refactor: move claudear's http client onto abnegate-http

[abnegate]

Depends on #161 (Docker toolchain).

Build and deploy breaks

• The Docker image cannot build this until chore(docker): build the image on the toolchain the shared crates need #161 lands. abnegate-http 0.1.2 declares rust-version = "1.97", and the Dockerfile pins RUST_VERSION=1.93. CI uses dtolnay/rust-toolchain@stable and is unaffected.
• claudear::http, claudear_core::http and the claudear_integrations::github::{HttpClient, ReqwestHttpClient} re-export are gone. Nothing in the workspace used them after this change. Anything outside the workspace that imported them should import abnegate_http instead.

No env var or config key changes.

What changed

claudear-core's local HTTP client (crates/claudear-core/src/http.rs, 242 lines) is replaced by abnegate-http 0.1.2. The client is deleted, and its 22 consumers use abnegate_http::{HttpClient, HttpResponse, ReqwestHttpClient}. The migration plan split this into three steps (CLA1: core, CLA2: integrations, CLA3: analysis). They ship as this one PR because deleting the local client breaks all 22 importers at once, so separate PRs off main would not build. Each of the five commits builds on its own.

1. refactor(core): map abnegate-http errors into the app error
   • Adds abnegate-http = "0.1.2" as a workspace dependency.
   • Adds one explicit From<abnegate_http::Error> for claudear_core::Error in error.rs: three arms, with a test for each crate variant (see the table below).
   • Adds watcher tests that pin retry_trigger_error_is_transient: a JSON error and an oversized body do not give the retry back, and a real refused connection does.
   • Moves SentryHttpClient (in types.rs, which feat(analysis): so similar-issue suggestions reuse published trial memory #163 also edits) to the crate's HttpResponse. The edit is two lines. Its two implementers move in the same commit so the commit builds: source/sentry.rs, and the regression/sentry.rs test double.
   • Pins tokio-macros at 2.7.0, as refactor(analysis): discover dependencies through abnegate-vcs #168 does, so the lockfile gains no syn 3.
2. refactor(integrations): move the HTTP consumers onto abnegate-http
   • Moves the 15 integrations consumers: github.rs, gitlab.rs, deploy_qa.rs, discord/{client,thread_manager}.rs, notifier/{discord,slack,sms,telegram,whatsapp}.rs and source/{github,gitlab,helpscout,jira,linear}.rs.
   • Builds every response with HttpResponse::new, because the crate struct is non_exhaustive.
   • The HttpClient test doubles now answer with abnegate_http::Result.
   • GitHub and GitLab clients read bodies up to scm::BODY_LIMIT (64 MiB).
3. refactor(analysis): move the HTTP consumers onto abnegate-http
   • Covers deploy_qa/tracker.rs, regression/{linear,sentry}.rs and release/{github,tracker}.rs.
   • claudear-analysis now names reqwest directly, for the constructor fallback described below.
   • MockErrorHttpClient fails with a real refused connection (127.0.0.1:1), converted into abnegate_http::Error::Request. It makes the connection through a client it builds once, without a proxy.
   • Adds release::BODY_LIMIT = 8 * ReqwestHttpClient::DEFAULT_BODY_LIMIT (64 MiB).
     • ReleaseClient::new applies it.
     • scm::BODY_LIMIT in claudear-integrations re-exports it, so the GitHub, GitLab and release clients share one value. claudear-analysis cannot depend on claudear-integrations, but integrations already depends on analysis.
4. refactor(core): delete the local http client
   • Removes http.rs, pub mod http and the src/lib.rs re-export.
5. test(http): pin the body limits by what the clients read, not by debug output (review follow-up)
   • GitHubClient::new, GitLabClient::new and ReleaseClient::new build the default transport and hand it to a private with_transport, which applies BODY_LIMIT. Production behaviour is unchanged.
   • The tests that read body_limit: … out of ReqwestHttpClient's derived Debug are gone. Each client's limit is now tested by what it reads from a loopback server (see "New tests").
   • The loopback helpers (serve_once, ok_response, loopback_transport) move from claudear-integrations/src/test_support.rs to claudear-analysis/src/test_support.rs, compiled under #[cfg(any(test, feature = "test-support"))]. claudear-integrations turns the feature on for its dev-dependency only, so the release client and the SCM clients share one server. loopback_transport() builds the reqwest client with no_proxy() and leaves the crate's default limit, so the limit under test comes from with_transport.

Error mapping (claudear-core/src/error.rs, one place)

The impl has three arms: Request | UnreadableBody → Http, TooManyRedirects → Network, and everything else → Other(error.to_string()). The tests pin each row:

abnegate_http::Error	claudear_core::Error	Retried by the watcher	Before
Request (transport)	Http(reqwest::Error)	yes	Http, same
UnreadableBody	Http(reqwest::Error)	yes	an empty successful body
TooManyRedirects	Network	yes	Http (only PublicClient raises this variant; the trusted transport still reports redirect loops as Request)
Json	Other("JSON parse error: …")	no	Other("JSON parse error: …"), same
Unsupported(method)	Other("POST is not supported by this HTTP client")	no	Other("POST not supported by this HTTP client")
OversizedBody	Other	no	no cap
InvalidUrl, InvalidHeaderName, InvalidHeaderValue, UnsupportedScheme, EmbeddedCredentials, MissingHost, PrivateAddress, InternalHost, UnfetchableResolution	Other	no	not raised: they come only from PublicClient URL checks, which claudear does not use
a variant added in a later crate release	Other	no	n/a (the crate enum is non_exhaustive)

Json maps to Other instead of Error::Json so the variant and message stay exactly what the local HttpResponse::json produced. A blanket mapping into Http would have made a malformed body look transient.

Constructing the client

The crate's ReqwestHttpClient::new() returns a Result, and the crate has no Default.

• GitHubClient::new, GitLabClient::new, HelpScoutSource::new, ReleaseClient::new and LinearRegressionChecker::new stay infallible. When the configured client fails to build, they fall back to ReqwestHttpClient::from(reqwest::Client::new()). The local ReqwestHttpClient::new did exactly the same.
• LinearRegressionChecker::with_embeddings already returned a Result, so it now builds the client first and propagates the build error before it spends an embedding call.
• Every site uses the trusted transport, never PublicClient, because self-hosted GitLab, Sentry and Jira run on private addresses.

Behaviour differences

These apply to requests that go through the crate's ReqwestHttpClient:

• Integrations: GitHub, GitLab and HelpScout.
• Analysis: the release client and tracker, the deploy QA tracker and the Linear regression checker.

The Discord, Slack, SMS, Telegram, WhatsApp, Jira, Linear-source and Sentry clients make their own reqwest calls. They only build the crate's HttpResponse, so their transport behaviour does not change.

• Body cap. Bodies were read whole; the crate caps them at 8 MiB (DEFAULT_BODY_LIMIT).
  • The GitHub and GitLab clients fetch PR diffs, so they read up to BODY_LIMIT, which is 64 MiB (eight times the crate default).
  • ReleaseClient also reads up to BODY_LIMIT. It is shared by the release tracker and the deploy QA tracker.
    • is_commit_in_release fetches compare/{commit}...{tag}, which carries up to 250 commits and 300 files with patches.
    • Under the 8 MiB default, an oversized comparison would be a non-retried Error::Other, and the release watch would fail on every poll.
  • A body over the cap is now an error (Error::Other, not retried) instead of a success.
  • HelpScout and the Linear regression checker keep the 8 MiB default. Their responses are paginated listings, Linear GraphQL pages and GitHub issue searches.
  • For PR diffs, get_pr_diff errors on a diff over 64 MiB, and the watcher logs "Failed to fetch PR diff for analysis" at debug level and skips the diff analysis.
• Unreadable bodies. A body that breaks off was returned as an empty successful body; it is now Error::Http.
  • Calls that parse JSON already failed in that case, with a non-retried JSON error. They now fail with a retried Http error.
  • Calls that only check the status now report an error where they reported success. These are merge_pr, close_pr, delete_branch, close_issue and their GitLab counterparts, after a response whose body breaks off.
• Decoding. Bodies decode as UTF-8 with invalid sequences replaced. reqwest's .text() honoured a charset from Content-Type. Every API involved sends UTF-8.
• Error text. Transport errors no longer carry the request URL, because the crate strips it so query-string secrets stay out of logs. The unsupported-method message reads "POST is not supported by this HTTP client".
• Unchanged. The 30s timeout and 10s connect timeout, reqwest's 10-hop redirect policy, proxy environment handling and the absent user agent.
• Dependency graph.
  • abnegate-http needs tokio 1.53, so tokio moves 1.50.0 → 1.53.2. tokio-macros moves 2.6.0 → 2.7.0, pinned there as in refactor(analysis): discover dependencies through abnegate-vcs #168 because 2.7.2 pulls in syn 3.
  • Patch bumps: mio 1.1.1 → 1.2.4, socket2 0.6.2 → 0.6.5, once_cell 1.21.3 → 1.21.4.
  • abnegate-http 0.1.2 and dashmap 6.2.1 are added.
  • abnegate-http turns on reqwest's multipart and query features workspace-wide.

Crate gaps (follow-ups in abnegate/crates, no shim added here)

• (a) No Default and no infallible constructor for ReqwestHttpClient. Five constructors repeat ReqwestHttpClient::new().unwrap_or_else(|_| ReqwestHttpClient::from(reqwest::Client::new())). A crate Default would collapse each one to ReqwestHttpClient::default().
• (b) No convenience constructor for a transport failure. A real Request error is constructible: the regression/linear.rs MockErrorHttpClient (which used to return Error::network("connection refused")) now makes a real refused connection to 127.0.0.1:1 and converts the reqwest error with abnegate_http::Error::from, which maps to the same retried Error::Http. That costs a loopback connect per call. A constructor such as Error::transport(message) would let a test double fail without touching the network.

How it was verified

Review follow-up (aad70a3, on top of the four commits below; no rebase was needed, main is still 95fab34). claudear-verify.sh <worktree> http ready passed: cargo fmt --all -- --check, cargo clippy --all-targets -- -D warnings -A clippy::double_must_use, and cargo test -p claudear-core -p claudear-config -p claudear-storage -p claudear-analysis -p claudear-integrations -p claudear-engine -p claudear (default features, which include sqlite): 10,473 passed, 0 failed, 3 ignored. The red runs for the new limit tests are described under "New tests". This machine's Homebrew now ships OpenSSL 4, which llama.cpp's vendored cpp-httplib does not compile against, so these local runs set CMAKE_DISABLE_FIND_PACKAGE_OpenSSL=TRUE for the llama-cpp-sys-2 build. That only drops HTTPS from llama.cpp's own downloader, which claudear does not use; CI is unaffected.

The original verification of the four migration commits:

All commands ran in the worktree with CARGO_BUILD_BUILD_DIR=/Users/jakebarnby/.cargo/build/shared/claudear-migrate CARGO_INCREMENTAL=0 CARGO_PROFILE_DEV_DEBUG=0 CARGO_PROFILE_TEST_DEBUG=0, default features plus sqlite for tests. The fastembed model came from a local cache copy, and vectorlite from /opt/homebrew/lib/vectorlite.dylib.

• cargo check --workspace --all-targets passes on each of the four commits on its own.
• cargo fmt --all -- --check passes.
• cargo clippy --workspace --all-targets -- -D warnings -A clippy::double_must_use passes.
• cargo test -p claudear-core -p claudear-integrations -p claudear-analysis -p claudear-engine -p claudear --features sqlite passes (exit 0).
  • 9,353 passed, 0 failed, 3 ignored (the ignored tests were already ignored in integrations).
  • By target: claudear lib 315, claudear bin 11, tests/e2e_real_repo.rs 2, tests/retries.rs 1, tests/shutdown.rs 13, claudear-analysis 2,885, claudear-core 452, claudear-engine 1,367, claudear-integrations 4,305, other targets 2.
  • No doc tests ran.
• Every test that touches a socket (the loopback server, the refused connections) builds its client with no_proxy(), so an HTTP_PROXY/HTTPS_PROXY in the environment cannot reroute it. I re-ran those tests with HTTP_PROXY, HTTPS_PROXY and ALL_PROXY (upper and lower case) pointing at a dead proxy on 127.0.0.1:9. All 142 selected tests passed: core 7, engine 3, integrations 7, analysis 125 (the regression::linear module, including every MockErrorHttpClient test).

New tests:

• claudear-core error.rs: one test per crate variant, including that a JSON error keeps the "JSON parse error: " text. The transport variants use a real refused connection.
• claudear-engine watcher.rs: retry_trigger_error_is_transient is false for a JSON error and for an oversized body, and true for a real refused connection.
• github.rs, gitlab.rs and release/github.rs: each client is built through the same with_transport that new() uses, with a loopback-only reqwest client, and reads from a loopback server.
  • A body of 8 MiB + 1 byte (one byte over the crate default) is read whole.
  • A declared length of BODY_LIMIT + 1 is refused as OversizedBody { limit: BODY_LIMIT }.
  • github.rs also checks that a body cut short becomes Error::Http.
  • Seen failing. With with_transport temporarily applying BODY_LIMIT / 8 (the crate default, as if the limit were removed), all six limit tests failed: the reads over 8 MiB were refused as OversizedBody { limit: 8388608 }, and the refusals reported the wrong limit. With it applying BODY_LIMIT * 2 (a raised limit), the three refusal tests failed, because the declared length was accepted and the read then failed as unreadable instead. Both mutations were reverted before the commit.

Clean-ups in touched files

• Section header comments (// ---, // ===, // ═══ banners) are removed from every file this PR touches, with two exceptions:
  • crates/claudear-core/src/types.rs belongs to feat(analysis): so similar-issue suggestions reuse published trial memory #163; this PR keeps to its two-line edit there.
  • crates/claudear-integrations/src/notifier/telegram.rs keeps its headers because a follow-up PR (Telegram HTML escaping) stacks on this branch and edits that file.
  • None of the removed headers sit in or next to a hunk of an open PR (checked against each overlapping PR's patch).
• Abbreviated names in and next to the changed lines are spelled out:
  • k/v in the GitHub and GitLab doubles;
  • tok/resp/r/e throughout HelpScout's source (feat(mcp): fuzzy channel names, HelpScout status filter, result-cap notes #166 has merged, so nothing collides);
  • resp_body in WhatsApp, now text so it no longer shadows the body parameter;
  • idx and r in the test doubles.
• The narrating comments in the GitHub doubles and the Sentry status-boundary test are gone.

Not verified

• --all-features (CUDA) was not built locally.
• The Docker image was not built; it needs chore(docker): build the image on the toolchain the shared crates need #161.
• No request was made against live GitHub, GitLab, HelpScout or Linear APIs. Transport behaviour is pinned by the loopback tests above and by the crate's own tests.

Overlapping open PRs

From gh pr view <n> --repo appwrite/claudear --json files, intersected with this PR's files:

• feat(analysis): so similar-issue suggestions reuse published trial memory #163: Cargo.toml, Cargo.lock, crates/claudear-analysis/Cargo.toml, crates/claudear-core/Cargo.toml, crates/claudear-core/src/types.rs (two-line SentryHttpClient return-type edit).
• refactor(analysis): discover dependencies through abnegate-vcs #168: Cargo.toml, Cargo.lock, crates/claudear-analysis/Cargo.toml, crates/claudear-engine/src/watcher.rs. Both PRs move tokio to 1.53.2 with tokio-macros 2.7.0.
• refactor(config): write .env secrets through abnegate-config #167: Cargo.toml, Cargo.lock, src/lib.rs.
• feat(watcher): per-source workers + global Claude session ceiling #170: crates/claudear-engine/src/watcher.rs.
• feat(reports): support threads digest and reviewed answer drafts #162: crates/claudear-engine/src/watcher.rs, crates/claudear-integrations/src/notifier/discord.rs.
• feat(agent): give fix agents live Grafana telemetry via MCP #147: crates/claudear-integrations/src/source/sentry.rs.
• Add Windows support for IPC with TCP and port file discovery #45 (stale):
  • Cargo.toml, Cargo.lock, crates/claudear-core/Cargo.toml, crates/claudear-core/src/lib.rs, crates/claudear-core/src/http.rs (deleted here);
  • crates/claudear-engine/Cargo.toml, crates/claudear-engine/src/watcher.rs;
  • crates/claudear-integrations/src/discord/client.rs, notifier/{discord,slack,sms,telegram,whatsapp}.rs, source/{jira,linear,sentry}.rs;
  • src/lib.rs.
• chore(docker): build the image on the toolchain the shared crates need #161: no shared files; it is the Docker toolchain this change needs.
• Dependabot:
  • chore(deps): bump the rust-minor group across 1 directory with 33 updates #169, chore(deps): bump lettre from 0.11.19 to 0.11.22 #137, chore(deps): bump quinn-proto from 0.11.14 to 0.11.15 #130, chore(deps): bump openssl from 0.10.76 to 0.10.80 #88, chore(deps): bump rustls-webpki from 0.103.10 to 0.103.13 #81 and chore(deps): bump rand from 0.10.0 to 0.10.1 #78 touch Cargo.lock.
  • chore(deps): bump sentry from 0.47.0 to 0.48.2 #101, chore(deps): bump sysinfo from 0.38.4 to 0.39.3 #100, chore(deps): bump rusqlite from 0.39.0 to 0.40.1 #99, chore(deps): bump sha1 from 0.10.6 to 0.11.0 #98, chore(deps): bump sha2 from 0.10.9 to 0.11.0 #75 and chore(deps): bump hmac from 0.12.1 to 0.13.0 #74 touch Cargo.toml and Cargo.lock.
• Follow-up: a PR fixing Telegram HTML escaping stacks on this branch. notifier/telegram.rs changes here only in its imports and HttpResponse construction.

Rebased on #154, #160 and #166.

🤖 Generated with Claude Code

[hansi-codes]

🟡 Tier B · Needs changes before merging

Limited by an open major finding: Update the Docker toolchain before adding this dependency

Replaces claudear’s local HTTP client with abnegate-http across core, integrations, and analysis, including explicit application error mapping and shared response-body limits for SCM and release requests. Updates the HTTP consumers, test doubles, dependency configuration, and retry-classification tests while removing the former HTTP module and re-exports.

Latest changes: The follow-up replaces Debug-based body-limit assertions with loopback behavior tests, moves the shared loopback helpers into analysis behind a test-support feature, and routes the GitHub, GitLab, and release clients through private transport constructors that apply their limits.

Verdict	New comments	Fixed	Still open
💬 Commented	0	0	1

Fix with agent prompt

### Issue 1
Cargo.toml:154
**Update the Docker toolchain before adding this dependency**

abnegate-http 0.1.2 requires Rust 1.97, but Dockerfile:2 still pins Rust 1.93, so the default image build fails at Cargo's rust-version check. Please land the toolchain update first or include it here rather than merging a dependency that the deployment build cannot compile.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

📂 Walkthrough · 6

File	Change
Cargo.toml, Cargo.lock, crates/claudear-{core,analysis,engine,integrations}/Cargo.toml	Adds abnegate-http and updates dependency and test-support feature configuration.
crates/claudear-core/src/{error.rs,http.rs,lib.rs,types.rs}	Maps abnegate-http errors, updates the Sentry response type, and removes the local HTTP client module and exports.
crates/claudear-engine/src/watcher.rs	Adds retry-classification coverage for HTTP, JSON, and oversized-body errors.
crates/claudear-analysis/src/{deploy_qa/tracker.rs,regression/linear.rs,regression/sentry.rs,release/github.rs,release/mod.rs,release/tracker.rs,test_support.rs,lib.rs}	Migrates analysis HTTP clients and test doubles, sets the release body limit, and adds shared loopback test support.
crates/claudear-integrations/src/{deploy_qa.rs,discord/client.rs,discord/thread_manager.rs,github.rs,gitlab.rs,notifier/discord.rs,notifier/slack.rs,notifier/sms.rs,notifier/telegram.rs,notifier/whatsapp.rs,scm.rs,source/github.rs,source/gitlab.rs,source/helpscout.rs,source/jira.rs,source/linear.rs,source/sentry.rs}	Migrates integration consumers and test doubles to abnegate-http, applies shared SCM body limits, and adopts the analysis loopback test helpers.
src/lib.rs	Removes the former HTTP re-export.

⏳ Still open from earlier reviews · 1

• 🟠 Update the Docker toolchain before adding this dependency · Cargo.toml:154

---

_{Reviewed the commits since 1c13990 · Details · Comment @hansi-codes review to re-run, or mention @hansi-codes with a question.}

[hansi-codes]

🟡 Tier B · 1 blocking finding to address. Summary

[hansi-codes]

Update the Docker toolchain before adding this dependency

abnegate-http 0.1.2 requires Rust 1.97, but Dockerfile:2 still pins Rust 1.93, so the default image build fails at Cargo's rust-version check. Please land the toolchain update first or include it here rather than merging a dependency that the deployment build cannot compile.

Prompt To Fix With AI

This is a comment left during a code review.
Path: Cargo.toml
Line: 154

Comment:
**Update the Docker toolchain before adding this dependency**

abnegate-http 0.1.2 requires Rust 1.97, but Dockerfile:2 still pins Rust 1.93, so the default image build fails at Cargo's rust-version check. Please land the toolchain update first or include it here rather than merging a dependency that the deployment build cannot compile.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

_{🟠 Major · bug · Reply if this doesn't apply.}

[abnegate]

Agreed, the image cannot build this until the toolchain moves. That change lives in #161, which sets RUST_VERSION=1.98 in the Dockerfile (abnegate-http 0.1.2 needs 1.97). This PR depends on #161 and must merge after it; the body now says "Depends on #161 (Docker toolchain)" at the top and keeps the build-break note. I kept the toolchain bump out of this PR so every crate migration shares one Dockerfile change rather than each carrying its own.

[hansi-codes]

This is still an outstanding build break in this PR’s current tree: the Dockerfile pins Rust 1.93 and uses that version for the builder (Dockerfile, builder stage), while the new dependency is declared here (Cargo.toml). Calling #161 a dependency documents the required merge order, but does not make this tree buildable. Please keep this blocked until #161 lands and this PR is based on the toolchain update, or include that update here; I’ll leave the finding open until then.