test(outpost): cover event get, attempt get, tenant token and retry

Raises CLI-level acceptance coverage from 22/30 to 26/30 leaf commands. All
four reuse data the existing tests already create, so they add coverage without
adding setup.

The tenant token assertion checks shape rather than contents — three JWT
segments, and that the raw tenant id is not readable in it. The token is a real
credential, so a test should not print or match on its payload.

The four commands still uncovered are the tenant portal and its custom domain.
They are not omitted casually: `custom-domain set` configures a real DNS-verified
hostname on the shared project, and `tenant portal` returns 404 until one exists.
Covering them safely needs a dedicated throwaway domain. They are the least
proven surface and should be called out as such in beta release notes.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wajsob136PyRh6nc92w5L3

fix: v2.5.0 bug fixes across listen, config, and validation (#342)

* fix: make the CLI safe to run without a terminal

Six triaged issues plus three related bugs found while scoping them. They share
one failure shape: the CLI silently does something other than what the caller
asked, and the first symptom is missing traffic rather than an error.

Design rule applied throughout: a prompt must never be the only path. Without a
terminal, reads pick the safe default and writes fail loudly; rendering degrades
rather than dies; exit codes tell the truth.

- #332 `ci --local` / `login --local` no longer write the global config. The
  login flows saved the profile before --local was inspected, so the flag added
  a second write instead of redirecting the first, silently repointing the
  machine's active project. Gated at the single writeConfig chokepoint.

- #333 `listen` falls back to compact output when stdout is not a terminal. The
  interactive renderer opens /dev/tty itself, so in CI, Docker, nohup or an AI
  agent it died at startup and the command exited 0 having forwarded nothing.
  Renderer failures now propagate as a non-zero exit rather than being logged
  and discarded.

- #334 `listen` honours HOOKDECK_API_KEY. It previously fell through to a guest
  account: traffic arrived locally so it looked like it worked, but none of it
  was in the caller's project. HOOKDECK_API_KEY holds a Project API key, which
  /cli-auth/validate rejects with 401, so it is exchanged for a CLI client key
  via POST /cli-auth/ci exactly as `hookdeck ci` does, then saved. Precedence:
  --cli-key, stored login, HOOKDECK_API_KEY, guest.

- #335 Explicitly empty secret and identity flags are rejected. `"$UNSET_VAR"`
  expands to "", and hasAny() is a pure != "" test, so the value was dropped and
  the source created with no verification at all while looking configured. An
  empty value never cleared anything: the API expresses "no verification" as a
  null auth object, and `--config '{"auth": null}'` still does that.

- #336 The release skill's CI gate used the legacy commit-status API, which
  GitHub Actions never writes, so it returned pending off total_count 0 and
  blocked every release. Replaced with the GraphQL statusCheckRollup. The skill
  was also duplicated byte-for-byte under .agents/; that is now a symlink, as
  .claude/skills and .cursor/skills already were.

- #331 Dropped go-github v28, whose only use was one unauthenticated call and
  whose only other effect was linking x/crypto/openpgp through package init
  (GO-2026-5932, no upstream fix). Replaced with a net/http call, now testable
  and tested. Added a govulncheck job, since nothing enforced the scan.

Also fixed, previously unreported:

- Unauthenticated commands no longer hang. Any command failing for want of
  credentials dropped into interactive login: blocking on Enter, opening a
  browser, then polling for ~4 minutes. Now gated on a terminal, and reports how
  to authenticate without one.

- Destructive commands no longer exit 0 on a skipped confirmation. The five
  delete/dismiss commands called fmt.Scanln and discarded its error, so without
  a terminal they printed "cancelled" and returned nil: a CI job deleted nothing
  and reported success. They now fail and name --force.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RRp3XjWGqiYSun29SSNVJt

* test: cover the real-world usage each fix was reported from

The previous acceptance tests asserted configuration and error text. These
assert the outcomes users actually reported, and each one has been verified to
fail against unmodified main and pass on this branch.

- listen forwards a real event end to end with no terminal. Starts a local HTTP
  server, creates a source and CLI connection, runs `listen` with no --output,
  POSTs to the source URL and waits for the payload to arrive locally. This is
  the scenario #333 was found in: start a service, start a tunnel, send an
  event, check it arrives. Asserting only on the absence of a TTY error would
  have passed even if nothing was being forwarded.

- Unauthenticated commands fail fast. Asserts elapsed time, not just the error:
  against main this test takes 261s, which is the ~4 minute browser-login poll
  #337 describes. A message-only assertion would pass while still hanging.

- `ci --local` leaves the active project alone, checked through `whoami` rather
  than file contents. The reported symptom was not "a file changed", it was that
  every other hookdeck invocation on the machine silently moved project (#332).

- delete without --force now covered on destination, connection and
  transformation as well as source. The five commands share one confirmation
  helper, but each has to call it; a command that quietly stops doing so keeps
  working and just stops catching the bug (#338).

Issue dismiss is covered by unit tests only: it shares the same helper, and
creating a real issue takes ~40s of setup for a one-line wiring check.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RRp3XjWGqiYSun29SSNVJt

* fix: address Copilot review — guest-profile precedence, test race, wording

Five review findings, all valid. One was a real bug.

- HOOKDECK_API_KEY was ignored once a guest profile existed. GuestLogin
  persists its key, so Profile.APIKey is non-empty after a single guest run and
  the "no credentials stored" test treated the throwaway account as a real
  login. Every later run then stayed on the guest project despite an exported
  Project API key — #334 again, one run later. A guest profile is now
  identified by guest_url alongside the key, and the environment key takes
  precedence over it while still losing to a real stored login. Replacing a
  guest profile is announced rather than silent, since discarding a sandbox
  link is exactly the kind of side effect this release is about surfacing.

- The acceptance helpers raced. os/exec writes subprocess output from its own
  goroutine while the tests read the buffer mid-run, so a bare bytes.Buffer
  could return partial output. Reads and writes now go through a mutex.

- "Check it is exported" misdiagnosed the reported cause. In #335 the secret
  was in a workspace .env that application code loaded but the shell never did,
  so the variable was not set in the invoking shell at all; exporting only
  matters for a variable that is set locally but not visible to child
  processes. The error and the README now say "set in this shell" and note that
  .env values are not loaded automatically.

- REFERENCE.md advertised --cli-key as a global flag. The root persistent flag
  was registered unhidden while --api-key beside it is hidden, and regenerating
  the docs surfaced it — contradicting README "CLI authentication keys" and
  AGENTS.md, which state authentication is command-specific. It is now hidden
  like --api-key: still functional for existing callers, no longer advertised.
  The command-specific `login --cli-key` and `listen --cli-key` remain
  documented.

Adds the guest-profile acceptance case the review asked for, using a synthetic
guest config so it stays deterministic and does not create a real guest account
on every CI run.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RRp3XjWGqiYSun29SSNVJt

* refactor: make .agents/skills the canonical skills location

.agents/skills is the cross-harness location — it is not tied to any one tool,
so it is the right home for the real files. This inverts what landed earlier in
this branch, which kept the files at a root skills/ and pointed .agents/skills
at it.

- .agents/skills/hookdeck-cli-release/ now holds the files (tracked as a rename,
  so history is preserved).
- .claude/skills and .cursor/skills are symlinks to ../.agents/skills.
- Root skills/ is gone; there is one canonical directory and no second copy to
  drift out of sync, which was the underlying problem in #336.

The skill sits one directory deeper than before, so its relative links needed
../../ -> ../../../. All four (release workflow, .goreleaser/, .goreleaser/mac.yml,
README) verified to resolve. AGENTS.md, README.md and CLAUDE.md updated to point
at the new location.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RRp3XjWGqiYSun29SSNVJt

* fix: address second Copilot pass — error context, help text, issue coverage

Three findings, all valid.

- The HOOKDECK_API_KEY guidance was being discarded. CILogin returns a 401
  APIError, and Execute classifies it with IsUnauthorizedError — which matches
  both through errors.As and through a "status code: 401" substring check — so
  any wrapping lost the message and the user saw the generic "API key is invalid
  or expired" instead of the one thing that helps: that HOOKDECK_API_KEY takes a
  Project API key and a CLI client key belongs in --cli-key. Added an
  actionableError type that Execute checks before its 401 handling, so a command
  that has already explained the failure keeps its message.

- listen's help documented a precedence it no longer follows. It said stored
  credentials always beat HOOKDECK_API_KEY, which is untrue for a guest profile
  since the review fix. Help text and README now state the exception and that
  replacing a guest profile discards the sandbox link. REFERENCE.md regenerated.

- issue dismiss was the only one of the five destructive commands without a
  non-interactive test. The earlier justification for skipping it — that
  creating a real issue costs ~40s of setup — was wrong: runIssueDismissCmd
  confirms before it calls DismissIssue, so a non-existent ID exercises the
  guard in ~2s. All five callers are now covered.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RRp3XjWGqiYSun29SSNVJt

* test: cover the actionableError classification ordering

The actionableError type added in 570b71f changes the root error path but had no
test. The assertion that matters is the precondition: a wrapped 401 IS still
recognised by IsUnauthorizedError, so it is purely Execute's case ordering that
keeps the specific guidance. Writing that down means the day the precondition
stops holding, the test says so and the extra case can be removed rather than
lingering as cargo.

Also covers the other direction: an unmarked error must still receive the
generic recovery message, or every 401 would print raw API text instead of
telling the user how to sign in.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RRp3XjWGqiYSun29SSNVJt

* fix: identify the CLI to GitHub, and pin govulncheck

Two suppressed comments from the third Copilot pass. Neither is breaking today,
both are worth doing.

- The release check sent Go's default "Go-http-client/1.1". GitHub accepts it
  (verified: 200), but asks callers to identify themselves and applies rate
  limits per User-Agent — and go-github used to set one, so dropping it was an
  unintended regression from that removal. Now sends "hookdeck-cli/<version>".
  Built inline rather than via pkg/useragent, which imports pkg/version and
  would make this an import cycle.

- The govulncheck job installed @latest, so a new release could change behaviour
  or fail CI with no change to the repo. Pinned to v1.6.0. The vulnerability
  database is still fetched at run time, so newly disclosed issues are picked up
  without a version bump.

Verified end to end: a binary stamped 2.0.0 still detects v2.4.0 against the
real API, and govulncheck v1.6.0 reports no findings.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RRp3XjWGqiYSun29SSNVJt

---------

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

chore(deps): bump golang.org/x/crypto and build with Go 1.26.5 (#330)

* chore(deps): bump golang.org/x/crypto to v0.52.0

Clears all 13 open Dependabot alerts on the default branch: 7 critical,
2 high and 4 moderate, every one of them golang.org/x/crypto and every
one fixed in 0.52.0.

golang.org/x/text comes along as a transitive requirement of the new
version. Nothing else in go.mod moves.

x/crypto is an indirect dependency and `go mod why` reports that the
main module does not import it, so exposure was limited, but the alerts
are real and the bump is free.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HHJQ1QSdKmJMivqw7SER6t

* ci: build with Go 1.26.5

The pinned 1.24.9 toolchain carries standard-library vulnerabilities that
govulncheck flags as reachable from this code, including crypto/x509,
crypto/tls, net/http and net/textproto. release.yml builds the published
binaries, so the pin decides what ships to users.

Verified against both toolchains on the same tree:

  go1.26.1  10 vulnerabilities (1 module + standard library)
  go1.26.5   1 vulnerability  (1 module, no standard library)

The remaining one is GO-2026-5932, the unmaintained x/crypto/openpgp
package reached transitively through go-github. It is marked "Fixed in:
N/A" and no version bump resolves it.

Note the pin was already misleading: go.mod declares `go 1.25.0`, above
the pinned 1.24.9, so Go was auto-downloading a newer toolchain anyway.
test-homebrew-build.yml already derives its version from go.mod and is
left alone.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HHJQ1QSdKmJMivqw7SER6t

---------

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

fix(listen): populate team_id in dashboard deep-links for user-scoped…

… CLI keys (#320)

* fix(listen): populate team_id in dashboard deep-links for user-scoped CLI keys

User-scoped CLI keys are not tied to a single project, so cfg.ProjectID
was empty and the TUI rendered links like /events/cli?team_id= (and the
per-event open-in-dashboard action did the same).

Resolve the effective project id from the team that owns the fetched
connections when the profile has no active project, and use it for the
renderer and the compact/quiet-mode printer. The websocket config is
intentionally left on the profile value. As a final guard, the TUI link
builders now omit the team_id parameter entirely when the project id is
unknown instead of emitting an empty value.

Fixes #315

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tt8HTLV1iCoQyxH9vozKra

* refactor(listen): share deep-link construction; scope printer links in console mode

Review feedback on the compact/quiet printer: it linked the dashboard
root when the project id was unknown, and console mode dropped team_id
even when the id was known — leaving compact-mode console links
unscoped, the very bug this PR fixes for the TUI.

Move link construction into pkg/listen/links, used by both the TUI and
the printer, so all output modes produce identical URLs: /events/cli
(or the console base) with team_id appended only when known.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tt8HTLV1iCoQyxH9vozKra

---------

Co-authored-by: Claude <noreply@anthropic.com>

fix(release): run npm publish job on Node 22 so npm@latest installs (#…

…316)

The publish-npm job pinned Node 20 and then ran `npm install -g npm@latest`.
npm@12 dropped Node 20 support (requires Node ^22.22 || ^24.15 || >=26), so the
upgrade aborts with EBADENGINE before `npm publish` runs — which is why v2.3.0
built and released everywhere (GitHub artifacts, Homebrew, Scoop, Docker) but
was never published to npm. `npm publish --provenance` needs a recent npm, so
the fix is to run this job on Node 22.x rather than downgrade npm.


Claude-Session: https://claude.ai/code/session_016QCuRsTPjpi6hbu9WSaE5J

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

chore: v2.3.0 release prep — dependency bumps + MCP list filter parity (

#314)

* feat(mcp): add full list filter parity for events and requests

Expose payload search (body, headers, parsed_query, path), date range
after/before params, and remaining CLI list filters on hookdeck_events
and hookdeck_requests. Document ISO 8601 date windows and gte/lte mapping
for agents; add unit and MCP acceptance smoke tests.

Co-authored-by: Cursor <cursoragent@cursor.com>

* Update package.json version to 2.3.0-beta.1

* chore(deps): bump golang.org/x/sys from 0.45.0 to 0.47.0

Bumps [golang.org/x/sys](https://github.com/golang/sys) from 0.45.0 to 0.47.0.
- [Commits](golang/sys@v0.45.0...v0.47.0)

---
updated-dependencies:
- dependency-name: golang.org/x/sys
  dependency-version: 0.47.0
  dependency-type: direct:production
  update-type: version-update:semver-minor
...

Signed-off-by: dependabot[bot] <support@github.com>

* chore(deps): bump golang.org/x/term from 0.43.0 to 0.45.0

Bumps [golang.org/x/term](https://github.com/golang/term) from 0.43.0 to 0.45.0.
- [Commits](golang/term@v0.43.0...v0.45.0)

---
updated-dependencies:
- dependency-name: golang.org/x/term
  dependency-version: 0.45.0
  dependency-type: direct:production
  update-type: version-update:semver-minor
...

Signed-off-by: dependabot[bot] <support@github.com>

* chore(deps): bump github.com/modelcontextprotocol/go-sdk

Bumps [github.com/modelcontextprotocol/go-sdk](https://github.com/modelcontextprotocol/go-sdk) from 1.6.1 to 1.7.0.
- [Release notes](https://github.com/modelcontextprotocol/go-sdk/releases)
- [Commits](modelcontextprotocol/go-sdk@v1.6.1...v1.7.0)

---
updated-dependencies:
- dependency-name: github.com/modelcontextprotocol/go-sdk
  dependency-version: 1.7.0
  dependency-type: direct:production
  update-type: version-update:semver-minor
...

Signed-off-by: dependabot[bot] <support@github.com>

* docs(listen): surface --cli-key/--api-key on listen for ad-hoc auth

What: document that `listen` accepts a key directly — add an authentication
example to `hookdeck listen --help` and a section in the README listen docs
covering `--cli-key` (user-scoped CLI key) and `--api-key` (project-scoped key).

Why: passing a key to `listen` already works (both are global persistent
flags bound to the profile key), but the flags are hidden, so the capability
was undiscoverable. Users authenticating in CI or switching accounts had no
signal that `listen --cli-key <key>` is supported. A CLI key is user-scoped
and can navigate across projects; a project API/CI key is scoped to one
project — the docs now make that distinction explicit.

No behavior change: this is help-text and README only. Docs are hand-maintained
(no cobra doc-gen in the repo), so the README is updated directly.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016QCuRsTPjpi6hbu9WSaE5J

* docs: name the Event Gateway product correctly in the listen intro

The listen overview said "Hookdeck works by routing events…", attributing a
product capability to the company/brand. Hookdeck is the company; the Event
Gateway is the product that routes events. Reword so the product is named
correctly, and note that `hookdeck listen` is a standalone command that works
with whichever product you're authenticated with (Hookdeck Console or the
Event Gateway). Also tidy the duplicated `destination`, `i.e.`→`e.g.`, and
`Github`→`GitHub`.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016QCuRsTPjpi6hbu9WSaE5J

---------

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: github-actions <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

feat(mcp): add full list filter parity for events and requests

Expose payload search (body, headers, parsed_query, path), date range
after/before params, and remaining CLI list filters on hookdeck_events
and hookdeck_requests. Document ISO 8601 date windows and gte/lte mapping
for agents; add unit and MCP acceptance smoke tests.

Co-authored-by: Cursor <cursoragent@cursor.com>

chore: prep for homebrew-core submission (#296)

* chore(release): disable CGO for darwin builds in mac.yml

The codebase has no CGO usage (no `import "C"`, no `//#cgo` directives),
so the CGO_ENABLED=1 flag on the darwin amd64 build produced a libc-linked
binary for no benefit. The arm64 darwin build, all linux builds, and the
npm package builds (since commit fd2338b) already use CGO_ENABLED=0.

This change makes the GoReleaser-built tap binaries statically linked
and consistent with every other build path.

Refs #295

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* docs: regenerate REFERENCE.md for connection pause/unpause command

The `gateway connection pause` and `gateway connection unpause` commands
accept an ID or name (added in #276), but REFERENCE.md still documented
the older ID-only argument. Regenerated from cobra command metadata to
sync.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(completion): output completion script to stdout

Previously, `hookdeck completion --shell bash` wrote the completion script
to a file in the current working directory and printed multi-step shell
setup instructions. This was non-standard — every other major Cobra-based
CLI (gh, goreleaser, kubectl, helm, terraform) outputs the script to
stdout so it can be piped, redirected, or sourced directly.

Outputting to stdout enables the canonical pattern:

  source <(hookdeck completion --shell bash)

The Long help text now documents the redirection patterns; OS-specific
instruction blobs have been removed from the command output.

Also unblocks the idiomatic `generate_completions_from_executable` helper
in Homebrew formulae, which assumes stdout output.

BREAKING CHANGE: `hookdeck completion --shell <shell>` no longer creates
a file in the current directory or prints setup instructions. Use shell
redirection to write to a file. Release notes will call this out.

scripts/completions.sh updated to use redirection. README.md "Completion"
section rewritten to show the new pattern. REFERENCE.md regenerated.

Refs #295

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

refactor(config): sync cached API client from Profile instead of rese…

…tting

Replace exported ResetAPIClient with RefreshCachedAPIClient that updates the singleton in place (matches how MCP already mutates credentials on the shared *hookdeck.Client). Keep resetAPIClient unexported for tests only.

MCP login continues to assign client.APIKey/ProjectID explicitly: tests use a dedicated client pointer, not the global singleton.

Made-with: Cursor

fix(mcp): auth-aware gateway MCP, stdio stderr, login reauth (#280)

* fix(mcp): auth-aware gateway MCP, stderr for Execute, login reauth

- Skip requireGatewayProject for gateway mcp when no API key; enforce when key present
- Route gateway mcp errors and login-required messages to stderr; no interactive login
- argvContainsGatewayMCP skips global flags (e.g. --profile, -p)
- Always register hookdeck_login with optional reauth; clear credentials then device login
- Hedged list-projects failure hint suggesting reauth for 401/403-style errors
- ClearMCPProfileCredentials for memory-only or persisted config
- Acceptance: subprocess gateway mcp with stdin pipe; stdout JSON-RPC hygiene

Made-with: Cursor

* refactor(config): centralize silent credential clear for logout and MCP

- Rename ClearMCPProfileCredentials to ClearActiveProfileCredentials; document disk vs memory paths
- Logout uses the same helper as MCP reauth; zeroProfileCredentialFields helper
- RemoveAllProfiles clears in-memory credential fields after wiping the file (parity with logout -a)
- Rename config test to match new API

Made-with: Cursor

* Tighten error matching, improve login context handling, and tests (#281)

* fix(review): tighten error matching, add safety comments, revert noise

- tool_projects_errors.go: match "status code: 4xx" instead of bare
  "401"/"403" to avoid false positives on IDs or timestamps
- tool_login.go: remove `_ = ctx` suppression, add TODO for context
  propagation to polling goroutine, document happens-before on
  loginState.err via channel close
- root.go: add maintenance comment linking flagNeedsNextArg to init()
- helpers.go: revert unrelated Attempt struct alignment change

https://claude.ai/code/session_01EpXZqTmgybtjgmSALukH8d

* fix(review): ListProjects returns *APIError, add reauth + argv tests

- projects.go: use checkAndPrintError instead of manual status check so
  ListProjects errors are structured *APIError — the errors.As path in
  shouldSuggestReauthAfterListProjectsFailure now matches directly
- tool_projects_errors_test.go: unit tests for reauth hint logic covering
  APIError 401/403, plain error fallback, and false-positive resistance
- root_argv_test.go: document boolean-flag-between-subcommands limitation

https://claude.ai/code/session_01EpXZqTmgybtjgmSALukH8d

* fix(mcp): cancel login polling goroutine on MCP session close

Thread the request context into the login polling goroutine so it
stops promptly when the MCP transport closes, instead of running for
up to ~4 minutes after the client disconnects.

WaitForAPIKey blocks with time.Sleep and doesn't accept a context, so
we run it in an inner goroutine and select on both its result channel
and ctx.Done(). The inner goroutine is bounded by loginMaxAttempts.

https://claude.ai/code/session_01EpXZqTmgybtjgmSALukH8d

---------

Co-authored-by: Claude <noreply@anthropic.com>

* fix(mcp): review fixes for auth-aware gateway MCP (#282)

* fix(review): tighten error matching, add safety comments, revert noise

- tool_projects_errors.go: match "status code: 4xx" instead of bare
  "401"/"403" to avoid false positives on IDs or timestamps
- tool_login.go: remove `_ = ctx` suppression, add TODO for context
  propagation to polling goroutine, document happens-before on
  loginState.err via channel close
- root.go: add maintenance comment linking flagNeedsNextArg to init()
- helpers.go: revert unrelated Attempt struct alignment change

https://claude.ai/code/session_01EpXZqTmgybtjgmSALukH8d

* fix(mcp): cancel login polling goroutine on MCP session close

Thread the request context into the login polling goroutine so it
stops promptly when the MCP transport closes, instead of running for
up to ~4 minutes after the client disconnects.

WaitForAPIKey blocks with time.Sleep and doesn't accept a context, so
we run it in an inner goroutine and select on both its result channel
and ctx.Done(). The inner goroutine is bounded by loginMaxAttempts.

https://claude.ai/code/session_01EpXZqTmgybtjgmSALukH8d

* fix(mcp): use session-level context for login polling, not per-request

The per-request ctx passed to MCP tool handlers is cancelled when the
handler returns. The previous commit selected on that ctx in the login
polling goroutine, which killed the poll immediately after returning
the browser URL — breaking in-MCP authentication.

Fix: add a sessionCtx field to Server, set it in Run() (called by
RunStdio and tests), and select on that instead. The session context
is only cancelled when the MCP transport closes (stdin EOF), which is
the correct signal to abandon login polling.

Also adds TestLoginTool_PollSurvivesAcrossToolCalls: a regression test
that starts a login flow, lets the mock auth complete between tool
calls, and verifies the client is authenticated on the second call.
This would have caught the per-request ctx bug.

https://claude.ai/code/session_01EpXZqTmgybtjgmSALukH8d

---------

Co-authored-by: Claude <noreply@anthropic.com>

* fix(test): MCP login poll timing, profile save guard, retry API 500

- Wait loginPollInterval before second hookdeck_login in PollSurvives test
- SaveActiveProfileAfterLogin: skip disk persist when viper nil (tests)
- CLIRunner: retry transient HTTP 500 like 502 for acceptance flakes

Made-with: Cursor

* fix(cli): validate without stale project headers; sync profile from API

Omit X-Team-ID/X-Project-ID on GET /cli-auth/validate via clientForCLIAuthValidate
so a valid CLI key is not rejected when config.toml has a stale project_id.

After successful validate on the existing-key login path, persist project_id,
project_mode, and project_type from the response.

Add Profile Apply* helpers for validate, poll, and CI responses; use them from
login, interactive login, and MCP hookdeck_login. requireGatewayProject now
applies the full validate response (including project_id) without clearing
guest_url.

Tests: auth validate headers, profile apply helpers, LoadConfigFromFile,
gateway resolve-from-validate integration.

Made-with: Cursor

* docs(agents): go test in Cursor, module cache, out-of-sandbox runs

Document GOMODCACHE for agent shells, prompt-first elevated execution, and
point CLAUDE.md at AGENTS.md as the single source for agent instructions.

Made-with: Cursor

* chore: send command flag names in CLI telemetry

Include changed flag names (not values) as command_flags in
X-Hookdeck-CLI-Telemetry JSON. Wire CollectChangedFlagNames from root
initTelemetry. Add unit tests and acceptance test for login --api-key
and --cli-key over the recording proxy.

Made-with: Cursor

---------

Co-authored-by: Claude <noreply@anthropic.com>