Adds **Amazon Bedrock** as an opt-in built-in provider, over Bedrock's
unified **Converse / ConverseStream** API. One codec serves every
Converse-capable family — Claude, Amazon Nova, Meta Llama, Mistral,
DeepSeek, Moonshot Kimi, Z.AI GLM, MiniMax, NVIDIA Nemotron, and OpenAI
gpt-oss — because AWS translates the envelope to each model's native
dialect server-side. Auth is either **AWS SigV4** (the default
credential chain — env / profile / IMDS / IRSA / SSO, resolved per
request so sessions refresh) or a **Bedrock API key**
(`AWS_BEARER_TOKEN_BEDROCK`, bearer). Disabled by default (the Ollama /
OpenRouter opt-in pattern).
This is the redo of #459's original Claude-only `InvokeModel` adapter,
rebuilt on the gateway-refactor seams (#481–#497). @depopry's SigV4
signer, AWS event-stream frame decoder, `BedrockAuth`, the `aws_sigv4`
credential grammar, `AdapterKind::Bedrock`, region-from-base_url, and
the lean-deps decision are preserved and authored by him on the first
two commits; the per-family `BedrockCodec` trait he wrote turned out to
be the crate-wide `Codec` seam in miniature, so the refactor promoted
exactly that shape. The original Claude-only description is preserved in
a comment below.
## What's here
- **`AdapterKind::Bedrock` × `CodecKind::BedrockConverse`** on the
route, plus the `aws_sigv4` credential source (no static secret — the
adapter signs at request time; `fabro-auth` stays AWS-free).
*(@depopry)*
- **SigV4 signer + AWS event-stream `FrameDecoder`** on the lean AWS
stack (no `aws-sdk-bedrockruntime`; transport stays on `fabro-http`).
Re-targeted at Converse's direct-JSON stream frames; the signer resolves
credentials per request. *(@depopry)*
- **`bedrock_converse` codec** — Converse envelope (`system[]`, typed
content blocks, `inferenceConfig`, `toolConfig`), prompt caching via
`cachePoint`, thinking-signature round-trip through `reasoningContent`,
usage mapped onto the disjoint `TokenCounts` buckets,
`provider_options.bedrock` passthrough. Plus the adapter shell and an
event-stream byte loop beside the transport's shared SSE loop.
- **Catalog**: `bedrock.toml` (Claude incl. Fable 5, Nova 2, Llama 4,
Mistral, DeepSeek, Kimi, GLM, MiniMax, Nemotron, gpt-oss — cross-region
inference-profile ids, per-model `billing_policy` so Claude bills
Anthropic-style) and a companion **`bedrock-openai`** provider for
GPT-5.5/5.4 over the `bedrock-mantle` Responses endpoint (pure config
over the existing `openai_responses` codec, zero new code).
- Secrets registry (`AWS_BEARER_TOKEN_BEDROCK`), gitleaks rules for both
Bedrock key formats, the `docs/integrations/bedrock` guide, and live e2e
tests.
## Live verification (confirmed end-to-end against a real AWS account)
Verified on a real Bedrock account (us-east-2, SigV4 + bearer):
- **SigV4 + Converse** — multiple families (Claude, Nova, DeepSeek, …)
via the full settings → catalog → route → adapter → codec path.
- **ConverseStream** — streaming deltas through the workflow engine.
- **Multi-turn tool use** — agent loop with tool calls round-tripping
(no-arg tools included).
- **Multi-model routing** — Claude + DeepSeek pinned in one run through
the single Converse codec.
- **mantle Responses** — `openai.gpt-5.5` answered via the
`bedrock-openai` provider (bearer auth).
The exercise caught and fixed several issues that unit tests (static
creds, mocked transports) could not — see the follow-up commits below.
## Follow-up fixes from live testing (commits on top of the foundation)
1. **Worker AWS env** — the workflow worker scrubs its env to an
allowlist, so SigV4 (which re-resolves from the ambient chain per
request) couldn't work through `fabro run`. The AWS credential-chain
inputs now cross into the worker.
2. **Vault bearer key** — Bedrock was the only key-based provider
missing a `vault:` credential ref, so `fabro secret set
AWS_BEARER_TOKEN_BEDROCK` silently didn't feed it. Now resolves env →
vault → SigV4.
3. **Converse tool-encoding hardening** — a no-arg tool call's
`toolUse.input` is now a `{}` object (Bedrock rejects null), and every
tool `inputSchema` gets a top-level `type: "object"` (strict families
like DeepSeek reject a typeless schema Claude tolerates).
4. **Nova output cap** — `amazon.nova-2-lite` max_output 65536 → 65535
(Bedrock's per-request limit).
Earlier fixes already folded into the foundation commits: the
`aws-config` sleep-impl (default chain panicked) and AWS error-body
decoding (top-level `message`/`Message`/`__type` → proper messages
instead of "Unknown error").
## Manual testing & setup
See `docs/integrations/bedrock` — now documents the non-obvious account
setup that live testing surfaced: the per-Region Anthropic use-case
approval, `aws-marketplace:Subscribe` for third-party models, the Fable
5 / Mythos-class data-sharing opt-in, and the bearer-vs-SigV4 precedence
override for running Converse + mantle side by side.
## Open decision / discussion
- **Model-id naming** — Bedrock rows use dotted ids mirroring Bedrock's
native inference-profile ids (`us.anthropic.claude-sonnet-4-6`,
`openai.gpt-5.5`), which also makes them the wire `api_id`. Third scheme
alongside bare ids and OpenRouter's `vendor/model` slashes. No collision
risk (enforced at catalog build). Open to a uniform scheme if preferred.
- **`BEDROCK_API_KEY` alias** — see the comment thread; the AWS console
hands some users `export BEDROCK_API_KEY=` while the SDK-standard var is
`AWS_BEARER_TOKEN_BEDROCK`. Question of whether to accept both.
## Deferred (named follow-ups)
- **`qwen.qwen3-coder-next`** — omitted pending a verified Bedrock
model/inference-profile id (its fabro id isn't a valid Bedrock
identifier; needs an explicit `api_id`). Re-add once confirmed via `aws
bedrock list-inference-profiles`.
- **Claude Mythos 5** — Anthropic-Messages-only on `bedrock-mantle`
(limited preview).
- **Converse structured output** (`response_format` rejected with a
clear error).
- **`reasoning_effort` on Converse rows** via
`additionalModelRequestFields` (the `bedrock-openai` GPT rows already
accept effort levels).
- **CountTokens** route (`count_input_tokens` returns `None`).
## Verification
`cargo nextest run --workspace`: green except the pre-existing
environment-dependent fabro-workflow failures (identical on main).
clippy `-D warnings` + pinned-nightly fmt clean. Codec unit tests +
adapter httpmock tests + frame-decoder/signer locks.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: Scott Werner <scott@sublayer.com>
Co-authored-by: Scott Werner <stwerner@vt.edu>
Replay retryable stream failures from the last committed turn state, clear partial visible output before retry or terminal failure, and map Anthropic stream error events into structured provider errors.
## Summary
Moves provider-specific facts out of `AdapterKind` metadata and into
provider catalog data, leaving adapters responsible for runtime protocol
behavior. This makes providers that share an adapter mostly TOML-driven
while still surfacing adapter construction failures during readiness
checks.
## What Changed
- Provider TOML now owns auth mode, API-key/header policy, billing
policy, agent profile, base URLs/env overrides, extra headers, and probe
markers.
- Auth, install, config, diagnostics, and server flows resolve provider
credentials from catalog auth config, including API-key, header-only,
and no-auth providers.
- LLM client registration now reports adapter construction failures,
validates final adapter requests before HTTP dispatch, and preserves
custom primary auth headers.
- Billing and docs now use provider-owned billing policy instead of
adapter metadata, and the old adapter metadata surface is removed.
## Reviewer Notes
OpenAI-compatible `base_url` validation now happens during
adapter/client registration rather than catalog build. That keeps
catalog parsing adapter-agnostic while still letting readiness and model
listing reflect providers that cannot register.
## Verification
- `cargo check -p fabro-model -p fabro-auth -p fabro-llm -p fabro-server
-p fabro-cli`
- `cargo nextest run -p fabro-llm -- adapter_registry`
- `cargo nextest run -p fabro-model -- catalog`
- `cargo nextest run -p fabro-auth -- api_key`
- `cargo nextest run -p fabro-server -- install`
- `cargo +nightly-2026-04-14 fmt --check --all`
---
[](https://github.com/EveryInc/compound-engineering-plugin)
🤖 Generated with GPT-5 via [Codex](https://openai.com/codex)
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
## Summary
- allow `fabro exec` to use configured custom provider IDs from the
resolved LLM catalog
- route direct exec sessions through the same catalog-aware
provider/profile resolution used by workflow runs
- stop `fabro-client::list_models` from rejecting non-built-in provider
filters client-side
- update CLI snapshots and add regression tests for custom-provider exec
and model listing
## Repro
With a configured provider like:
```toml
[llm.providers.bedrock]
adapter = "openai_compatible"
base_url = "https://.../v1"
[cli.exec.model]
provider = "bedrock"
name = "bedrock-claude-sonnet-4-6"
```
these paths diverged:
- `fabro run ... --model bedrock-claude-sonnet-4-6` worked
- `fabro model list` showed `bedrock-*` models
- `fabro exec "..."` failed with `unknown provider: bedrock`
- `fabro model test --provider bedrock` failed with the same client-side
error
## Root cause
There were two separate built-in-only assumptions:
1. `fabro-agent` direct CLI paths parsed provider strings into the
built-in `Provider` enum and built a default catalog, so configured
provider IDs from `settings.toml` were invisible.
2. `fabro-client::list_models()` parsed the optional provider filter
into the same built-in enum before calling the server, so custom
provider filters never reached the API.
## Validation
- `cargo check -p fabro-cli -p fabro-agent -p fabro-client`
- `cargo test -p fabro-agent
resolve_provider_accepts_custom_catalog_provider -- --nocapture`
- `cargo test -p fabro-client list_models_allows_custom_provider_filters
-- --nocapture`
- `cargo test -p fabro-cli
exec_accepts_configured_custom_provider_from_settings -- --nocapture`
- `cargo test -p fabro-cli list_invalid_provider_errors -- --nocapture`
- `cargo test -p fabro-cli help -- --nocapture`
---------
Co-authored-by: Bryan Helmkamp <bryan@brynary.com>
Replace 16 inline copies of the `[cli.target] type = "http"` settings TOML across CLI integration tests with a single `set_http_target(&base_url)` method on `TestContext`. Removes a brittle format string that was maintained in ten files but only meaningfully asserted-against in one.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replace remaining expensive CLI lifecycle checks with seeded fixtures or focused unit coverage so the concurrent suite spends less time on duplicate full-process setup.
Persist dev-token credentials in auth.json alongside OAuth entries so CLI targets resolve credentials consistently across TCP and Unix socket flows.
Move install-time token minting to runtime storage, add auth login --dev-token, and refresh the embedded SPA after updating the stale dev-token hint.
Wrap root CLI errors at the main boundary so fatal diagnostics use miette's styled renderer while preserving existing telemetry, exit codes, and auth help hints.
Returns exit code 4 whenever the CLI fails because the user needs to run
fabro auth login, so scripts and the install wizard can distinguish
re-auth from generic failures.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Enable clippy::allow_attributes_without_reason at the workspace level.
Add concise, callsite-specific reasons to existing allow attributes, including generated code paths.
Phase 2/3 of the std::fs lint initiative (Phase 1 refactors landed in
commit 9d1c0d98c).
clippy.toml additions (appended to disallowed-methods):
std::fs::read, read_to_string, write, read_dir, copy, canonicalize
std::fs::File::open, File::create, File::create_new
std::fs::OpenOptions::open
File::options was deliberately excluded — it returns an OpenOptions
builder with no syscall. OpenOptions::open is where the block happens.
Non-blocking std::fs items (metadata, exists, create_dir_all, remove_*,
rename, and all std::fs types) remain legal.
Annotation policy (per updated plan):
- Mixed async/sync production source: function- or statement-scoped
#[expect(...)] so future accidental Tokio-path regressions in the
same file still fire.
- Fully-sync production source, test modules, integration tests,
build.rs: file-level #![expect(...)].
- Every #[expect] has a specific reason identifying the sync context.
Annotations added in ~90 files across the workspace. Notable narrow
placements: fabro-server server.rs current_server_target,
build_disk_usage_response, create_test_app_state_with_session_key;
fabro-server install.rs read_to_string rollback snapshot;
fabro-sandbox local.rs list_recursive; fabro-agent cli.rs FOLLOW-UP on
the JSON-stdout writer; fabro-llm providers/common.rs FOLLOW-UP for
load_file_as_base64 (7 translator call sites; revisit if file:// URL
usage grows).
build.rs blanket allows: fabro-api/build.rs, fabro-util/build.rs.
Pre-existing unrelated nightly-clippy warnings fixed under scope:
fabro-sandbox sandbox_spec.rs (unused_imports, unused_async),
reconnect.rs (unused_variables, unused_async).
Verified: cargo +nightly-2026-04-14 clippy --workspace --all-targets
-- -D warnings passes; fmt clean; 4129/4131 tests pass (two known
flakes under parallel nextest load, both pass individually and are
unrelated to this change).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Integration tests spawn the fabro binary as a subprocess and call
env_clear() for isolation, which strips LLVM_PROFILE_FILE. Under
cargo-llvm-cov this dropped subprocess coverage into orphaned
default.profraw files in tempdirs instead of the merged profile.
Add a preserve_coverage_env! macro in fabro-test and call it after
each env_clear() in apply_test_isolation, LightweightCli, and the
exec.rs sites. No-op when the env var is unset (normal test runs).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- rust.yml: move clippy to nightly-2026-04-14 (was stable); also pin
fmt to the same nightly date for consistency. Both jobs now use the
dated nightly and the run-step uses `cargo +nightly-2026-04-14 ...`.
- AGENTS.md: update developer commands to match CI.
- Duration constructors: replace `Duration::from_secs(N * 60)` /
`Duration::from_millis(N * 1000)` with `from_mins` / `from_secs` /
`from_hours` across the workspace to satisfy clippy's new
`duration_suboptimal_units` lint. std::time::Duration only — custom
`settings::duration::Duration` sites kept on `from_secs`.
- map/unwrap_or cleanup: `.map(f).unwrap_or(v)` → `.map_or(v, f)`,
`.map(f).unwrap_or(false)` on Result → `.is_ok_and(f)`, per
`clippy::map_unwrap_or`.
- Misc lints: collapse nested `if` into match guard in
handler/llm/api.rs and run_state.rs; replace `columns.len() > 0`
with `!columns.is_empty()`; switch a pair of `sort_by` calls to
`sort_by_key`.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add the shared fabro-http transport crate and route hand-written HTTP client construction through it.
Use FABRO_HTTP_PROXY_POLICY for test no-proxy defaults, remove direct reqwest deps from ordinary crates, and add clippy bans for raw reqwest entrypoints.
Collapse expensive CLI smoke coverage into scenario tests, replace the
slow doctor no-color integration check with a unit-level render test,
and remove duplicate attach coverage. Also fix local Unix-socket
autostart so missing daemons don't spend the full 5s readiness wait
before startup.
The commit includes the measured slow-test report updates for the work
landed here.
Consolidate CLI and server machine defaults under settings.toml,
including loader renames, writer preservation fixes, same-machine
manifest handling, and docs/test updates for the new config model.
Move secret storage, diagnostics, and repo/provider validation behind the
server API so credentials live under the server storage dir and take effect
immediately without process env mutation.
This also removes the old .env runtime path, rewires doctor/install/secret/
provider login/repo init around the server contract, and regenerates the
TypeScript client for the new endpoints.
Move --storage-dir and --server-url off GlobalArgs and onto the
leaf commands that actually honor them.
This aligns help, parser behavior, and env-var wiring with the
current command architecture while preserving the intended model
and exec targeting semantics.
The server subcommand and related code were gated behind
cfg(feature = "server"). This removes the feature flag entirely,
making fabro-server a required dependency so the server command
is always available.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add shared twin scenario helpers and use them to cover OpenAI-backed
CLI, agent parity, workflow, and exec integration paths. This brings the
worktree implementation back into the main checkout as a single commit.
Tests were flaky because command() inherited the real repo as the
working directory. When the repo was clean, detached runs attempted
git worktree creation against it, sometimes failing and injecting
extra warning lines into snapshots.
Now command() defaults to the non-git temp_dir, eliminating this
class of flakiness. Tests needing a specific directory override with
.current_dir(). Also canonicalizes fixture paths and adds a
[FIXTURES] snapshot filter via test_context!() macro.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace predicates::str::contains checks with full snapshots in
single-command tests: repo deinit failure, repo init help, secret
get/rm missing key, exec missing API key, config show missing workflow.
The remaining predicate usages are in multi-step CRUD tests and legacy
config tests where programmatic assertions are still the better fit.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace manual std::fs::create_dir_all + std::fs::write boilerplate
with context.write_temp() and context.write_home() in repo, workflow,
exec, and run tests.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace all ad-hoc arc()/fabro() command helpers with TestContext methods
(command(), run_cmd(), validate(), exec_cmd(), etc.) across all cmd/
test files. This eliminates 9 duplicate helper definitions and gives
every test consistent isolation (HOME, NO_COLOR, FABRO_STORAGE_DIR,
FABRO_NO_UPGRADE_CHECK).
Also removes init_cli_home() helper — TestContext's env-var-based
FABRO_STORAGE_DIR makes it unnecessary.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Redistribute all tests from the monolithic cli.rs (2190 lines) and the
standalone tests from scenario.rs into their respective cmd/ files,
completing the migration to the one-file-per-subcommand structure.
- Delete cli.rs entirely; move tests to cmd/{run,config,llm,exec,doctor,serve}.rs
- Move scenario.rs standalone tests to cmd/{repo,secret,workflow,doctor}.rs
- Slim scenario.rs to only 6 parametrized E2E workflow scenarios + run lifecycle
- Create new cmd/serve.rs and cmd/workflow.rs modules
- Remove duplicate tests already covered by snapshot tests
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adopt uv's testing pattern: a shared `fabro-test` crate with TestContext
and fabro_snapshot! macro, plus one test file per subcommand under
tests/it/cmd/. This replaces the trycmd-based tests which were hard to
read and didn't compose well with programmatic assertions.
- Create lib/crates/fabro-test with TestContext, run_and_format,
apply_filters, INSTA_FILTERS, and test_context!/fabro_snapshot! macros
- Add 42 snapshot tests across 16 subcommand files
- Delete trycmd.rs and all tests/cmd/ trycmd files
- Remove trycmd dependency, add fabro-test dev-dependency
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
