Add a fabro-llm input token counting API with provider-native counting for Anthropic, Gemini, and OpenAI. Include strict fallback behavior, deterministic local estimates, count-specific request filtering, and README documentation for privacy and billing semantics.
Use raw sandbox reads for memory and skills, keep line-numbered reads focused on display, and share retry-delay handling across agent and LLM code.
Trim task tool descriptions, bound multi-file read concurrency, restore Docker's text read path, and add the reviewed implementation plan docs.
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
OpenAI-profile agents now receive `apply_patch` as a Codex-compatible
freeform custom tool instead of a JSON function with a `patch` field.
The model sends raw patch text validated by the vendored Lark grammar,
and Fabro round-trips OpenAI custom tool calls/results through the
Responses API.
## What Changed
- Added `fabro-llm` support for function and custom tool definitions
while preserving existing JSON function behavior for normal tools.
- Translated OpenAI custom tool calls, custom tool outputs, and
streaming `response.custom_tool_call_input.delta` events into Fabro tool
calls with raw arguments.
- Ported the Codex apply-patch grammar and adapted Codex-style patch
parsing/application semantics to Fabro's `Sandbox` trait, including
strict envelopes, fuzzy context matching, move/delete/add/update
behavior, trailing-newline normalization, and Codex-style
summaries/errors.
- Updated OpenAI agent prompt/tool registration so `apply_patch` is
freeform, and skipped JSON-schema validation/repair for custom tool
calls only.
- Updated file tracking to read Codex-style `A`/`M` result lines from
successful patch output.
## Verification
- `cargo nextest run -p fabro-llm -p fabro-agent`
- `cargo +nightly-2026-04-14 fmt --check --all`
- `cargo +nightly-2026-04-14 clippy --no-deps -p fabro-llm -p
fabro-agent --all-targets -- -D warnings`
- `git diff --check`
Note: the full non-`--no-deps` clippy command still surfaces an
unrelated existing `fabro-sandbox` `large_enum_variant` warning in
`lib/crates/fabro-sandbox/src/sandbox_spec.rs`.
---
[](https://github.com/EveryInc/compound-engineering-plugin)
🤖 Generated with GPT-5 via [Codex](https://openai.com/codex)
## 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
Prepares the model catalog tests for catalog-only built-in providers so
a future provider can be added with just its catalog TOML.
## Changes
- Replaces the closed-enum round-trip guardrail with a catalog metadata
guardrail, allowing built-in TOML providers that do not have `Provider`
enum variants.
- Makes the all-model `fabro model` CLI tests assert stable table
structure instead of snapshotting every built-in catalog row.
- Renames synthetic custom-provider and missing-provider fixtures away
from provider names that can become real catalog entries.
## Verification
- `cargo +nightly-2026-04-14 fmt --check --all`
- `cargo nextest run -p fabro-model -p fabro-auth -p fabro-llm -p
fabro-server -p fabro-workflow -p fabro-config`
- `cargo nextest run -p fabro-cli cmd::model`
---
[](https://github.com/EveryInc/compound-engineering-plugin)
🤖 Generated with GPT-5 (unknown context, medium reasoning) via
[Codex](https://openai.com/codex)
---------
Co-authored-by: Jesse <606+jesseproudman@users.noreply.github.com>
## Summary
Finish phase 9 of the configurable LLM provider/model work by aligning
public docs, release notes, and guardrails with the implementation
already landed in phases 0-8.
- documents settings-driven providers/models, OpenAI-compatible gateway
examples, typed `extra_headers`, model `api_id`, controls, and per-speed
costs
- adds the 2026-05-13 changelog entry and provider string migration note
- updates the internal phase plan ledger to reflect current
implementation status
- adds a workspace policy test blocking direct production
`Catalog::builtin()` usage outside catalog owner/test code
- clarifies `Provider` as a built-in compatibility enum while open-ended
identity is `ProviderId`
## Verification
- `cargo nextest run -p fabro-dev --features dev --test it policy`
- `cargo dev docs check`
- `cargo nextest run -p fabro-model -p fabro-config -p fabro-auth -p
fabro-llm`
- `cargo build --workspace`
- `cargo nextest run --workspace` (5717 passed, 182 skipped, nextest
reported 1 leaky test)
- `cargo +nightly-2026-04-14 fmt --check --all`
- `cargo +nightly-2026-04-14 clippy --workspace --all-targets -- -D
warnings`
- `git diff --check`
## Summary
This PR advances the catalog-driven LLM work from fabro-sh/fabro#210 by
making the resolved model catalog the source of truth for provider
registration, request control validation, and billing identity. Runs now
preserve canonical provider/model/speed identity through pricing and API
responses instead of collapsing billing around provider API aliases or
model IDs alone.
## What Changed
- Register LLM provider adapters from the resolved catalog, including
custom OpenAI-compatible providers and their credential resolution
paths.
- Validate effective model request controls, including run-level
defaults and node overrides, before dispatching LLM requests.
- Add catalog-aware billing lookup that prices canonical `ModelRef`
values, uses base model costs for standard speed, applies per-speed cost
overrides, and returns an unknown estimate instead of silently billing
zero for unsupported combinations.
- Move Anthropic Opus fast-mode pricing into the built-in catalog for
`claude-opus-4-6` and `claude-opus-4-7`.
- Thread the injected catalog and effective speed controls through
workflow billing, including API-mode and CLI-mode handlers.
- Update billing APIs, server aggregation, generated clients, and the
web billing view to expose provider/model/speed billing identity and
keep standard and fast usage in separate rows.
## Notes for Review
Billing lookup intentionally uses canonical catalog model IDs. Provider
`api_id` substitution remains limited to provider request construction,
so aliases can be used on the wire without changing billing identity.
Event conversion paths that do not have catalog access now preserve
token counts with a null dollar estimate rather than falling back to the
bootstrap catalog.
## Verification
- `cargo build -p fabro-api`
- `cd lib/packages/fabro-api-client && bun run generate`
- `cargo +nightly-2026-04-14 fmt --check --all`
- `cargo +nightly-2026-04-14 clippy --workspace --all-targets -- -D
warnings`
- `ulimit -n 4096 && cargo nextest run -p fabro-model -p fabro-workflow
-p fabro-server -p fabro-api -p fabro-cli --no-fail-fast`
- `ulimit -n 4096 && cargo nextest run --workspace --no-fail-fast`
- `cd apps/fabro-web && bun run typecheck`
- `cd apps/fabro-web && bun test`
- `git diff --check`
---
[](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
This PR moves Fabro’s provider/model catalog toward settings-driven
provider identity by replacing the closed provider schema at the
API/auth/model boundary with `ProviderId`, then loading built-in
provider and model metadata from embedded per-provider TOML files.
The immediate result is that built-ins now use the same settings-shaped
catalog data that custom providers will use later, while request-serving
paths still keep the existing bootstrap/default catalog behavior until
the resolved-catalog plumbing lands.
## Changes
- Replaces API-facing provider enum usage with string-backed
`ProviderId`, including OpenAPI/progenitor replacements and regenerated
TypeScript client models.
- Routes model, auth, billing, CLI, server, and workflow call sites
through provider IDs where they cross product identity boundaries.
- Builds `Catalog` from settings-shaped provider/model data with
validation for adapter keys, OpenAI-compatible `base_url`, duplicate
aliases, provider defaults, disabled entries, model controls, and
per-speed cost rows.
- Replaces `catalog.json` with embedded provider TOML files under
`lib/crates/fabro-model/src/catalog/providers/`.
- Adds an explicit `fabro_model::bootstrap_catalog` hatch for
setup/install paths and extends the dev policy test to keep bootstrap
access contained.
- Preserves public training and knowledge-cutoff labels in LLM model
settings while still accepting bare TOML dates.
## Verification
- `cargo nextest run -p fabro-model -p fabro-config -p fabro-api` — 416
passed
- `cargo nextest run -p fabro-dev --features dev
bootstrap_catalog_references_stay_in_allowlist` — 1 passed
- `cargo +nightly-2026-04-14 fmt --check --all`
- `cargo +nightly-2026-04-14 clippy --workspace --all-targets -- -D
warnings`
- `cargo build --workspace`
- `git diff --check`
---
[](https://github.com/EveryInc/compound-engineering-plugin)
🤖 Generated with GPT-5 via [Codex](https://openai.com/codex)
## Summary
Adds the Phase 1 settings surface for gateway-backed LLM providers. This
was prompted by @haroldolivieri's Portkey/Bedrock field report on PR
#207, which showed that gateway auth and routing often live in custom
headers rather than the adapter's primary API-key header.
This PR is schema and seam work only. It does not make settings-defined
providers runnable yet; later phases still own ProviderId migration,
catalog construction, auth resolution, and production adapter
registration.
## Changes
- add typed `extra_headers` values to `[llm.providers.<id>]`
- support explicit `literal`, `env`, and `credential` header value forms
while rejecting bare strings, empty values, ambiguous tables, and
unknown keys
- cover whole-map header merge behavior and adapter header pass-through
tests
- update the settings-driven LLM plan with the Phase 1 gateway header
attribution and completion notes
## Non-goals
- does not make settings-defined providers runnable yet
- does not migrate ProviderId/OpenAPI/auth resolver/runtime catalog
plumbing
- does not route Codex OAuth through custom provider settings
## Tests
- `cargo nextest run -p fabro-config -p fabro-llm`
- `cargo +nightly-2026-04-14 fmt --check --all`
- `cargo +nightly-2026-04-14 clippy -p fabro-config -p fabro-llm
--all-targets -- -D warnings`
- `git diff --check origin/main...HEAD`
## Post-Deploy Monitoring & Validation
No additional operational monitoring required. This is schema and
adapter-seam coverage only; production provider registration and runtime
credential/header resolution remain deferred.
## Attribution
Motivated by @haroldolivieri's Portkey/Bedrock report on PR #207:
https://github.com/fabro-sh/fabro/pull/207#issuecomment-4377929769
Commits include `Co-authored-by: Haroldo Olivieri
<6575718+haroldolivieri@users.noreply.github.com>`.
---
Compound Engineered: Codex, `ce:work`.
---------
Co-authored-by: Haroldo Olivieri <6575718+haroldolivieri@users.noreply.github.com>
## Summary
This lays the groundwork for settings-driven LLM providers and models
without switching production routing yet. The new schemas and shared
vocabulary let later catalog construction treat provider/model identity
as data while keeping adapter behavior and control values Rust-owned.
## What changed
- Added `[llm.providers]` and `[llm.models]` settings layers with sparse
per-entry merging, whole-array replacement for credential/alias/control
lists, TOML date support for `knowledge_cutoff`, and typed `credential:`
/ `env:` references that reject literal secrets.
- Added `ProviderId`, `ModelId`, and a shared `ReasoningEffort` enum in
`fabro-model`, plus adapter metadata for `anthropic`, `openai`,
`gemini`, and `openai_compatible`.
- Added a matching `fabro-llm` adapter factory registry with parity
tests to keep metadata keys and factory keys in sync.
- Added `[run.model.controls]` defaults through config resolution and
runtime settings types.
- Added a workspace policy test to prevent future `bootstrap_catalog`
use outside install/test-support paths.
### Plan Summary
- This is the foundation slice of the settings-driven catalog plan.
- Production still uses the existing `Provider` enum and
`Catalog::builtin()` call paths.
- ProviderId routing, OpenAPI regeneration, auth resolver changes,
resolved `Arc<Catalog>` injection, typed request speed, and per-speed
billing are deferred follow-ups.
⚒️ Generated with [Fabro](https://fabro.sh)
---------
Co-authored-by: Fabro <noreply@fabro.sh>
Co-authored-by: fabro-bot <fabro-bot@fabro.sh>
Co-authored-by: Bryan Helmkamp <bryan@brynary.com>
Always request encrypted reasoning content for Responses API requests with store:false so GPT-5.5 reasoning items can be sent back on later turns without relying on server-side item persistence.
Use a lightweight basic probe target for preflight instead of fabricating catalog models, run configured model probes with bounded concurrency, and keep expensive model choices opt-in for defaults and live tests.
Propagate OpenAI Responses SSE error and response.failed events as structured provider errors, and treat response.incomplete as a normal length finish with partial output preserved.
Also preserve those stream errors through Codex-mode complete_via_stream and add agent coverage proving quota failures do not replay the turn.
GPT-5.5 (released 2026-04-23) replaces 5.4 as the default OpenAI model.
Live integration tests confirm both new IDs respond on the OpenAI API;
they require default temperature like other reasoning models.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add configured to the model API contract and server responses so clients can see whether provider credential material exists before testing. Use that signal in bulk model tests to skip unconfigured providers before printing progress and treat post-list skips as race failures.
Promotes per-run observability events (stage start/complete, edge
selection, checkpoint, fidelity resolution, agent session, LLM stream
finish, tool calls, sandbox cleanup, PR build/create) from debug to
info so default-level operators see end-to-end run progress.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Remove redundant as_str/from helper methods on provider, reasoning, model-test, safe URL, and interview types. Migrate call sites to Display, IntoStaticStr, and FromStr while keeping wire-format coverage in tests.
Move secret redaction and DisplaySafeUrl into fabro-redact so credential handling has a narrow ownership boundary. Update direct consumers and docs to depend on fabro_redact instead of fabro_util::redact.
Add DisplaySafeUrl under fabro-util::redact so URL Display and Debug output redact credentials by default. Migrate token-bearing GitHub, OAuth, server, LLM, sandbox, and workflow paths to use the wrapper at logging/error boundaries while keeping raw URLs explicit for wire and shell transit.
Add fabro-static::EnvVars as the shared registry for fixed environment variable names and migrate env reads, clap env bindings, and subprocess/test allowlists to use it.
Add clippy bans for raw std::env lookup APIs so future dynamic env facades must be documented explicitly.
Let callers decide whether to wrap in Arc. Also consolidates the two
state() fetches in build_pr_body into one.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replace hand-written Display/FromStr/as_str boilerplate with strum
derives on Provider, RunStatus, StatusReason, Speed, ReasoningEffort,
SandboxProvider, Fidelity, ModelTestMode, ModelTestStatus. Update a few
downstream callers whose FromStr::Err = String assumption no longer
holds. Net -172 lines, zero wire-format change.
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.
Replace the remaining blocking filesystem touches in shared async code with
Tokio-native I/O or explicit blocking boundaries. This keeps provider file
loading, workflow metadata rebuilds, and related export paths compatible with
the stricter clippy async-fs rules without changing their external behavior.
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>
API keys in query strings leak to access logs, proxies, and request
traces. Move to the header form Google documents as equivalent for both
generateContent and streamGenerateContent endpoints.
A recent CI flake surfaced as bare "error: No such file or directory
(os error 2)" with no chain, because the failing operation lived behind
a raw `?` on a `std::fs::` / `File::create` / `Command::spawn` call. The
error had no verb, no path, no hint at which step in server startup
broke. Retry loops were explicitly rejected -- the goal is to diagnose
the next occurrence, not mask it.
Wraps 50+ such sites across fabro-cli, fabro-server, fabro-workflow,
fabro-util, fabro-vault, fabro-telemetry, fabro-interview, fabro-llm,
and fabro-devcontainer with `.with_context(|| format!("<verb> {path}"))`
so anyhow's error chain carries both the operation and the path when
an io error escapes.
Where the enclosing function returns `io::Result` (fabro-util run_log,
fabro-interview recording, fabro-llm attachment loader), the error is
re-wrapped via `io::Error::new` to keep the signature stable. Where a
crate uses its own thiserror enum, either a new `io_context` helper
was added (fabro-vault) or the path was folded into the existing
`Error::Io(String)` message (fabro-workflow).
No retry loops. No behavior changes. Skipped sites documented:
`.ok()`-swallowed, `match ErrorKind::NotFound`, `let _ = ...`, typed
error variants that already carry the path, and test modules.
Verified: cargo build --workspace, cargo +nightly clippy --workspace
--all-targets -- -D warnings, cargo nextest run --workspace (3991/3991
pass).
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>
The fabro-http crate's proxy policy mechanism was not being used in
tests. http_api.rs used #[cfg(test)] to call .no_proxy(), but cfg(test)
only applies within the crate being tested — downstream crates like
fabro-workflow and fabro-cli hit the production path with system proxy
discovery, adding ~900ms per reqwest client per process.
- Set FABRO_HTTP_PROXY_POLICY=disabled in .cargo/config.toml so all
test HTTP clients skip proxy discovery automatically
- Remove dead #[cfg(test)] branch in http_api.rs; it now relies on the
env var like every other fabro-http consumer
- Remove kind(test) from nextest overrides so timeout budgets apply to
unit tests too, not just integration tests
- Remove unused SessionCookie import in web_auth.rs
Eliminates all 11 flaky nextest timeouts under parallel load.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Share provider display names and OAuth expiry helpers across auth, CLI,
and server code, and simplify the small match arms and helper plumbing
that full-workspace clippy surfaced during final verification.
- Replace unwrap_or_default() with expect() in hooks/llm HTTP client
builders — Default silently discards all config (timeouts, TLS, proxy)
- Route fabro-mcp through fabro_http instead of raw reqwest, respecting
FABRO_HTTP_PROXY_POLICY for MCP HTTP transport connections
- Deduplicate HttpClientBuilder / BlockingHttpClientBuilder via macro
- Extract helpers for repeated http_client error handling in diagnostics
and web_auth
- Remove duplicate test_http_client() in fabro-cli and fabro-llm
Co-Authored-By: Claude Opus 4.6 (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.
No production deployments exist, so there's no need for migration shims.
Remove all six backwards-compat type aliases (AgentError, SdkError,
CoreError, GraphvizError, StoreError, FabroError) and migrate ~880
callsites to use the canonical Error name directly within each crate,
or qualified imports (e.g., `use fabro_llm::Error as LlmError`) for
cross-crate references. Also fix a pre-existing absolute-path clippy
lint in fabro-server error.rs.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
