Addresses a Copilot review comment on #653.
The source array is runtime data, usually produced by a model, so its
length is not something a workflow author reviewed. Two changes, so an
over-long array degrades into a clear error rather than memory pressure.
Cap the item count at 1000. Above that the stage fails deterministically
before `parallel.started`, alongside the other for_each contract
violations, and the message says how to reduce the array.
Fork the parent context inside the branch task, after it acquires a
`max_parallel` slot, instead of at dispatch time. Live context copies now
track `max_parallel` rather than item count. Only the branch's own
preamble entry is moved into the task, so the shared stash is not cloned
per branch either.
The reviewer also suggested replacing spawn-all with `max_parallel`
workers pulling from a queue. Not done here: with the fork deferred, a
pending task holds little beyond its item, and reshaping the dispatch
loop would change cancellation and scope-reservation ordering, which
deserves its own review.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Restore the documented SDK env credential facade without reintroducing run fallback behavior. Fail closed on GitHub permission resolution, require worker storage at the CLI boundary, and align interpolation names and generated docs.
Conflict in apps/fabro-web/app/components/interview-dock.tsx. Main moved
the dock onto the shared collapsible `RunDockShell` and replaced the
local button constants with shared ones.
Kept main's structure whole and re-applied the review target rendering
onto it: the question paragraph in the shell's `body` becomes the linked
`ReviewTargetQuestion` when the target passes `safeReviewTarget`, and
plain text otherwise. Both now share main's paragraph classes through
`QUESTION_TEXT`, so the two renderings stay visually identical.
`peek` keeps using `question.text`, which is the correct plain-text
collapsed summary for a review target question.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The review question sentence was written in four places and the URL
safety rules in three. Collapse each to one definition.
- Add `ReviewTarget::question_text_with_link` as the single definition of
the question wording. `question_text()` and the Slack header both use
it, so a wording change is now one edit.
- Delete `ReviewTargetKind::noun()`. The enum already derives
`strum::Display` with the same snake_case output.
- Share one `review_target_line` helper between the console interviewer
and the CLI attach client, which held a byte-identical copy. Print only
the URL: `question.text` already carries the label and the noun.
- Trim the web-side check to the URL scheme, host, and credentials, which
are what a raw `href` can act on. Label length and control characters
cannot affect the DOM and stay server-side.
- Split validation from presentation in the web UI. `safeReviewTarget`
returns the target or null, and each caller picks its own fallback, so
an unsafe target now falls back to the same Markdown rendering as a
question with no target.
- Derive the resource noun from `kind` in the web UI instead of
hardcoding "document".
- Use `ReviewTargetKind.DOCUMENT` and the shared `isRecord` guard when
parsing events, instead of a raw string and a hand-rolled object check
that accepted arrays.
- Drop `deny_unknown_fields` from the wire struct. The OpenAPI schema
leaves `additionalProperties` permissive, so an added field would
otherwise make persisted events unreadable.
- Import `ReviewTarget` by name, and stop naming Slack in a fabro-types
error message.
- Document that `review_target=true` replaces the gate's `label`.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The process environment is no longer a configuration source. `{{ vars.NAME }}`
(non-sensitive, server-stored) and `{{ secrets.NAME }}` (vault-backed) cover
both cases, and reading the worker's ambient environment made a run's inputs
depend on how its process happened to be launched.
`Namespace::Env` is kept but wired to nothing, so `{{ env.NAME }}` still
parses and fails with a message naming its replacement rather than reaching
a consumer as literal text. `ResolveCtx::with_env` is gone, so no call site
can opt back in.
Two long-standing warts were env-only and go with it:
- `InterpString::resolve_or_source`, the "fall back to the raw template
source on failure" path, which let an unresolved token reach a sandbox or
the GitHub API as literal `{{ ... }}` text. Its own comment noted it was
slated for hard-error semantics.
- `RunEnvironmentSettings::resolve_env`'s matching source fallback for
env-only values.
Both carried `#[expect(clippy::disallowed_methods)]` escape hatches. Every
run-boundary resolver — sandbox env, prepare steps, MCP transports, GitHub
permissions, Slack channels, run goal files, provider extra_headers — now
fails closed instead.
Hooks lose their `allowed_env_vars` allowlist, `resolve_header`, and
`HeaderResolveError` along with the `E: Env` generic threaded through the
executor. They keep `{{ vars.* }}`, which `RunSettings::substitute_variables`
already substitutes server-side at run creation.
`allowed_env_vars` is removed from the OpenAPI spec and the generated
TypeScript client. The docs example showing `{{ env.* }}` in
`[server.slatedb.s3].bucket` was already wrong — that field is a plain
String and never interpolated — and is now a literal.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A node with no `shape` defaulted to `box`, which resolves to the agent
handler. That made a shapeless `script` node run as an LLM call prompted
with its own label, while the `script` was reported as inert — wrong
behavior behind a warning.
`script` is read by the command handler and by nothing else, so a
shapeless node that sets it is unambiguously a command node. `shape()`
now infers `parallelogram` in that case. An explicit `shape` still wins.
Two rules keep the inference honest:
- `script_prompt_conflict` — setting both `script` and `prompt` is an
error. No handler reads both. It fires regardless of shape so that
adding one cannot downgrade the error to a warning.
- `command_requires_script` — a command node without a script is an
error. Without this the original trap just moves: a node meant as a
command that omits its script silently becomes an agent again.
Also drops the `tool_command` alias in favor of `script` alone, routing
the six read sites through a new `Node::script()` accessor.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The DOT parser created a node for every edge endpoint, and nothing
recorded whether a node came from a declaration or was synthesized from
an edge. The edge_target_exists rule only checked whether the node id
was present in the graph, which was always true by then, so a misspelled
endpoint became an attribute-free node that defaulted to shape=box — an
LLM stage. Validation emitted a prompt_on_llm_nodes warning and exited 0.
Node now carries `implicit`, set only when the parser synthesizes the
node from an edge endpoint. A declaration anywhere in the workflow
clears it, so order does not matter and subgraph declarations count.
Node::new leaves it false, so programmatic construction and graphs
deserialized from older checkpoints read as declared.
edge_target_exists treats an endpoint as valid only when it exists and
is declared, reporting each undeclared node once. The near-identical
missing-source and missing-target branches collapse into one path. The
import transform copies the flag onto spliced nodes so an edge-only node
inside an imported fragment is caught too.
parse_and_validate_human_gate had two edge-only nodes and now declares
them; it was an instance of the bug rather than a casualty of the fix.
No shipped workflow, docs example, or CLI fixture relied on the old
behavior.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`ReadBeforeWriteSandbox` blocked writes to any existing file the agent
had not read, tracked by a session read set populated only by
`read_file`, `grep`, `read_many_files`, and the Kimi `Read`.
The gpt56 profile has none of those. It mirrors Codex's tool contract --
`shell_command`, `apply_patch`/`edit_file`, `update_plan`, `web_search`
-- and reads through the shell, so its read set stayed permanently
empty and every edit to an existing file failed. In run
01KYD4360GN6SED4BYEVGYP4XT all 28 `edit_file` calls failed, 25 of them
on the guard. The agent read `package.json` with `sed` and `cat`,
hex-dumped it trying to diagnose the rejections, then routed around the
guard with `sed -i`, which the guard never covered. It prevented no
blind write; it converted content-anchored edits into an unreviewed
in-place shell rewrite.
Neither Codex nor Kimi Code enforces read-before-write at runtime.
Codex's `apply_patch` `Add File` overwrites an existing path silently;
Kimi Code's `Write` has no check at all. Both rely on the exact-match
requirement in their edit tools, which is stronger proof of inspection
than a read set, plus per-write approval.
Tool descriptions and the Kimi prompt keep telling the model to read
before editing -- that guidance matches Kimi Code's own `edit.md` and
still prevents `old_string not found` -- but no longer claim the
workspace refuses unread writes.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Model default reasoning explicitly at the provider-route level so always-reasoning endpoints without effort controls receive summary headroom. Cap all summary requests at model output limits and bound retained visible summaries to the original allowance. Reuse builtin catalog fixtures and named budget constants in tests, and document the new model setting.
Fabro advertised Bash while its three backends implemented three
different contracts: Daytona evaluated commands through `sh`, and
Docker's streaming, stdio, and setup paths used a login shell. Bash-only
syntax silently misbehaved depending on provider and code path, and
login profiles could change PATH and command behavior per image.
Make `bash -c` the enforced interpreter for every command string the
Unix sandbox API accepts, on every production backend and through both
buffered and streaming execution. This selects the interpreter only —
no `errexit`, no `pipefail`, no login mode — so `false | true` still
succeeds and a workflow that wants other semantics writes them into its
own command.
Local resolves `bash` through the worker's PATH (NixOS has no
/bin/bash) and reuses that one executable across all three command
paths. Docker and Daytona require /bin/bash with no `sh` fallback.
Fresh initialization and resume/start now verify Bash through a shared
marker-validating probe before reporting the sandbox usable, so a
missing or non-Bash interpreter fails at the lifecycle boundary with
provider-specific remediation instead of on the first command. The
probe also rejects Bash in POSIX mode, which an image whose `bash` is
really `sh` would otherwise pass.
Sandbox MCP scripts and the detached launch wrapper move under the same
contract; host-side stdio MCP scripts, hooks, and interactive terminals
are separate executors and keep their existing `sh` behavior.
The `shell` tool's name and JSON schema are unchanged across providers;
only its prose now identifies `command` as Bash source.
BREAKING CHANGE: sandbox commands no longer load login-shell profiles,
so environment set in /etc/profile.d/*.sh, ~/.bash_profile, or
nvm/rbenv/sdkman initializers is gone. Move those exports into the
Dockerfile's ENV or the Daytona snapshot image.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Cumulative implement + simplify_fable diff recovered from the run's meta
branch (fabro/meta/01KY7YH7RYCJ1BDVTTP96ZA4HV, stage 006 diff.patch).
The run validated this tree clean: cargo nextest (7,007 passed), clippy,
fmt, TS client regen + typecheck, web tests (679 passed), docs check.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Document the contract of read_last_file_routing_json (terminal JSON
extraction only; routing validation happens downstream), extract a
shared sandbox_with_file test helper, and drop the misleading
"standalone" wording from the fallback docs.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Add two lint rules so the graph format stops silently accepting
attributes that nothing reads:
- inert_attribute: handler-specific attributes (script, language,
duration, join_policy, max_parallel, output_schema, prompt) placed on
node types that never read them. Attributes read by several handlers
(timeout), resolved for every node (fidelity, retry_policy), or
injectable via model stylesheets (model, reasoning_effort, ...) are
deliberately excluded.
- parallel_branch_inert_attribute: fidelity/thread_id on parallel
branch nodes and fork->branch edges. Branch dispatch bypasses the
fidelity lifecycle, so these are dead letters today; the warning
points at the parallel node, where fidelity does take effect.
Also reconcile the loop_restart docs with actual executor behavior:
taking a loop_restart edge restarts from the target with a fresh empty
context (visit counts preserved), on success as well as failure; the
transient_infra guard applies only to failure crossings.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Add Poolside as a built-in OpenAI-compatible provider and expose Laguna S 2.1 and XS 2.1 both directly and through OpenRouter. Include vault/env credential registration, secret redaction, live coverage, catalog tests, and user documentation.
## Summary
Provider `extra_headers` previously required values to be typed TOML
tables (`{ env = "X" }`, `{ literal = "Y" }`, `{ vault = "Z" }`). This
PR migrates them to the project's standard interpolation string format:
plain text for literals, `{{ env.NAME }}` tokens for environment
variables, and `{{ secrets.NAME }}` tokens for vault secrets. This
brings `extra_headers` in line with the rest of the interpolation system
and unlocks mixed-segment values like `Bearer {{ secrets.GATEWAY_TOKEN
}}`.
### What changed and why
**Config authoring surface (`fabro-config`):**
`ProviderSettings.extra_headers` changes from `Option<HashMap<String,
HeaderValueRef>>` to `Option<HashMap<String, InterpString>>`. The
`Combine` impl and all re-exports are updated accordingly.
**Catalog layer (`fabro-model`):**
`ProviderCatalogSettings.extra_headers` and
`CatalogProvider.extra_headers` become `HashMap<String, String>` — raw
interpolation source strings. This is required by the crate dependency
direction: `fabro-types` (which owns `InterpString`) depends on
`fabro-model`, so `fabro-model` cannot hold `InterpString` without
creating a cycle. The source string is re-parsed and resolved in
`fabro-auth` at credential-build time.
**Credential resolution (`fabro-auth`):** Both `CredentialResolver`
(vault-backed) and `EnvCredentialSource` (env-only) are rewritten to
parse each header source string as an `InterpString` and resolve it with
a `ResolveCtx` scoped to `env` + `secrets` only. A new
`resolve_extra_headers` helper is shared between the two paths. Vault
resolution uses `vault_token_lookup`, which wraps `vault_get_token` and
maps any non-Token vault entry to `None` — so file and OAuth vault
entries fail closed rather than resolving incorrectly. `vars.*` and
`inputs.*` tokens are not in scope and produce `Unavailable` errors
automatically.
**New error variant:** `ResolveError::Interpolation { provider, source
}` surfaces header resolution failures as diagnosable auth issues. The
inner `source` (an `InterpResolveError`) names only the token namespace
and name — never a resolved value.
**`{ literal = "..." }` guardrail removed:** `HeaderValueRef`
deliberately rejected bare string header values to discourage pasting
credentials. `InterpString` accepts any string. This is an intentional
change; the mitigation is documentation — use `{{ secrets.NAME }}` for
credential-shaped values, not bare literals.
**Redactor registration gap (noted, not fixed here):** Secrets resolved
into provider headers at the credential boundary do not flow through the
run boundary's exact-match redaction registry. Exposure is low (headers
are host-side and outbound-only, never logged), but a follow-up should
thread a registering lookup through `VaultCredentialSource`. A code
comment at the resolution site marks the gap.
### Breaking change
Existing `extra_headers` config using `{ env = "X" }`, `{ literal = "Y"
}`, or `{ vault = "Z" }` table syntax **will fail to parse** after this
change. Users must migrate to the token form: plain strings for
literals, `{{ env.X }}` for env vars, `{{ secrets.X }}` for vault
secrets. A changelog entry is included.
### Plan Summary
- Update `ProviderSettings.extra_headers` → `InterpString` in
`fabro-config`
- Collapse authoring `InterpString` → source `String` in
`provider_settings_to_catalog` (allowlisted `as_source()` call)
- Delete `HeaderValueRef` and its serde/display/parse machinery from
`fabro-model`
- Rewrite both auth resolution paths to use `InterpString::parse +
resolve_with`; add `Interpolation` error variant
- Add `vault_token_lookup` helper for token-only fail-closed vault
resolution
- Update test TOML in `fabro-llm`, builtin catalog comment in
`openrouter.toml`, and all hand-written + generated docs
### Fabro Details
<details>
<summary>Ran 8 stages in 108m 43s for $34.44</summary>
| Stage | Duration | Cost | Retries |
|---|---|---|---|
| start | 0s | – | 0 |
| toolchain | 1s | – | 0 |
| preflight_compile | 5m 33s | – | 0 |
| preflight_lint | 5m 59s | – | 0 |
| implement | 43m 23s | $17.34 | 0 |
| simplify_fable | 32m 49s | $13.09 | 0 |
| simplify_gpt | 6m 25s | $4.01 | 0 |
| verify | 13m 58s | – | 0 |
| **Total** | **108m 43s** | **$34.44** | **0** |
</details>
<details>
<summary>Ran <code>ImplementPlan.fabro</code> (11 nodes and 14
edges)</summary>
```dot
digraph ImplementPlan {
graph [
goal="Implement and simplify",
model_stylesheet="
* { model: claude-opus-4-8; }
"
]
rankdir=LR
start [shape=Mdiamond, label="Start"]
exit [shape=Msquare, label="Exit"]
toolchain [label="Toolchain", shape=parallelogram, script="command -v cargo >/dev/null || { curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y && sudo ln -sf $HOME/.cargo/bin/* /usr/local/bin/; }; cargo --version 2>&1", max_retries=0]
preflight_compile [label="Preflight Compile", shape=parallelogram, script="cargo check -q --workspace 2>&1", max_retries=0]
preflight_lint [label="Preflight Lint", shape=parallelogram, script="cargo +nightly-2026-04-14 clippy -q --workspace --all-targets -- -D warnings 2>&1", max_retries=0]
fix_lints [label="Fix Lints", prompt="The preflight lint step failed. Read the build output from context and fix all clippy lint warnings.", max_visits=3]
implement [label="Implement", prompt="Read the plan file referenced in the goal and implement every step. Make all the code changes described in the plan. Use red/green TDD. Be sure to use the rust-style-guide skill to help you follow this repo's Rust style conventions.", model="gpt-55", reasoning_effort="xhigh"]
simplify_fable [label="Simplify (Fable)", prompt="@prompts/simplify.md", model="claude-fable-5", reasoning_effort="xhigh"]
simplify_gpt [label="Simplify (GPT-55)", prompt="@prompts/simplify.md", model="gpt-55"]
verify [label="Verify", shape=parallelogram, timeout="1800s", script="git fetch origin main 2>&1 && git merge --no-edit --no-stat origin/main 2>&1 && cargo +nightly-2026-04-14 fmt --all 2>&1 && cargo dev docs refresh 2>&1 && cargo +nightly-2026-04-14 fmt --check --all 2>&1 && { command -v rg >/dev/null 2>&1 || { echo 'rg is required for verify'; exit 127; }; } && ! rg -n 'AuthMode::Disabled|RunAuthMethod|RunSubjectProvenance|\bActorRef\b|\bActorKind\b|AuthenticatedSubject|AuthenticatedService|AuthorizeRunScoped|AuthorizeRunBlob|AuthorizeStageArtifact|AuthorizeCommandLog|auth_method\s*==\s*\"disabled\"' lib/crates apps lib/packages docs/public/api-reference/fabro-api.yaml 2>&1 && cargo +nightly-2026-04-14 clippy --workspace --all-targets -- -D warnings 2>&1 && cargo nextest run --workspace --status-level slow --profile ci 2>&1 && cargo dev docs check 2>&1 && bun install --frozen-lockfile 2>&1 && (cd apps/fabro-web && bun run typecheck) 2>&1 && (cd apps/fabro-web && bun run test) 2>&1 && (cd lib/packages/fabro-api-client && bun run typecheck) 2>&1 && cargo dev build -- -p fabro-cli --release 2>&1", goal_gate=true, retry_target="fixup"]
fixup [label="Fixup", prompt="The verify step failed. Read the build output from context and fix all format, clippy, Rust test, docs, TypeScript typecheck/test, and build failures.", max_visits=3]
start -> toolchain
toolchain -> preflight_compile [condition="outcome=succeeded"]
toolchain -> exit
preflight_compile -> preflight_lint [condition="outcome=succeeded"]
preflight_compile -> exit
preflight_lint -> implement [condition="outcome=succeeded"]
preflight_lint -> fix_lints
fix_lints -> preflight_lint
implement -> simplify_fable -> simplify_gpt -> verify
verify -> exit [condition="outcome=succeeded"]
verify -> fixup
fixup -> verify
}
```
</details>
⚒️ Generated with [Fabro](https://fabro.sh)
---------
Co-authored-by: Fabro <noreply@fabro.sh>
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>
The first feature payoff of the gateway refactor series (#481–#496):
OpenRouter lands as **pure configuration over the `openai_compatible`
codec** — no new adapter, no new `AdapterKind`, no OpenRouter codec
fork. Redone from #438, which prototyped this pre-refactor as ~2,500
lines including a dedicated adapter and parallel codec plumbing; this
PR's fabro-llm diff is the usage-superset decode plus a TOML file.
## What's here (3 commits)
**Per-model `billing_policy` override (fabro-model)** — a model row may
override its provider's billing family: the aggregator case, where
Claude served through an OpenAI-compatible provider bills
Anthropic-style cache reads/writes. `pricing_for`/`billing_facts_for`
and the resolved `Route` read the model-effective policy; unknown
passthrough model ids keep the provider policy. Pinned by a pricing test
(cache writes bill at 1.25× input under the override, $0 under the
provider's OpenAI default).
**Aggregator usage superset in the `openai_compatible` codec** — the
wire usage struct gains tolerant optional fields:
- `prompt_tokens_details.cached_tokens` / `cache_write_tokens` and
`completion_tokens_details.reasoning_tokens` normalize into their
disjoint `TokenCounts` buckets with the same subtraction convention as
the `openai_responses` codec
- in-band `usage.cost` (OpenRouter returns it on every response)
surfaces as `Response.cost_usd` with `cost_source = authoritative`, on
both blocking and streamed responses — #494's client-side estimate
stamping already defers to it by construction
- **deliberate behavior change owned here**: compat providers that
report cached-token details now see them split out of `input_tokens`
(previously ignored — the wire pin placed in PR 0 anticipating exactly
this change flips, and two new OpenRouter-shaped wire pins land)
**The provider package** — `openrouter.toml` (disabled by default, the
Ollama opt-in pattern; curated vendor-namespaced model list; Claude rows
set `billing_policy = "anthropic"`; attribution headers deliberately not
sent unless the operator opts in via `extra_headers`),
`OPENROUTER_API_KEY` env/secret registry entries, a gitleaks rule for
`sk-or-v1-` keys, a live e2e test asserting authoritative cost, and docs
(integration guide + models concept + config reference).
## Deliberate scope cuts (fidelity follow-ups, per the plan)
- `reasoning_details[]` parse + verbatim multi-turn echo,
`cache_control` multipart emission, `provider`/`native_finish_reason`
field reads — the new wire pin proves they're tolerated and ignored
today
- Typed reasoning-param-style / routing codec params — no catalog row
can request reasoning effort yet (no `controls.reasoning_effort`
declared), and routing prefs already pass through
`provider_options.openrouter` verbatim via the existing
adapter-name-keyed merge; typed params land when an operator-level knob
actually needs them
- The OpenRouter Anthropic skin (`/api/v1/messages`) — a future pure
config row pairing the existing `anthropic_messages` codec with bearer
transport
## Verification
- `cargo nextest run --workspace --no-fail-fast`: 6724 passed; only the
known 5 pre-existing environment-dependent fabro-workflow failures
(identical on main)
- Wire snapshots: one deliberate flip
(`decode_usage_ignores_token_details` →
`decode_usage_parses_token_details`) + two new OpenRouter pins (blocking
cost/cache-write, streamed cost); all other snapshots unmodified
- clippy `-D warnings` + pinned-nightly fmt clean
- Builtin catalog unchanged for existing providers: OpenRouter is
`enabled = false`, so the #493 route-equivalence table is untouched
Credit to #438 for the provider research, catalog curation, gitleaks
rule, and docs structure.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
## Summary
Adds Anthropic Claude Fable 5 as a first-class Fabro model without
changing the default Anthropic model. The catalog now exposes
`claude-fable-5` with `fable` and `claude-fable` aliases, 1M context,
128k max output, effort levels, vision/tools, prompt caching, and the
documented pricing.
The Anthropic adapter now handles Fable's API behavior directly: it uses
the `claude-fable-5` API ID, omits the legacy 1M context beta header,
avoids injecting default `thinking`, preserves `output_config.effort`,
omits deprecated `temperature`/`top_p` sampling fields for Fable, and
rejects unsupported manual enabled/disabled thinking configs locally.
Fable refusals are converted into content-filter LLM errors with
`stop_details` preserved. Those refusal errors are fallback-eligible, so
existing `run.model.fallbacks` chains work for both prompt and agent
paths, while no-fallback refusals surface clearly as LLM errors.
## Live QA
Manually exercised the PR branch against a live Anthropic API key from
`~/.fabro.bak/.env.bak` using a temporary local harness that was removed
before commit. The run covered non-streaming completion via `fable`,
token counting via `claude-fable`, streaming completion, the deep
model-test path with tools/reasoning, local rejection of manual thinking
config, and a live refusal probe. The live run initially exposed
Anthropic's Fable rejection of `temperature`; this PR now strips
deprecated sampling fields for Fable and the live harness then passed
6/6 checks.
## Testing
- `cargo test -p fabro-llm --test live_fable_manual -- --nocapture
--test-threads=1` -> 6 passed against live Anthropic, temporary harness
removed afterward
- `cargo nextest run -p fabro-llm
encode_fable_uses_api_id_effort_and_omits_1m_beta`
- `cargo nextest run -p fabro-model -p fabro-llm -p fabro-workflow` ->
1808 passed, 41 skipped
- `cargo +nightly-2026-04-14 fmt --check --all`
- `cargo insta pending-snapshots` -> no pending snapshots
- `git diff --check`
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
