assertion represents AG2 v1.
+- The provider README and public docs install `ag2[openai]`, but their examples
+ still import Classic `AssistantAgent`, `LLMConfig`, and `UserProxyAgent`.
+- `python/providers/autogen` is not a root uv workspace member, so this package
+ change must not alter root `uv.lock`.
+
+Target AG2 1.0.1 or newer within major version 1. Keep the published
+`composio-autogen` package, `composio_autogen` module, and `AutogenProvider`
+class names in this migration so existing Composio imports remain valid.
+
+## Commands
+
+| Purpose | Command | Expected result |
+| ------------------ | ------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------- |
+| Focused tests | `cd python && nox -s tst -- tests/test_autogen_provider.py tests/test_provider.py` | all focused tests pass |
+| Python checks | `cd python && nox -s chk` | exit 0 |
+| Type inference | `cd python && nox -s type_inference` | exit 0 |
+| Provider build | `mise exec -- uv build python/providers/autogen` | wheel and sdist build successfully |
+| Stale-import guard | `rg -n -e 'from autogen' -e 'import autogen' -e autogen_core -e flaml python/providers/autogen python/tests/test_type_inference_autogen.py` | no matches |
+| Docs links | `cd docs && bun run lint:links` | exit 0 |
+| Docs build | `cd docs && bun run build` | exit 0 |
+| Lockfile guard | `git diff --exit-code -- uv.lock` | exit 0; root lock unchanged |
+
+## Scope
+
+**In scope**:
+
+- `python/providers/autogen/composio_autogen/provider.py`
+- `python/providers/autogen/pyproject.toml`
+- `python/providers/autogen/setup.py`
+- `python/providers/autogen/README.md`
+- `python/tests/test_autogen_provider.py` (new)
+- `python/tests/test_provider.py`
+- `python/tests/test_type_inference_autogen.py`
+- `docs/content/docs/providers/autogen.mdx`
+
+**Out of scope**:
+
+- Renaming the published package, module, provider class, or docs route.
+- Supporting AG2 Classic and AG2 v1 in one provider release.
+- Root `uv.lock` or root uv workspace membership.
+- Unrelated Python provider migrations or historical changelogs.
+- Network model calls, Composio API requests, or real API keys in tests.
+
+## Git workflow
+
+- Branch from the latest `origin/next`.
+- Commit 1: `test(autogen): characterize AG2 v1 tool contracts`
+- Commit 2: `feat(autogen): migrate provider to AG2 v1`
+- Commit 3: `docs(autogen): update provider guide for AG2 v1`
+- Do not push or open a PR unless instructed.
+
+## Steps
+
+### Step 1: Replace private Classic assertions with AG2 v1 contracts
+
+Create `python/tests/test_autogen_provider.py` using local tool fixtures and a
+fake `execute_tool` callback. Update the two existing test files so they import
+and assert AG2 v1 types. Cover these contracts offline:
+
+1. `wrap_tool` returns `ag2.tools.final.FunctionTool`, and
+ `wrap_tools` returns `list[FunctionTool]` statically and at runtime.
+2. `wrapped.schema.function` contains the deterministic name, Composio tool
+ description, parameter types, required fields, and defaults expected from
+ the generated signature. Assert the public schema instead of private
+ framework attributes.
+3. `schema_config={"skip_defaults": True}` makes defaulted parameters required
+ in the AG2 schema; the default configuration preserves their defaults.
+4. A reserved Python keyword is exposed under its safe generated parameter and
+ restored before calling the fake executor.
+5. Invoke the wrapped tool through AG2's async `ToolCallEvent` and `Context`
+ protocol. Assert the original Composio slug, normalized arguments, and
+ structured result without constructing a real model client.
+6. A long tool slug produces the same valid name on repeated wraps and stays at
+ or below 64 characters.
+7. A model-free or stub-configured `ag2.Agent` accepts wrapped tools through
+ `Agent(..., tools=tools)`, and the migrated `register_tools(agent, tools)`
+ helper delegates to `Agent.add_tool` for post-construction attachment.
+
+The initial test commit may fail because the source still uses Classic imports;
+record that expected failure before migrating the implementation.
+
+### Step 2: Migrate wrapping and registration to AG2 v1
+
+In `provider.py`:
+
+1. Replace Classic imports with `ag2.Agent`, `ag2.tool`, and the public AG2 v1
+ `FunctionTool` type.
+2. Keep the existing name hashing, schema-to-signature conversion, reserved-key
+ restoration, argument normalization, and Composio execution closure.
+3. Wrap the generated callable with `ag2.tool`, passing the processed name and
+ description. Let AG2 derive its function schema from the generated signature
+ so `skip_defaults` continues to control required/defaulted parameters.
+4. Replace `register_tools(caller, executor, tools)` with
+ `register_tools(agent, tools)` implemented through `Agent.add_tool`. Retain
+ the helper for dynamic attachment while making `Agent(..., tools=tools)` the
+ primary documented API. Do not preserve the obsolete executor parameter or
+ depend on Classic runtime objects.
+5. Keep tool execution synchronous inside the generated function and rely on
+ AG2's default `sync_to_thread=True` behavior when the async tool protocol
+ invokes it.
+
+The old caller/executor signature is an intentional breaking change caused by
+the upstream major version. Document its exact replacement in Step 4 rather
+than adding a compatibility dependency on the Classic package.
+
+### Step 3: Align package metadata with the migrated implementation
+
+Change both metadata sources to the same `ag2>=1.0.1,<2.0` requirement. Remove
+the direct `flaml==2.6.0` and `autogen_core>=0.7.5` requirements after the source
+and tests no longer import them. Do not add the `autogen` distribution.
+
+Build the provider, inspect wheel and sdist metadata, and confirm both artifacts
+declare the same AG2 bound with no direct `autogen`, `autogen-core`, or FLAML
+requirement. Leave package versioning to the existing Python release workflow.
+
+### Step 4: Rewrite the quickstart for AG2 v1
+
+Update the provider README and MDX page to:
+
+1. Continue installing `composio`, `composio-autogen`, and `ag2[openai]`.
+2. Import `Agent` from `ag2` and the current OpenAI model config from
+ `ag2.config`; do not use Classic `AssistantAgent`, `UserProxyAgent`, or
+ `LLMConfig`.
+3. Pass `session.tools()` to `Agent(..., tools=tools)` and run the current async
+ AG2 interaction method shown in upstream v1 documentation.
+4. Add a short migration note mapping the old caller/executor registration call
+ to constructor tools or `Agent.add_tool`.
+5. Keep the note about deterministic 64-character tool names.
+
+Validate the example against installed AG2 1.0.1 without sending a model request.
+Follow `docs/AGENTS.md`; do not edit generated docs data.
+
+### Step 5: Prove clean installs on every supported Python version
+
+After building, run this import check independently for Python 3.10, 3.11, and
+3.12:
+
+```bash
+mise exec -- uv run --isolated --no-project --python 3.10 \
+ --with-editable ./python \
+ --with-editable ./python/providers/autogen \
+ python -c "import ag2; import composio_autogen; from ag2 import Agent"
+```
+
+Repeat with `--python 3.11` and `--python 3.12`. On one supported version, add
+`--with 'ag2[openai]>=1.0.1,<2.0'` and import the exact model config used by the
+quickstart. Expected: every command exits 0, resolves AG2 1.x, and does not need
+the Classic `autogen` namespace.
+
+Then run every command in the Commands table and require the repository's
+Python 3.10, 3.11, and 3.12 CI jobs to complete.
+
+## Test plan
+
+The focused tests exercise AG2's public schema, execution event, agent tool
+registration, and Composio's existing name/default/argument behavior without
+network access. Static type inference, fresh-version imports, artifact metadata,
+Python checks, and docs builds cover packaging and documentation regressions.
+
+## Done criteria
+
+- [ ] Both metadata sources require `ag2>=1.0.1,<2.0` and do not add `autogen`.
+- [ ] Provider runtime and type tests use only AG2 v1 public imports.
+- [ ] Tool schema, defaults, reserved arguments, long names, execution, and
+ agent attachment pass focused offline tests.
+- [ ] README and public docs use AG2 v1 imports and explain the caller/executor
+ migration.
+- [ ] Fresh imports pass on Python 3.10, 3.11, and 3.12.
+- [ ] Wheel and sdist metadata agree and contain no stale direct Classic
+ dependencies.
+- [ ] Root `uv.lock`, package/module/class names, and unrelated providers remain
+ unchanged.
+
+## STOP conditions
+
+- AG2's public 1.x `tool`, `FunctionTool`, `Agent`, or tool-event contract differs
+ from the interfaces confirmed during the drift check.
+- Preserving Composio's `skip_defaults`, reserved-key restoration, or execution
+ result contract requires AG2 private attributes.
+- The current quickstart cannot be constructed without a network call or secret;
+ replace only the validation technique, not the documented public API.
+- Fresh resolution selects AG2 2.x, installs the Classic `autogen` distribution,
+ or requires a direct dependency not declared by AG2.
+- Root `uv.lock` or another provider changes during resolution.
diff --git a/advisor-plans/007-upgrade-tomli-after-crewai.md b/advisor-plans/007-upgrade-tomli-after-crewai.md
new file mode 100644
index 000000000..adfa8aa1b
--- /dev/null
+++ b/advisor-plans/007-upgrade-tomli-after-crewai.md
@@ -0,0 +1,133 @@
+# Plan 007: Upgrade Tomli after CrewAI permits a newer line
+
+> **Executor instructions**: This plan is currently blocked. Run the upstream
+> trigger check first. If CrewAI still requires Tomli 2.0.x, do not edit the
+> repository; leave the plan `BLOCKED` in `advisor-plans/README.md`. Once the
+> trigger clears, follow every step and verification below.
+>
+> **Drift check (run first)**:
+> `git diff --stat 03291bfaa..HEAD -- python/pyproject.toml python/noxfile.py python/providers/crewai/pyproject.toml python/providers/crewai/setup.py python/scripts/bump.py uv.lock`
+> Re-audit the complete resolved graph if any in-scope dependency changed.
+
+## Status
+
+- **Priority**: P3
+- **Effort**: S after upstream unblock
+- **Risk**: LOW
+- **Depends on**: a CrewAI release that no longer requires `tomli~=2.0.2`
+- **Category**: maintenance
+- **Planned at**: commit `03291bfaa`, 2026-08-03
+- **Source PR**: [GitHub #4007](https://github.com/ComposioHQ/composio/pull/4007) ([Glen review](https://app.tryglen.com/review/ComposioHQ/composio/4007))
+- **Current status**: BLOCKED — CrewAI 1.15.10 still constrains Tomli to 2.0.x
+
+## Why this matters
+
+PR #4007 widens Composio's direct upper bound but cannot update the installed
+package: the unified uv workspace includes CrewAI, whose current metadata still
+forces Tomli 2.0.x. Merging that diff would claim an update while `uv.lock`
+remains on 2.0.2. Waiting for a compatible CrewAI release keeps dependency
+metadata truthful and avoids an override.
+
+## Current state
+
+- `python/pyproject.toml:36-37` documents CrewAI's constraint and requires
+ `tomli>=2.0.2,<2.1`.
+- `python/providers/crewai/pyproject.toml` and `setup.py` support CrewAI
+ `>=1.15.7,<2.0.0`; `python/noxfile.py` tests exactly 1.15.7.
+- `python/scripts/bump.py` imports Tomli to parse release metadata.
+- `uv.lock` resolves Tomli 2.0.2.
+- Live PyPI metadata for CrewAI 1.15.10 still includes `tomli~=2.0.2`.
+- There is no known high/critical advisory against Tomli 2.0.2 in this usage;
+ this is maintenance, not an emergency override.
+
+## Commands
+
+| Purpose | Command | Expected result |
+| ---------------- | ---------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------- |
+| Upstream trigger | `curl -fsSL https://pypi.org/pypi/crewai/json | jq -r '.info.version, (.info.requires_dist[]? | select(test("tomli"; "i")))'` | newest version and a Tomli requirement that allows the intended newer line |
+| Generate lock | `mise exec -- uv lock` | exit 0; lock generated by uv |
+| Lock check | `mise exec -- uv lock --check` | exit 0 |
+| Resolution | `mise exec -- uv tree --locked --package tomli --invert` | selected Tomli is newer than 2.0.2 and all dependents are satisfied |
+| Parser smoke | `cd python && uv run python -c 'import pathlib, tomli; tomli.loads(pathlib.Path("pyproject.toml").read_text())'` | exit 0 |
+| Python checks | `cd python && make chk` | exit 0 |
+| Tests | `cd python && make tst` | all pass |
+| Sanity | `cd python && make snt` | exit 0 |
+
+## Scope
+
+**In scope after the trigger clears**:
+
+- `python/pyproject.toml`
+- `python/noxfile.py`
+- `python/providers/crewai/pyproject.toml`
+- `python/providers/crewai/setup.py`
+- `uv.lock` through uv only
+
+**Out of scope**:
+
+- Dependency overrides or patched CrewAI metadata.
+- Removing CrewAI from the uv workspace.
+- Rewriting the bump script to avoid Tomli.
+- Any change while upstream still requires Tomli 2.0.x.
+
+## Git workflow
+
+- Branch from the latest `origin/next` only after the trigger clears.
+- Commit: `chore(deps-dev): update Tomli and CrewAI baseline`
+- Do not push or open a PR unless instructed.
+
+## Steps
+
+### Step 1: Prove the upstream block is gone
+
+Run the Upstream trigger command and inspect the newest stable CrewAI release,
+including any split CrewAI Core/CLI distributions it requires. Proceed only if
+the full published requirement graph accepts the intended Tomli version without
+an override. Record the qualifying CrewAI version in the PR description.
+
+If the output still contains `tomli~=2.0.2`, stop with no repository changes.
+
+### Step 2: Raise the tested CrewAI baseline consistently
+
+Update the CrewAI lower bound in both provider metadata files and the exact
+test version in `python/noxfile.py` to the first stable release that relaxes
+Tomli. Keep `pyproject.toml` and `setup.py` identical. Do not raise the provider
+baseline if the older CrewAI range can coexist with the selected Tomli and the
+owner wants to retain it; in that case stop for an explicit compatibility
+decision rather than silently narrowing support.
+
+### Step 3: Select and resolve the newer Tomli line
+
+Update the direct Tomli range and its comment in `python/pyproject.toml`. Set the
+minimum to the selected tested release so the lockfile must actually advance;
+choose an upper bound supported by both Tomli and CrewAI metadata. Run `uv lock`
+from the repository root and never hand-edit `uv.lock`.
+
+### Step 4: Prove parsing and the Python graph
+
+Run the remaining Commands table checks. In CI, require the Python 3.10, 3.11,
+and 3.12 matrix because Tomli is especially relevant on Python versions before
+the standard-library `tomllib` module.
+
+## Test plan
+
+No new test is needed. The package is used by the release bump parser; a direct
+parse smoke check plus Python lint, unit, sanity, and version-matrix jobs cover
+the changed dependency contract.
+
+## Done criteria
+
+- [ ] Upstream CrewAI metadata accepts the selected Tomli line without overrides.
+- [ ] Both CrewAI provider metadata sources and the nox baseline agree.
+- [ ] `uv.lock` resolves a Tomli version newer than 2.0.2.
+- [ ] `uv lock --check`, parser smoke, checks, tests, and sanity pass.
+- [ ] No workspace member, override, or bump-script behavior changed.
+
+## STOP conditions
+
+- Any current CrewAI distribution still constrains Tomli to 2.0.x.
+- Updating Tomli requires a resolver override, removing a workspace member, or
+ dropping a supported Python version.
+- The tested CrewAI baseline cannot be raised without a product compatibility
+ decision.
+- `uv.lock` still resolves Tomli 2.0.2 after the manifest changes.
diff --git a/advisor-plans/README.md b/advisor-plans/README.md
new file mode 100644
index 000000000..0593dc415
--- /dev/null
+++ b/advisor-plans/README.md
@@ -0,0 +1,77 @@
+# Dependabot remediation plans
+
+Generated from live GitHub state and `origin/next` on 2026-08-03. The audit
+covered all eight open Dependabot pull requests and all 25 dependency updates
+named by them. Execute each plan from a fresh branch based on the then-current
+`origin/next`; never reuse a stale Dependabot lockfile.
+
+## Audit result
+
+All eight pull requests are Git-mergeable and review-blocked. Mergeability is
+not approval: four heads have substantive failed checks, two green heads need
+only a fresh base and rerun, and two green or partially green heads do not prove
+the proposed compatibility change.
+
+| PR | Packages | Live result | Disposition |
+| --------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------- | --------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------- |
+| [#4001](https://github.com/ComposioHQ/composio/pull/4001) ([Glen](https://app.tryglen.com/review/ComposioHQ/composio/4001)) | `actions/stale` 10.4.0 to 11.0.0 | 10 checks, no failures; scheduled action itself did not run | Refresh and approve through Plan 001 |
+| [#4002](https://github.com/ComposioHQ/composio/pull/4002) ([Glen](https://app.tryglen.com/review/ComposioHQ/composio/4002)) | 12 npm production updates | Examples fail on incompatible AI SDK internal types | Supersede through Plan 004 |
+| [#4003](https://github.com/ComposioHQ/composio/pull/4003) ([Glen](https://app.tryglen.com/review/ComposioHQ/composio/4003)) | 7 npm development updates | 22 checks, no failures | Refresh and approve through Plan 002 |
+| [#4004](https://github.com/ComposioHQ/composio/pull/4004) ([Glen](https://app.tryglen.com/review/ComposioHQ/composio/4004)) | `@effect/printer-ansi` 0.50.0 to 1.0.0 | Unit and CLI E2E fail; 1.0.0 belongs to an obsolete Effect generation | Close and supersede through Plan 003 |
+| [#4005](https://github.com/ComposioHQ/composio/pull/4005) ([Glen](https://app.tryglen.com/review/ComposioHQ/composio/4005)) | `openai` 6.49.0 to 7.x on current `next` | Build is green, but tests, typecheck, examples, peer checks, and changesets did not run | Supersede through Plan 005 |
+| [#4006](https://github.com/ComposioHQ/composio/pull/4006) ([Glen](https://app.tryglen.com/review/ComposioHQ/composio/4006)) | six explicit AI SDK 6 compatibility pins to AI SDK 7 | 11 jobs fail at frozen install because the lockfile was omitted | Close; Plan 004 advances the AI SDK 7 catalog while retaining this intentional v6 lane |
+| [#4007](https://github.com/ComposioHQ/composio/pull/4007) ([Glen](https://app.tryglen.com/review/ComposioHQ/composio/4007)) | `tomli` 2.0.x to `<2.5` | Green, but CrewAI still forces the lock to 2.0.2 | Close; Plan 007 is blocked on CrewAI relaxing its requirement |
+| [#4008](https://github.com/ComposioHQ/composio/pull/4008) ([Glen](https://app.tryglen.com/review/ComposioHQ/composio/4008)) | `ag2` 0.x to 1.x | Python 3.10 import fails; 3.11 and 3.12 were cancelled | Supersede with an AG2 v1 provider migration through Plan 006 |
+
+## Package inventory
+
+Plan 004 covers the 12 production-group packages: `@anthropic-ai/claude-agent-sdk`,
+`@ai-sdk/mcp`, `@anthropic-ai/sdk`, `langchain`, `@openai/agents`,
+`@langchain/anthropic`, `typebox`, `@ai-sdk/openai`,
+`@cloudflare/workers-types`, `@mastra/core`, `@modelcontextprotocol/sdk`, and
+`hono`. It also advances the existing AI SDK 7 catalog dependency in lockstep
+with the two `@ai-sdk/*` packages.
+
+Plan 002 covers the seven development-group packages: `@types/node`, `oxlint`,
+`oxlint-plugin-eslint`, `turbo`, `tar`, `@earendil-works/pi-coding-agent`, and
+`eve`. Plans 001, 003, 005, 006, and 007 cover `actions/stale`, the coordinated
+Effect family, `openai`, AG2 v1, and Tomli respectively.
+
+## Execution order and status
+
+| Plan | Title | Priority | Effort | Depends on | Status |
+| ---- | ------------------------------------------------------------------------------------- | -------- | ------ | ------------------------ | --------------------------------------------------- |
+| 001 | [Refresh `actions/stale` safely](./001-refresh-actions-stale.md) | P2 | S | none | TODO |
+| 002 | [Refresh development tooling](./002-refresh-development-tooling.md) | P2 | S | none | TODO |
+| 003 | [Align the Effect CLI family](./003-align-effect-cli-family.md) | P1 | M | none | TODO |
+| 004 | [Rebuild the npm production rollup](./004-rebuild-npm-production-rollup.md) | P1 | M | none | TODO |
+| 005 | [Migrate published OpenAI consumers to v7](./005-migrate-openai-v7.md) | P1 | M | none | TODO |
+| 006 | [Migrate the AutoGen provider to AG2 v1](./006-migrate-autogen-provider-to-ag2-v1.md) | P1 | M | none | TODO |
+| 007 | [Upgrade Tomli after CrewAI permits it](./007-upgrade-tomli-after-crewai.md) | P3 | S | upstream CrewAI metadata | BLOCKED — CrewAI 1.15.10 still requires Tomli 2.0.x |
+
+Status values: `TODO`, `IN PROGRESS`, `DONE`, `BLOCKED`, or `REJECTED`.
+
+## Landing strategy
+
+- Plans 001 and 002 may share one low-risk rollup PR, with one conventional
+ commit per plan.
+- Plans 003, 004, 005, and 006 should use separate PRs because each has a
+ distinct compatibility contract and rollback boundary.
+- Every TypeScript plan must generate `pnpm-lock.yaml` through pnpm and prove a
+ frozen install. Never copy the lockfile from an old Dependabot head.
+- Do not run the stale workflow manually: its production dispatch can relabel
+ or close issues and pull requests.
+
+## Findings considered and rejected
+
+- Merging #4004 and adding an import patch was rejected. The proposed 1.0.0
+ package is from the wrong Effect generation; the maintained coordinated line
+ is 0.51.x.
+- Merging #4006 after merely regenerating the lockfile was rejected. The six
+ explicit AI SDK 6 pins provide Mastra and Zod compatibility coverage; the
+ normal catalog already exercises AI SDK 7.
+- Merging #4007 was rejected. Widening the direct Tomli range does not change
+ the resolved version while CrewAI requires `tomli~=2.0.2`.
+- Replacing AG2 with the Classic `autogen` distribution was rejected. AG2 1.x is
+ the maintained forward line; the dependency update needs a provider/API
+ migration because it no longer provides the old `autogen` import path.