**Stacked on #3927** (A1 — direct-to-PostHog + install→Apollo identity).
Targets that branch; will auto-retarget to `next` when #3927 merges.
Review only this diff.
## What
Makes every CLI event self-describe its funnel stage, so the adoption
funnel can be queried without string-matching event names — which is how
it silently broke in the first place.
- **`journey_stage`** on every event: `install | setup | login | connect
| execute | other`. Declared as an exported typed const plus a single
exhaustive map, `as const satisfies Record<CliAnalyticsEventName,
CliJourneyStage>` — so **the compiler refuses to build if a future event
has no stage**. No per-event literals; stamped in one shared builder.
- **`cli_channel`** alongside `cli_version`, reusing the existing
`inferSkillReleaseChannel()` / `CLI_RELEASE_CHANNELS`. Makes "did this
fix reach stable users?" answerable — the question that went unanswered
in the 0.2.30 assetless-release incident.
- **Installer origin fix:** `install.sh` ran `"$exe" install` *without*
`COMPOSIO_CLI_INVOCATION_ORIGIN=installer` (the `setup` call on line 410
had it), so a genuine install was indistinguishable from a user running
`composio install` by hand. Now tagged, with a regression test that
reads `install.sh` and asserts it.
## Stage mapping (reviewed against 30d volume)
`install` = `CLI_INSTALL_*` · `setup` = `CLI_SETUP_*` + `CLI_PLUGIN_*` ·
`login` = `CLI_LOGIN_*` · `connect` = `CLI_LINK_*` · `execute` =
`CLI_EXECUTE_*` + `CLI_TOOL_INVOCATION_*` · `other` = `CLI_COMMAND_*`,
`CLI_SEARCH_*`, `CLI_PROXY_*`, `CLI_RUN_*`, `CLI_LOGOUT_*`
Judgment calls, made with data rather than by default:
- **`proxy` → `other`.** 2,060 of 2,140 proxy users (96%) already
execute; only 80 are proxy-only. Reclassifying would move the activated
cohort ~1%, so `execute` stays precisely "tool execution".
- **`run` → `other`.** `run` children emit their own execute events
carrying `parent_run_id`; staging the parent as `execute` would
double-count.
- **`search` → `other`.** Not a required funnel stage. Candidate for a
"discovery" step later if the flow page wants one.
## Note for whoever builds the flow page
`CLI_TOOL_INVOCATION_*` and `CLI_EXECUTE_FAILED` can both fire for a
single failure, so the `execute` stage **double-counts at the event
level**. Count **distinct users per stage**, not events.
## Known coverage limit
npm and Homebrew installs don't run `install.sh`, so they emit no
`CLI_INSTALL_*`. The `install` stage means "installed via install.sh",
not all installs.
## Verification
`pnpm typecheck` clean · `validate:boundaries` OK (46 disables, **zero
added**) · CLI suite **107 files / 971 passed** (+15 tests) ·
`dispatch.ts` identity/transport untouched.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
## What
Sends CLI telemetry **directly to PostHog** (project 196278) instead of
through Apollo/Segment, and stitches the anonymous install identity to
the Apollo user at login. This unlocks the previously-dark top of the
CLI adoption funnel (install → setup → login → connect → execute).
### Why
Two live funnel breaks in PostHog today:
1. **Install/setup/plugin events were dark.** Pre-login, `dispatch.ts`
had no API base URL (→ endpoint null → dropped), and Apollo's ingest
401s anonymous requests. So `CLI_INSTALL_*`/`CLI_SETUP_*` never arrived.
2. **Identity split.** CLI/server events landed on *email*-keyed PostHog
persons (via Segment); dashboard onboarding on *UUID*-keyed persons —
the two never joined.
The CLI already knows the canonical Apollo UUID (`org_member.id` from
session info), so it can self-stitch with no platform/dashboard change.
## How
- **Direct transport.** New PostHog capture path (`captureToPostHog`)
reusing the existing detached best-effort worker. No dependency on the
API base URL, so **pre-login events send**.
- **Identity.** Pre-login `distinct_id = install_id`; at login
`linkApolloIdentityForAnalytics(org_member.id)` persists the UUID to
`~/.composio/analytics.json` and fires one `$create_alias` (`install_id
→ org_member.id`); post-login events key on the persisted
`apollo_user_id`. Because onboarding is already UUID-keyed, the CLI
funnel now joins web onboarding.
- **Public key, baked at build.** The write-only `phc_` project key is
injected via `bun build --define` from the `COMPOSIO_POSTHOG_KEY` env
(CI Actions **variable**), never committed to source. `--define` (not a
second `--env`) is deliberate — Bun honors only the last `--env`, which
would clobber the existing `DEBUG_OVERRIDE_*` bake. For this OSS repo
this keeps the key out of git history and means **forks / local builds
compile a no-op binary** (only official releases carry the key).
- **Direct-only:** the Apollo `/cli/analytics` send is dropped for CLI
events; the CLI attaches its own `org_id` (project_id is per-session; a
follow-up).
## Guards
No-op (no send, no worker spawn, no identity-file write) when telemetry
is disabled
(`COMPOSIO_CLI_TELEMETRY_DISABLED`/`TELEMETRY_DISABLED`/`COMPOSIO_DISABLE_TELEMETRY`/`CI`/`NODE_ENV=test`)
or when no project key is configured. Alias fires once per install
(idempotent). No user API key or tool-arg values in event properties.
## Verification
- `pnpm typecheck`, `validate:boundaries`, full CLI vitest (**954 passed
/ 1 skipped**) — green.
- Independent code review pass; findings addressed (spawn-gated-on-key,
disabled=no-file, idempotent alias).
- **Live cold-install trace:** built the real compiled binary with the
key baked, ran it from a throwaway `$HOME` → fresh `install_id` → 2
pre-login events delivered to `us.i.posthog.com` with **HTTP 200** →
confirmed queryable in PostHog 196278 within ~10s, keyed by
`install_id`.
## Notes
- CLI is Changesets-ignored; human note added to `CHANGELOG.md`
directly.
- Requires the `COMPOSIO_POSTHOG_KEY` Actions variable to be set for
official release builds (done).
🤖 Generated with [Claude Code](https://claude.com/claude-code)
## Summary
Automated docs update triggered by SDK source changes on `next`.
- Claude reviewed the SDK diff and updated guides, FAQs, or examples
that reference changed APIs or features.
- The docs `@composio/*` dependencies were realigned to their latest
published releases so Twoslash snippets and example apps validate
against versions users can actually install.
## Review checklist
- [ ] Changes accurately reflect the new SDK behavior
- [ ] No unrelated docs were modified
- [ ] Code examples are correct and complete
- [ ] If a documented feature is not published yet, the Twoslash build
will fail — wait for the release instead of working around it
Generated by Claude Code via GitHub Actions.
Co-authored-by: jkomyno <12381818+jkomyno@users.noreply.github.com>
This PR:
- fixes https://github.com/ComposioHQ/composio/issues/3932
- replaces the private CLI package semver with the non-release
`0.0.0-development` sentinel
- injects the exact GitHub release tag version into every standalone
binary and verifies it during the build
- makes every push to `next` a beta, with explicit semver selection for
intentional minor or major betas and stable releases only through beta
promotion
- resolves skill installation, setup repair, companion repair, and
upgrades from packaged `release-tag.txt` metadata
- adds regression coverage for release selection, compiled version
identity, and packaged-tag asset lookup; all 1,009 CLI tests and the
monorepo typecheck pass
- updates the direct CLI changelog and release guidance; no Changeset is
added because CLI packages are excluded from Changesets
This PR:
- fixes
https://github.com/ComposioHQ/composio/actions/runs/30535159056/job/90850055191
by keeping checksum generation independent of unbuilt CLI runtime
packages
- extracts the shared Effect teardown into a dependency-light module
while preserving existing script imports
- adds an executable release regression that generates a checksum
manifest from a clean fixture
- verifies the fix with `pnpm test:release-workflow`, `pnpm
build:packages`, CLI typecheck, the full CLI test suite, and `pnpm
validate:changesets`
## What
Restores the two docs-prose skills that were dropped from the skill
tree:
- `good-docs-writing` — the house documentation style guide (Voice,
Structure, Terminology, Punctuation, Code examples, Formatting), for
drafting or revising docs prose.
- `good-docs-audit` — the review process that audits a doc/README/prose
block against that guide and reports `file:line` findings with concrete
rewrites. Report-only by default.
## Why
Both existed at `d17a268d` under `.claude/skills/` and were deleted in
#3666 (`chore(agents): normalize repo guidance skills`), which renamed
the rest of the tree into the current taxonomy but did not carry these
two over. Nothing replaced them: there is no docs-voice guidance in the
repo today, and `docs-decisions` covers docs-site mechanics (Fumadocs,
changelogs, ADRs), not prose style.
## Shape of the restore
Rule content is restored unchanged from `d17a268d`. Only the file layout
differs, because the old flat single-file form no longer passes
`validate:agent-skills` (SKILL.md is now capped at 80 lines and a linked
`references/` directory is required):
```
.agents/skills/good-docs-writing/SKILL.md # routing + the rules that matter most
.agents/skills/good-docs-writing/references/style-guide.md
.agents/skills/good-docs-audit/SKILL.md # routing + non-negotiables
.agents/skills/good-docs-audit/references/audit-process.md
```
They live under `.agents/skills` (canonical); `.claude/skills` is a
symlink, so there is no second copy.
Wiring, all three places the taxonomy is asserted:
- `ts/scripts/validate-agent-skills.mjs` — added to the `expectedSkills`
gate
- `ts/scripts/test-skill-routing.mjs` — one routing probe each (the
validator fails on any skill without a probe)
- `AGENTS.md` — added to the Skill Routing list
## Test results
- `node ts/scripts/test-skill-routing.mjs` — passes, 18 probes over 18
skills. Both new probes resolve to a unique top match.
- `node ts/scripts/validate-agent-skills.mjs` — zero findings against
tracked files.
- `pnpm lint-staged` — no staged file matches a configured task
(patterns cover `ts/packages/**` and `python/**`).
- Markdown formatted with the repo Prettier config.
Two notes on how those were run. `pnpm validate:*` could not be used
locally: pnpm's pre-run dependency check tries to purge and reinstall
the workspace and aborts without a TTY, so the scripts were invoked
directly with `node` (same entrypoints the pnpm scripts use).
Separately, `validate-agent-skills` reports stale-guidance hits under
`docs/.eve/dev-runtime/snapshots/**` on my machine; those are untracked,
gitignored local eve snapshots, unrelated to this change, and won't
exist in CI. The commit used `--no-verify` because the pre-commit hook
is just the `pnpm lint-staged` wrapper that hits the same pnpm abort;
lint-staged was run manually instead.
No changeset: guidance and tooling only, no published package touched.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This PR was opened by the [Changesets
release](https://github.com/changesets/action) GitHub action. When
you're ready to do a release, you can merge this and the packages will
be published to npm automatically. If you're not ready to do a release
yet, that's fine, whenever you add more changesets to next, this PR will
be updated.
# Releases
## @composio/core@0.14.1
### Patch Changes
- 577a3d4: Replace the backtracking leading/trailing-slash-trim regexes
in the Cloudflare Workers/Edge platform path helpers with index-walk
loops, closing a polynomial-time regular expression denial-of-service
(CodeQL js/polynomial-redos) on long runs of slash characters. Output is
unchanged for every input.
- 503b50a: Refresh runtime dependencies across the TypeScript SDK
packages.
- 2f63fe5: Guard Tool Router session URL uploads against SSRF,
revalidate redirect targets, and enforce a streamed 100 MiB response
limit across TypeScript URL upload paths.
## @composio/experimental@0.2.1
### Patch Changes
- 503b50a: Refresh runtime dependencies across the TypeScript SDK
packages.
## @composio/mastra@0.10.3
### Patch Changes
- 503b50a: Refresh runtime dependencies across the TypeScript SDK
packages.
## @composio/slim@0.14.1
### Patch Changes
- 503b50a: Refresh runtime dependencies across the TypeScript SDK
packages.
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
This PR:
- prepares Python SDK core and provider packages for `0.18.1`, including
runtime and lockfile metadata
- adds the canonical combined changelog for Python `0.18.1` and the
TypeScript releases tracked by
https://github.com/ComposioHQ/composio/pull/3906
- documents `@composio/core` `0.14.1` so the generated release PR passes
the release-workflow guard
- refreshes `mise.lock` so the TypeScript audit workflow reaches the
dependency audit
- keeps the pending TypeScript Changesets untouched; the generated
release PR remains responsible for npm versioning and publication
- validates Python with Ruff/mypy, 944 tests, all 13 package builds, and
`twine check`
- validates TypeScript/docs with the full 26-task `pnpm test`,
high-severity production audit, 100 docs static tests, link checks, docs
typecheck, and production build
## Summary
Auto-generated Python SDK reference docs from `python/composio/`.
Regenerates pages at `docs/content/reference/sdk-reference/python/` to
reflect changes in the Python package's public API (new methods, updated
signatures, changed types).
Co-authored-by: jkomyno <12381818+jkomyno@users.noreply.github.com>
## Summary
- add an experimental guide for registering customer-hosted MCP servers
- document supported authentication, the registration and sync flow,
toolkit behavior, and customer-facing workarounds for known gaps
- add the guide to the Extending sessions navigation and cross-link it
from the sessions-over-MCP guide
## Why
Custom MCP toolkits now have a working platform lifecycle, but customers
need one place that clearly distinguishes remote MCP servers from
in-process custom tools. The guide also makes the current pilot
limitations explicit so customers know which API version, sync flow,
account selection, and hosting model to use.
## Validation
- `cd docs && bun run lint:links`
- `cd docs && bun run types:check`
- `git diff --check`
Docs-only change, so no changeset is required.
This PR:
- fixes a regression introduced by
https://github.com/ComposioHQ/composio/pull/3956: fumadocs-openapi 11's
`groupBy: 'tag'` silently skips operations whose tag is not declared in
the document's top-level `tags` array, which 404'd all 16 Projects and
Organization Management operation pages (e.g.
`/reference/api-reference/projects/postProjectUsageSummary`) on both
v3.1 and v3, and dropped them from the sitemap and search index
- adds `declareOperationTags()` (`docs/lib/openapi-tags.ts`), applied at
spec load time in `lib/openapi.ts` (covers build-time-fetched specs) and
at sync time in `scripts/fetch-openapi.mjs`, which now also warns when
the upstream payload omits tags; the checked-in specs are normalized
accordingly (purely additive)
- adds CI guards in `tests/static/api-reference-routes.test.ts`: spec
invariants (every visible operation has a tag and `operationId`),
bidirectional spec-vs-generated-routes set equality through the
production loader path, a negative test pinning fumadocs' silent-drop
behavior (fails when upstream fixes it, signaling the workaround can be
retired), and validation that every `ApiEndpointsTable` quick-link href
resolves to a generated page
- extends `Docs - Check Links` with a nightly (02:30 UTC) + manual
external-URL sweep via a new `lint:links:external` script; the validator
only fails on evidence a link is dead (404/410 or repeated network
errors, with UA/timeout/GET-fallback/retry and per-URL caching) and
scheduled failures file a deduplicated tracking issue; PR runs keep the
fast internal-only check
- fixes six dead external links the first sweep found (moved Claude
Agent SDK docs, stale dashboard settings URL ×2, a `master`-branch path
now pinned to tag `0.5.0+post.1`, and two removed `ComposioHQ` example
repos whose mentions were dropped — their code is embedded in the pages
via `RepoBrowser`)
## Context
fumadocs-openapi 10 generated a page for any tag string found on an
operation; v11 requires the tag to be declared top-level and skips
silently otherwise (`preset-auto.js`: `builder.fromTagName(tag)` →
`continue`, no warning). The backend spec generator omits `Projects`,
`Organization Management`, and `Invite Codes` from `tags`, so their
operation pages vanished while the checked-in MDX tag landing pages kept
rendering with dead quick links. With the normalizer, the generated URL
set is byte-identical to the pre-#3956 set (234 routes, verified via
`getReferenceSource().getPages()` diff); no revert needed.
Existing checks missed this because `lint:links` only extracts markdown
links and `Card` hrefs (the dead links live in `ApiEndpointsTable`'s
array prop) and validates against the same loader that shrank, and the
integration suite samples fixed routes under tags that survived. The new
completeness guard derives expectations from the specs themselves, so
routine data syncs don't churn it.
## What
Closes the PRDE-1138 gap: an unattended agent hitting plain `composio
login` in a pipe dead-ended at browser OAuth. Now the default headless
path either completes unattended or tells the agent exactly what to run
— without ever auto-creating an account for a human.
- **Offer**: non-interactive login instructions gain a second hint —
"For unattended agents: run `composio login --agent`" — alongside the
existing URL + `composio login --poll` flow.
- **Perform (reuse only)**: when `~/.composio/agent.json` already holds
a READY agent identity, plain headless `composio login` completes
unattended by reusing it. New `getStoredReadyAgent` in
`services/agents.ts` is the reuse-only counterpart of
`getOrSignupReadyAgent` — it never falls back to signup.
- **Guardrail** (the design constraint from the ticket): account
creation stays exclusively behind the explicit `composio login --agent`
/ `composio signup` commands. A human piping `composio login` with no
stored agent identity gets the same URL instructions as before and no
account. Asserted at both unit and E2E level (`/api/signup` is never
called on the default path).
- `install.sh --agent` login now sets
`COMPOSIO_CLI_INVOCATION_ORIGIN=installer` for analytics attribution.
- Root `.dockerignore`: host `node_modules` (pnpm store symlinks that
break outside the host) no longer enter the e2e image build context;
without this, e2e images cannot build from a checkout that has run `pnpm
install`.
## Verification
- `pnpm typecheck` green; `@composio/cli` suite green — 1001 passed, 1
skipped; `validate:boundaries` clean.
- New Docker E2E suite `ts/e2e-tests/cli/agent-signin` (10/10) against a
mock `agents.composio.dev` (`COMPOSIO_AGENTS_BASE_URL`): unattended
nothing→authenticated (`login --agent` then `whoami` in one container),
headless-human guardrail (instructions printed, zero signup calls),
stored-identity reuse (login completes + whoami reports `account_type:
agent`). No real accounts are created.
## Notes
- No VHS recordings: the visible change is piped-mode text output, which
VHS interactive recordings don't demonstrate.
- No changeset by design (`@composio/cli` is Changesets-ignored);
human-facing note added to `ts/packages/cli/CHANGELOG.md` under 0.3.0.
getStoredReadyAgent collapsed every whoami failure into Option.none via
Effect.option and fell back to the on-disk identity, so a revoked agent
key (HTTP 401/403) still completed headless login. Auth rejections now
return no agent; only transport-shaped failures keep the offline
fallback.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A root .dockerignore excluding **/node_modules broke Dockerfile.node's
runtime fixtures: suites without a setup phase (claude-agent-sdk) resolve
deps through node_modules copied in via COPY ts/e2e-tests/. BuildKit's
per-Dockerfile ignore keeps the exclusion for Dockerfile.cli, which
installs inside the image and must not see host pnpm symlinks.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
## What
Removes npm and Homebrew as Composio CLI install channels. `curl -fsSL
.../install.sh | bash` is now the only supported channel (plus manual
zip / from-source).
Channel decommission decision: KJ, 2026-07-27. Tickets: PRDE-1155 (npm),
PRDE-1156 (brew).
## Changes
- **INSTALL.md** — removed the npm/pnpm/yarn install section; Windows
guidance is now explicitly WSL
- **install.sh** — Windows error path no longer advises `npm install -g
@composio/cli` (dead channel); points to WSL
- **.github/workflows/build-cli-binaries.yml** — generated per-release
INSTALL.md no longer includes an npm section; comment no longer lists
Homebrew as a release consumer
- **.github/workflows/cli.install-health-check.yml** — dropped the stale
comment describing an npm dist-tag publish flow that no longer exists
- **.github/workflows/cli.test-installation.yml** — deleted the
`test-npm-fallback` job (a no-op that only checked `npm --version`) and
its `toolchain-versions` feeder job; pruned summary references
- **.github/workflows/cli.bump-homebrew-tap.yml** +
**.github/scripts/bump-homebrew-formula.py** — deleted. This automation
never fired reliably: releases are published with `GITHUB_TOKEN`, which
suppresses `release`-triggered workflows, and `HOMEBREW_TAP_TOKEN` is
dead (root cause recorded on PRDE-1156). Removal loses nothing.
- **ts/packages/cli/package.json** — dropped the unused `publishConfig`
block (package stays `private: true`)
- **.agents/skills/cli-release/** + **ts/packages/cli/AGENTS.md** —
removed Homebrew steps from release docs
- **toolchain-versions.json** — dropped `node_install_compat`, whose
only consumer was the deleted npm-fallback job
## Deliberately untouched
- SDK npm packages (`@composio/core` etc.) — only the CLI channel dies
- Historical CHANGELOG entries mentioning npm
- `brew install mise` / `brew install steipete/tap/peekaboo` —
third-party tools, unrelated to the CLI channel
## Cross-repo
Companion PR on the Homebrew tap repo replaces the formula with a
deprecation stub that migrates existing brew users (0.2.28) to the curl
installer.
## Verification
- `bash -n install.sh`
- `test/install-sh-release-resolution.test.sh` passes locally
- YAML/JSON parse checks on all edited workflow/manifest files
- grep sweeps: zero remaining user-facing npm/brew CLI-install
references
🤖 Generated with [Claude Code](https://claude.com/claude-code)
This PR:
- follows up https://github.com/ComposioHQ/composio/pull/3943, replacing
its hardcoded message with a fix for the cause it worked around
- stops `install.sh` redirecting `composio install`'s stderr to a temp
file — that redirect is what silenced the command, since the CLI gates
every log, note and outro on stderr being a terminal, so the captured
file was always empty and the replay printed nothing
- probes `"$exe" --version` before delegating, so a binary that cannot
run still falls back to inline setup without spilling loader errors —
the reason the redirect existed in the first place
- makes `composio install` fall back to plain `ui.error` lines when
there is nothing to decorate, so container and CI installs that pipe the
session to a build log still get the report; interactive output is
unchanged. `canDecorate` is constant for a run, so each channel binds to
its decorated or plain form once rather than re-testing per message
- drops the unconditional "Composio was added to your PATH" line, which
printed on three paths where nothing was written: shell-not-detected and
unsafe-install-dir (both exit 0 after reporting manual setup), and the
inline fallback's unsupported-shell and unwritable-rc branches
- adds regression tests using the production `makeTerminalUI`
implementation with captured streams, so the real decoration gate is
exercised — the shared test double reports `canDecorate: false` but
prints everything anyway, which is why this went unnoticed through two
prior attempts
## Context
Measured against the released binary, `composio install` writes 0 bytes
when its stderr is not a terminal, and exits 0 without touching any rc
file when `$SHELL` is unset. Both were invisible to the installer:
| case | exit | stderr bytes | PATH block written |
| --- | --- | --- | --- |
| `2>file`, as `install.sh` did | 0 | 0 | yes |
| `$SHELL` unset | 0 | 0 | no |
| stderr on a terminal | 0 | full report | yes |
With the redirect gone, the delegated path prints what the user was
missing — `Updated ~/.zshrc` and `Restart your shell to apply changes` /
`source ~/.zshrc` — and the manual-setup text on the paths where PATH
genuinely was not touched. The closing block no longer needs a PATH
claim of its own, and the fallback keeps printing its own
`refresh_command`.
067c67b695 already tried to surface this
output by capturing it to a temp file and replaying it on success. That
replay has been a no-op for the same reason, so this guidance has never
reached a user since the delegation was introduced.
## How Has This Been Tested?
- new tests fail with `expected '' to contain 'PATH: will add'` (empty
output — the production bug) without the `install.cmd.ts` change, and
pass with it
- full CLI suite: 957 passed, 1 skipped; `pnpm typecheck` 14/14;
`validate:boundaries`, eslint and prettier clean, no new eslint disables
- released binary in a sandboxed `$HOME`: stderr to a file yields 0
bytes; stderr on a pty yields the full report
- CLI from source with stderr captured: 155 bytes of plain guidance
where it previously wrote nothing; decorated output on a terminal
unchanged
- `install.sh` driven through a harness built from the real script under
a pty: a working binary's stderr reaches the terminal, a broken binary's
loader noise stays suppressed and the fallback writes the PATH block and
prints `source ~/.zshrc`. The same harness against `next` shows the
swallow.
No changeset: `@composio/cli` is Changesets-ignored and `install.sh` is
not a published package. The CLI note went into `CHANGELOG.md` directly.
`install.sh` is served from `next`, so the installer half of this
reaches new installs on merge and works with already-released binaries.
The `install.cmd.ts` half only affects captured-stderr installs and
rides the next CLI release.
This PR:
- restores TypeDoc generation after
https://github.com/ComposioHQ/composio/pull/3966 by resolving
`@composio/core`'s TypeScript dev dependency from `catalog:ts6`
- keeps core typechecking on the root TypeScript 7 compiler while
TypeDoc receives the TS6 compiler API
- documents the additional compiler-API consumer in the shared
TypeScript catalog
- commits the regenerated `session-files.mdx` output that could not be
produced while the workflow was broken
- verifies `generate:docs`, `typecheck`, and `build` for
`@composio/core`
Teach non-interactive login to reuse an existing READY agent identity without creating one implicitly. Preserve explicit --no-browser and --no-wait contracts, and branch on TerminalUI.canPrompt so piping stdout alone remains interactive.
Add unattended guidance, installer analytics attribution, and focused unit and Docker E2E coverage. Exclude host node_modules from Docker build contexts.
## Problem
TriggerSubscription._handle_chunked_events is bound directly as a Pysher
channel callback. Pysher channel dispatch has no local exception
boundary, so malformed chunk frames reached websocket-client error
handling and marked the realtime connection failed. The normal payload
parser already skips malformed frames, but the chunked path did not.
## Fix
- Validate that each decoded frame is an object with string id/chunk,
integer non-boolean index, and boolean final fields.
- Contain decoder, validation, and reassembly failures at the callback
boundary and log only a bounded frame preview.
- Clear partial state only when the frame carries a validated string id,
allowing that id to be reused safely.
- Preserve the valid chunk reassembly path and fix the Chunked docstring
typo.
## Regression coverage
Tests cover invalid JSON, non-object frames, missing fields, unhashable
ids, wrong field types, unexpected decoder failures, and successful
same-id reassembly immediately after a bad frame clears partial state.
## Verification
- nox -s tst -- tests/test_triggers.py: 83 passed
- make chk: Ruff and mypy passed
---------
Signed-off-by: Checo <sergio@checo.cc>
Co-authored-by: Checo <sergio@checo.cc>
Co-authored-by: jkomyno <alberto@composio.dev>
## Summary
`TerminalUI` classified the terminal with a single aggregate:
```ts
isInteractive = stdinIsTTY && stdoutIsTTY && stderrIsTTY;
```
and reused it for three unrelated decisions: whether Clack may prompt,
whether `ui.output()` emits machine data, and caller behavior in login,
setup, and the update notice. That let output routing change program
behavior:
- Redirecting stdout made `login` look non-interactive even though stdin
and stderr were still usable, so `composio login | tee log` silently
switched to the headless instructions path.
- Redirecting only stdin made `ui.output()` write machine data onto a
human-visible stdout terminal, because the output gate keyed off the
prompting aggregate instead of stdout.
- Capturing stderr changed stdout data emission, even though it is
purely a reporting-mode concern.
This PR separates the three capability axes so each stream owns exactly
one decision:
| Axis | Rule |
| --- | --- |
| Prompting | `canPrompt = stdinIsTTY && stderrIsTTY` — stdout is
irrelevant |
| Machine output | `ui.output(data)` writes only when stdout is **not**
a TTY, or with explicit `{ force: true }` (unchanged) |
| Decoration | `canDecorate = stderrIsTTY` (unchanged) |
The aggregate `isInteractive` capability is removed rather than renamed,
so no caller can reintroduce the conflation.
## Changes
- `services/terminal-ui.ts`: `TerminalCapabilities` now exposes
`canPrompt` instead of `isInteractive`; `output()` gates on
`stdoutIsTTY` alone. The live implementation is now built by an exported
`makeTerminalUI(streams)` factory so tests can exercise real behavior
against injected streams; `TerminalUILive` passes
`process.{stdin,stdout,stderr}` as before.
- `commands/login.cmd.ts`: waiting for browser login is a prompting
decision (`canPrompt`); intro/notes are stderr chrome (`canDecorate`).
Piping stdout no longer reroutes authentication.
- `commands/setup.cmd.ts`: the `--yes` guards for local changes now key
off `canPrompt`.
- `services/update-check.ts`: the update notice is advisory output for a
human watching stderr, not part of any command's result, so it now keys
off `canDecorate` — piping stdout no longer suppresses it, and it still
never touches stdout, so piped data stays clean.
- Telemetry continues to receive the raw per-stream TTY bits
(unchanged).
- `ts/packages/cli/AGENTS.md`: output conventions now document the three
independent stream contracts.
## Tests
- Full 8-combination TTY matrix at the classifier level.
- Behavioral tests with fake sink streams and mocked Clack prompts,
proving behavior rather than returned booleans:
- stdin redirected, stdout+stderr TTY: prompts fall back to defaults
without invoking Clack, `ui.output()` does **not** leak JSON to the
visible terminal, `force` still writes, decoration renders;
- stdout piped, stdin+stderr TTY: the Clack prompt actually runs,
machine data emits, decoration renders;
- stderr captured, stdin+stdout TTY: no prompts, no decoration, stdout
suppressed unless forced;
- everything redirected: machine data still emits.
- Login regression: with stdout piped but stdin/stderr TTYs, `composio
login --no-browser` enters the interactive wait (poll spinner) path and
does not print the headless/agent instructions.
- Update-notice regressions: still shown when stdout is piped or stdin
is redirected; suppressed only when stderr is captured.
## Notes
This is the shared prerequisite for
https://github.com/ComposioHQ/composio/pull/3951 (captured-stderr plain
renderer) and https://github.com/ComposioHQ/composio/pull/3945
(agent-friendly headless login). It intentionally implements neither
feature; both PRs should rebase onto this once merged. No changeset:
`@composio/cli` is Changesets-ignored.
## Verification
- `pnpm typecheck` (repo root)
- `pnpm --filter @composio/cli test` (validate:skills,
validate:boundaries, full vitest suite)
- `eslint` and `prettier --check` on all touched files
VHS recordings skipped: internal capability wiring, no visible
command-surface change.
This PR:
- fixes the `ts.build.yml` push trigger, which still pointed at
`master`: the repo's default branch is `next`, so the workflow has not
run on push since 2025-10-24 and never populated the turbo/pnpm caches
that PR runs restore from (GitHub cache isolation lets PRs read
base-branch caches, but not other PRs')
- fixes the same stale `master` push trigger in `py.check.yaml` (all
other workflows already list `next`)
- adds a `cache-save` input (default `'true'`, preserving current
behavior for every other consumer) to the `setup-node-pnpm-bun`
composite action, swapping `actions/cache` for `actions/cache/restore`
(same repo, same v6.1.0 SHA pin) when disabled
- wires `cache-save` in `ts.build.yml` so only push runs on `next`
persist caches: PR-scoped saves are invisible to other branches, so they
only cost post-job time and evict default-branch entries from the 10 GB
repo cache quota
## Context
Measured on run
[30364148082](https://github.com/ComposioHQ/composio/actions/runs/30364148082)
(PR run, ~90s wall time):
| Step | Time |
|---|---|
| Checkout | 5s |
| Setup Node.js, pnpm, Bun | 34s (mise 1s, setup-node 5s, setup-bun 2s,
pnpm 2s, **pnpm store restore 18s** for 658 MB, turbo restore 1s) |
| `pnpm install --frozen-lockfile` | 14s |
| `pnpm lint` | 1s |
| `pnpm run build:packages` | 24s, **0/19 turbo cache hits** |
| Post-job cache saves | ~11s (turbo save + a 658 MB pnpm store save
that failed with "Unable to reserve cache", racing the parallel ts.*
jobs on the same PR) |
The build restored a turbo cache only via the fallback restore-key
`Linux-turbo-build-` from an unrelated branch, hence 0 hits. Once push
runs fire on `next` again, every merge saves a fresh
`Linux-turbo-build-<sha>` cache that PR runs match through the
restore-key prefix, so unchanged packages become turbo hits (up to ~20s
off the build step) and the ~11s post-job save disappears from PR runs.
The per-commit key suffix is kept intentionally: the turbo cache
accumulates, so each default-branch save needs a fresh key and readers
match via the prefix (documented inline in the action).
This PR:
- replaces ESLint with oxlint across the pnpm workspace and the
Bun-based docs site, porting the rules to `.oxlintrc.json` /
`docs/.oxlintrc.json` with behavior parity (restricted-syntax selectors
kept via `oxlint-plugin-eslint`)
- migrates typecheck to TypeScript 7 (`typescript@^7.0.2` catalog) and
keeps a TS6 pin for JS compiler API consumers via a named `ts6` pnpm
catalog (`ts/scripts/validate-examples.ts`, the `@composio/cli` generate
pipeline). The CLI's `typescript` dependency rebinds only the
compiler-API import — its typecheck still runs the root TS7 `tsc`, since
the alias package only ships a `tsc6` bin (documented in
`ts/packages/cli/AGENTS.md`)
- removes the `paths` mappings that pointed `@composio/core` (and, in
`experimental`, `@composio/json-schema-to-zod` plus core-internal
`#`-imports) at sibling `src` directories: under TS7, tsdown's
tsgo-based dts step emitted stray `.d.ts` files next to those
out-of-root sources on every dependent package build. Workspace deps now
resolve through their built dist types, which turbo's `dependsOn:
^build` already guarantees exist — and which the deep-path exports
(`@composio/core/*`) always used anyway
- renames the cli boundary tooling `eslint-boundaries*` →
`lint-boundaries*` and hardens the scanner to reject `oxlint-disable`
spellings so the disable manifest cannot be bypassed
- rewrites inline `eslint-disable` comments to oxlint rule names
(comment-only; no runtime changes), and adds **one new** declared
boundary: `tool-file-uploads.ts` needs `no-restricted-imports` disabled
for `node:crypto` (MD5 for the presigned-upload checksum is not in Web
Crypto), because oxlint also catches dynamic `await import()` where
ESLint did not. The manifest grows 46 → 47 deliberately
- updates CI path filters, `turbo.jsonc` lint inputs, and the docs
typescript-check workflow (renamed to "Docs - Lint and TypeScript
Validation" since it now lints too); drops `eslint`,
`typescript-eslint`, `eslint-config-next`, and `globals` from the
dependency graphs
- ships no changeset: I built `@composio/core` and `@composio/anthropic`
on this branch and on the pre-migration base and diffed the emitted
`dist/**/*.d.mts`. The provider output is byte-identical. Core's output
is **semantically identical but not byte-identical**: TS7 changes quote
style (`"x"` → `'x'`), object-property and union-member ordering in
inferred types, and picks equivalent shorter re-export alias paths for
five signatures (e.g. `OpenAI.Beta.Threads.Runs.Run` →
`OpenAI.Beta.Threads.Run` — verified both names alias the same type in
the shipped typings). Chunk-name hashes shift as a consequence. No type
gains, losses, or shape changes; `attw` and `publint` pass on the TS7
build
## Context
First of a three-PR split of #3958. The type-safety refactors are
stacked on this branch and merge after it:
- docs: https://github.com/ComposioHQ/composio/pull/3967
- `@composio/core`: https://github.com/ComposioHQ/composio/pull/3968
This PR:
- codifies the schema-parsing convention in agent guidance: parse
untyped/external data once at the boundary with zod (`z.infer`
downstream) in the docs site and SDK packages, and with `effect/Schema`
in the CLI
- bans hand-rolled structural guards (`'x' in obj` / `typeof` chains),
`as`-casts of parsed JSON, and fake validators such as `z.custom(() =>
true)`
- updates `docs/AGENTS.md`, `ts/AGENTS.md`,
`ts/packages/core/AGENTS.md`, and the existing trust-boundary rule in
`ts/packages/cli/AGENTS.md`
## Context
Follow-up to the review feedback that led to the #3958 split (#3966 /
#3967 / #3968): the docs type-safety work initially shipped a suite of
conditional structural utility helpers instead of zod schemas. This
records the convention so agents and contributors reach for schemas by
default.
## Summary
`composio listen` crashes with `TypeError: Object is not a constructor`
immediately after realtime init in every compiled release binary that
bundles pusher-js 8.5.0 (reproduced on darwin-arm64 0.2.33-beta.291–298
and 0.2.32-beta.289, and linux-x64 0.2.33-beta.298). Full diagnosis in
#3918.
**Root cause (corrected after independent verification):** pusher-js
changed the shape of its CJS export between versions — 8.4.x exports the
constructor itself (`module.exports = Pusher`), **8.5.0's Node bundle
exports a namespace object** (`module.exports = { Pusher }`; its webpack
bundle ends with `module.exports.Pusher = __webpack_exports__`), and
8.6.0 exports the constructor again. This is an undocumented upstream
packaging regression — reported in
[pusher/pusher-js#935](https://github.com/pusher/pusher-js/issues/935),
fixed by
[pusher/pusher-js#936](https://github.com/pusher/pusher-js/pull/936),
first released in 8.6.0 (2026-07-23) whose release notes do not mention
it. With 8.5.0 (what the crashing published binaries bundled),
`pusherModule.default` is not callable under **either Node or Bun, from
source or compiled** — the constructor is only reachable via the named
`Pusher` export. The earlier "build-target-dependent Bun interop" theory
in this PR's first revision and in #3918 is not correct: compiled
interop probes built with bun 1.3.10 for darwin-arm64 and bun-linux-x64
embed byte-identical bundles, so the shape varies by pusher-js version,
not by build target. "Works from source" reports came from non-frozen
installs resolving `^8.5.0` to 8.6.0.
Note: `next` now pins pusher-js 8.6.0 (#3955), which incidentally
restores a callable `module.default` in new builds — and this branch has
since been merged up with `next`, so it resolves 8.6.0 too. This PR
makes that robustness deliberate instead of incidental: pusher-js has
changed its export shape twice within two minor versions, and the new
real-module regression test has now passed against both shapes (8.5.0
pre-merge, 8.6.0 post-merge).
Fixes#3918
## Changes
- `ts/packages/cli/src/services/triggers-realtime.ts`:
- `resolvePusherConstructor` probes the interop shapes in precedence
order — `module.default.default` (defensive double-wrap),
`module.default` (8.4.0/8.6.0 class export), the module object itself,
then the named `Pusher` export on either level (`module.default.Pusher`
is the 8.5.0 shape that crashed release binaries) — returning the first
callable candidate via `Arr.findFirst` with
`Predicate.hasProperty`/`Predicate.isFunction` narrowing (no `as`
casts).
- Resolution failure surfaces as the service's typed
`TriggerRealtimeSubscriptionError` instead of an untyped `TypeError`.
- `new Pusher(...)` itself is now wrapped in `Effect.try`, so a
callable-but-non-constructable candidate would also land on the typed
error channel instead of escaping as an untyped defect past the listen
commands' `catchTag`.
- `ts/packages/cli/test/src/services/triggers-realtime.test.ts`:
regression tests pinning every interop shape, precedence (including
`default.Pusher` isolated from the top-level `Pusher` fallback), the
module-object fallback, the no-constructor failure case, and the **real
installed pusher-js module** — so the next packaging drift fails the
unit suite rather than only the compiled binary.
- `ts/packages/cli/CHANGELOG.md`: Patch Changes entry under 0.3.0 (the
CLI is Changesets-ignored, so notes go in the changelog directly),
stating the verified root cause.
## Type of change
- [x] Bug fix
- [ ] New feature
- [ ] Refactor/Chore
- [ ] Documentation
- [ ] Breaking change
## How Has This Been Tested?
- `pnpm exec vitest run test/src/services/triggers-realtime.test.ts` — 9
passed.
- `pnpm --filter @composio/cli test` — full package suite green
(includes `validate:skills` and `validate:boundaries`).
- CLI package `pnpm typecheck` (src + test) and `prettier --check`
clean.
- Independent empirical verification with the release toolchain (bun
1.3.10, pinned in `mise.toml`):
- pusher-js **8.5.0**: `module.default` is `{ Pusher }` (not callable)
under Node 24, Bun from source, and `bun build --compile --production`
binaries — the pre-fix `new (module.default)(...)` throws the exact
#3918 `TypeError` in all three environments; the resolver picks the
named `Pusher` export and constructs successfully.
- pusher-js **8.4.0** and **8.6.0**: `module.default` is the
constructor; the resolver picks it directly and constructs successfully.
- darwin-arm64 vs cross-compiled bun-linux-x64 binaries embed
hash-identical pusher bundles — the interop shape does not vary by build
target, only by pusher-js version.
- Compiled-binary before/after on darwin-arm64 and linux-x64 as
described in the original report (crash reproduced from a pre-fix `next`
binary; fixed binary subscribes, holds the full `--timeout` window,
disables the temporary trigger on exit, exit 0); end-to-end Gmail
trigger delivery verified twice.
## Screenshots (if applicable)
N/A — no visible command-surface change (internal wiring fix).
## Remaining follow-up (out of scope for this PR)
- Release CI (`build-cli-binaries.yml`) smoke-tests binaries with only
`--version`/`--help`; a per-target step exercising the bundled
`import('pusher-js')` + constructor resolution would catch the next
interop drift before publishing (requested in #3918).
- The listen commands append network/login remediation hints to every
`TriggerRealtimeSubscriptionError`, which is misleading for packaging
failures, and the probed module shape is not logged on failure; a
`reason` discriminant on the error would make a future variant easier to
diagnose.
## Checklist
- [x] I have read the Code of Conduct and this PR adheres to it
- [x] I ran linters/tests locally and they passed
- [x] I updated documentation as needed (CLI `CHANGELOG.md`; no
docs-site change — behavior matches what `composio listen` already
documents)
- [x] I added tests (`test/src/services/triggers-realtime.test.ts`)
- [x] I added a changeset if this change affects published packages —
not applicable: `@composio/cli` is excluded from Changesets by design;
changelog entry added instead
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: jkomyno <alberto@composio.dev>
The analytics module resolved user_data.json and config.json from a
hardcoded ~/.composio, but the CLI writes both under the cache dir
(COMPOSIO_CACHE_DIR override, ~/.composio otherwise; see
setup-cache-dir.ts and cli-user-config.ts). With the override set,
telemetry silently lost the api key fingerprint, org_id, and the
keychain-mode signal it needs for identity attribution.
PostHog-bound events carried raw error messages, the user's search query,
and the proxied endpoint URL. Free text like this can embed API keys,
customer identifiers, and private URLs, and the client-side scrubber only
catches known token shapes. Record the shape instead: error_name only for
failures, query_length/query_term_count for search, has_endpoint for proxy.
The codact-failure path keeps its error_message: it goes to Composio's own
authenticated endpoint, not to PostHog.
Fixes#3858 (both bugs reported there).
Bug 1 — TriggerSubscription.stop() deadlocks when called from a trigger
callback. pysher's Connection.disconnect runs socket.close() then
join(timeout) on the websocket thread, but callbacks execute on that
same thread's dispatch path (_handle_event blocks on the callback's
future.result()), so joining the thread from within a callback that runs
on it deadlocks the whole process and _alive is never cleared, so a main
thread parked in wait_forever() never exits. Fix: clear _alive
synchronously (so wait_forever unblocks immediately) and run
disconnect() on a daemon thread, breaking the reentrancy while still
closing the socket.
Bug 2 — _SubcriptionBuilder.connect() leaks a websocket thread on
timeout. The deadline path raised ComposioSDKTimeoutError without
calling pusher.disconnect(), so pysher's run loop kept redialing every
reconnect_interval for the life of the process — one leaked thread per
timed-out subscribe(), and if a leaked connection later succeeded it
would authenticate and subscribe the channel of an abandoned
TriggerSubscription, silently consuming events. Fix: wrap the wait loop
in try/except so the timeout path tears the connection down before
re-raising.
Added regression tests (TestTriggerSubscriptionStop and
TestSubscriptionBuilderConnectTimeout) that reproduce both failure modes
offline with mocked pusher connections; all 73 tests in
tests/test_triggers.py pass, ruff check + format clean.
---------
Co-authored-by: Alberto Schiabel <jkomyno@users.noreply.github.com>
Co-authored-by: jkomyno <alberto.schiabel@gmail.com>
Kshitij Jhunjhunwala