feat: Add multi-provider LLM support via rig-core by zmanian · Pull Request #36 · nearai/ironclaw

zmanian · 2026-02-11T19:45:36Z

Summary

Add OpenAI, Anthropic, Ollama, and OpenAI-compatible LLM backends via rig-core v0.30 adapter, keeping NEAR AI as the default
New LlmBackend enum and per-provider config structs with env var resolution (LLM_BACKEND, OPENAI_API_KEY, ANTHROPIC_API_KEY, OLLAMA_BASE_URL, etc.)
Generic RigAdapter<M: CompletionModel> bridges rig-core to the existing Arc<dyn LlmProvider> trait with full message, tool, and response conversion
Session authentication is now conditional on the NearAi backend; other backends skip it entirely

Why rig-core?

Multi-provider support was the top community ask -- users want to bring their own OpenAI/Anthropic keys or run local models via Ollama without needing a NEAR AI account. We evaluated several approaches:

Option 1: Hand-roll each provider's HTTP API. Straightforward but high surface area -- each provider has its own auth, message format, tool calling convention, streaming protocol, and error semantics. Maintaining parity across 4+ providers is an ongoing burden.

Option 2: Use rig-core as an adapter layer. rig-core (v0.30, 5.9K stars, actively maintained) provides a unified CompletionModel trait across 19+ providers with correct tool calling, message conversion, and usage tracking already implemented. We write one adapter (RigAdapter) and get all providers for free.

Option 3: Other multi-provider crates (e.g. llm-chain, langchain-rust). These are either less mature, more opinionated about the agent loop (we already have our own), or pull in heavier dependency trees.

We went with Option 2 because:

Minimal integration surface: One generic struct (RigAdapter<M: CompletionModel>) implementing our existing LlmProvider trait. All agent, reasoning, and tool code is untouched.
Low dependency cost: Most of rig-core's transitive deps (reqwest, serde, tokio, thiserror, tracing, futures) already exist in our tree. The incremental addition is small.
Provider correctness for free: Tool calling schemas, message role mapping, and token usage extraction are already battle-tested across providers. We don't need to reverse-engineer each API's quirks.
Future providers at zero marginal cost: Any new provider rig-core adds (Gemini, Groq, DeepSeek, etc.) becomes available by adding a few lines to our factory function.
No lock-in: The adapter is ~300 lines. If rig-core ever becomes unsuitable, replacing it with direct HTTP calls is straightforward since the LlmProvider trait boundary isolates all consumers.

NEAR AI remains the default and recommended backend. The rig-core adapter is purely additive -- it doesn't change any existing NEAR AI code paths.

Test plan

cargo fmt -- clean
cargo clippy --all --benches --tests --examples --all-features -- only pre-existing warnings
cargo test -- 675 unit tests pass, 0 new failures
Manual test: LLM_BACKEND=openai OPENAI_API_KEY=sk-... cargo run
Manual test: LLM_BACKEND=ollama OLLAMA_MODEL=llama3 cargo run
Manual test: LLM_BACKEND=anthropic ANTHROPIC_API_KEY=sk-ant-... cargo run

Generated with Claude Code

Add support for OpenAI, Anthropic, Ollama, and OpenAI-compatible endpoints alongside the existing NEAR AI backend. Users can now bring their own API keys via environment variables (LLM_BACKEND, OPENAI_API_KEY, ANTHROPIC_API_KEY, etc.) while NEAR AI remains the default. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

matthiasdebernardini · 2026-02-11T20:28:41Z

Dont use ollama, use lmstudio or vllm

ollama is slow

Add support for OpenAI, Anthropic, Ollama, and OpenAI-compatible endpoints alongside the existing NEAR AI backend. Users can now bring their own API keys via environment variables (LLM_BACKEND, OPENAI_API_KEY, ANTHROPIC_API_KEY, etc.) while NEAR AI remains the default. Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Co-authored-by: Guillermo Alejandro Gallardo Diez <gagdiez@iR2.local>

Closes the remaining libSQL/Postgres/JSONL gaps surfaced in the latest serrrfirat review pass on PR #3171. libSQL production gate (#34, #36, #41): - Reject `:memory:` for Production via the InMemory error variant. - Case-insensitive scheme matching for `HTTP://` / `HTTPS://` / `LibSQL://`, so a mixed-case URL no longer falls through to Builder::new_local and silently creates a node-local SQLite path. - Reject bare hostname-like values (e.g. `db.example.com`) for Production via a new ProductionLibsqlAmbiguousTarget error; LocalDev still accepts ergonomic forms like `events.db`. libSQL backend hardening (#45, #49, #50): - Skip `create_dir_all("")` for `events.db` in cwd. - Enable `journal_mode=WAL` + `synchronous=NORMAL` once at build time for file-backed local stores so a long replay reader can no longer block writer commits past `busy_timeout`. - Mirror the v1 retry pattern (`src/db/libsql/mod.rs::connect`): three attempts with exponential backoff so concurrent transient "unable to open database file" errors do not surface as durable-log failures. Postgres backend hardening (#46): - Honor `sslmode=require` for loopback configs so a TLS-only local Postgres / loopback proxy is accepted instead of forced to NoTls. JSONL hardening (#38, #40, #42, #43, #44): - Make JSONL constructors crate-private so production composition cannot bypass the single-node-durable acceptance gate. - Hash path components (SHA-256 / 16-hex-char prefix + 32-byte URL- encoded hint) so case-distinct IDs (`Alice` vs `alice`) cannot collide on case-insensitive filesystems and 256-byte scope IDs no longer overflow the 255-byte filename limit. - Snapshot file length before write/flush/sync; truncate back on any error so a partial write never leaves a torn JSON tail that wedges every subsequent append. - Create directories with `0o700` and stream files with `0o600` on Unix so durable history is not world-readable under the typical `umask 022`. SQL replay correctness (#37): - After fetching filtered rows, run a small unfiltered COUNT over the scanned cursor window in both libSQL and Postgres backends; mismatch surfaces a missing entry row as `EventError::ReplayGap`. JSONL already detects this via line-by-line cursor sequencing. Regression tests: - libSQL/Postgres production gate (case-insensitive scheme, in-memory, bare hostname). - Hashed-path-component case distinctness and length boundedness. - Atomic JSONL append: failed serialise leaves file at pre-append length and the next append still advances cursor cleanly. - Unix `0o700` JSONL root permissions. - libpq quoted single-quote socket path is local; whitespace-around- `=` keyword strings classify remote correctly. - libSQL replay surfaces a deleted entry row as `ReplayGap`. Deferred with explicit acknowledgement: - #39 (`ReadScope.invocation_id`) — module-doc note; needs a cross-crate change to `ironclaw_events` plus every replay caller. - #48 (replay holds writer-blocking locks) — inline note at the JSONL read path; needs a stream-bytes-snapshot redesign coordinated with the durable-log contract. Test plan - cargo test -p ironclaw_reborn_event_store - cargo test -p ironclaw_reborn_event_store --features "libsql postgres" - cargo test -p ironclaw_host_runtime - cargo test -p ironclaw_events - cargo test -p ironclaw_architecture reborn_crate_dependency_boundaries_hold - cargo clippy -p ironclaw_reborn_event_store --all-targets -- -D warnings - cargo clippy -p ironclaw_reborn_event_store --features "libsql postgres" --all-targets -- -D warnings - cargo clippy -p ironclaw_reborn_event_store --no-default-features --all-targets -- -D warnings - cargo clippy -p ironclaw_host_runtime --all-targets -- -D warnings - cargo fmt --all -- --check Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* Add Reborn event store backends * Address Reborn event store review feedback * Address Reborn event store review feedback (round 2) Postgres TLS / fail-closed (comment 3178548634): the Postgres event-store client now uses tokio-postgres-rustls for any non-loopback URL, mirroring src/db/tls.rs (native certs with webpki-roots fallback). Local sockets and loopback hosts continue to use NoTls; unparseable URLs with a scheme are treated as remote so a typo cannot silently downgrade to plaintext. next_cursor advancement (comment 3178548646): JSONL, libSQL, and Postgres read paths now track the highest scanned cursor independently from the last matched cursor, and return max(last_matched, last_scanned). Without this, a matched record followed by a filtered-out record would leave next_cursor pinned to the matched cursor and the filtered record would be rescanned on every subsequent replay. JSONL bounded replay (comment 3178548670): replays now stream the JSONL file line-by-line via BufReader, decoding only the cursor envelope until a line crosses `after`, and stop as soon as `limit` matches are collected. A `limit = 1` request against a multi-gigabyte stream no longer pays full-file allocation or full-stream parse latency. JSONL cross-process locking (comment 3178548701): JSONL appends now take an OS-level exclusive advisory lock (std::fs::File::lock) for the entire read-tail-cursor + write window. Two IronClaw processes pointing at the same JSONL root will block on this lock and emit monotonically-sequenced cursors instead of corrupting the stream with duplicates. Readers take a shared lock to avoid observing partially-written tail lines. New deps on the event-store crate (postgres feature only): tokio-postgres-rustls, rustls, rustls-native-certs, webpki-roots, url. All four were already used elsewhere in the workspace; no new external crates are introduced. File locking uses stdlib (Rust 1.89+ stable). Regression tests added: - jsonl_replay_advances_next_cursor_past_trailing_filtered_records - jsonl_concurrent_appenders_emit_monotonic_cursors_through_file_lock - jsonl_bounded_replay_does_not_parse_the_whole_file - postgres_store::tests::{local_postgres_urls_are_recognised, remote_postgres_urls_require_tls, unparseable_postgres_url_with_scheme_falls_closed_to_remote} Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(reborn_event_store): close TLS bypass for libpq keyword strings and http:// libsql in production Two High-severity findings from PR #3171 review: 1. is_local_postgres_url() previously returned true for any string without "://", so libpq keyword form like "host=db.example.com user=..." took the NoTls branch and connected over plaintext. Walk the keyword list, look at the actual host, and only treat localhost / socket paths as local. Tests cover remote keyword strings, socket paths, and missing host=. 2. Production profile previously accepted http:// libSQL URLs and forwarded the auth token in cleartext. Reject http:// for RebornProfile::Production with a typed error before the build call. LocalDev / Test still accept http:// for running against local sqld instances. Tests cover both. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(reborn_event_store): use deadpool-postgres pool for transparent reconnection Replaces the single Arc<Client> in PostgresStore with a deadpool_postgres::Pool. The previous design built one tokio_postgres::Client at startup; if the underlying connection task exited (idle timeout, server restart, failover, transient network drop), all subsequent append/read operations would fail forever because nothing rebuilt the connection. deadpool's Manager replaces broken Clients on the next pool.get() call, so each call site sees a live connection without per-site reconnect logic. This mirrors v1's connection-management story (src/db/postgres.rs, src/db/tls.rs both use deadpool-postgres) and avoids two divergent reconnection paths in one binary. Addresses serrrfirat's Medium-severity finding on PR #3171. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * chore(tests): fix post-merge clippy warning * fix(reborn_event_store): close TLS-mode + libpq-config + JSONL fsync gaps Three follow-ups from PR #3171 round-3 review: 1. CRITICAL — remote Postgres TLS not actually enforced. Passing a rustls connector is not enough on its own: tokio-postgres only consults the connector when Config::ssl_mode is Prefer or Require, and `sslmode=disable` in the connection string returns a plaintext stream before TLS is attempted. Add `enforce_remote_ssl_mode`: reject Disable for non-local configs, and force Prefer (the default) up to Require so the server cannot decline TLS without failing the connection. 2. HIGH — local/remote detector missed valid libpq forms. Re-parsing the raw connection string failed on `hostaddr=10.0.0.5` (numeric-IP keyword, no `host=` entry), `postgresql:///db?host=db.example.com` (URL with empty authority + host in query), and `host=/var/run/postgresql, db.example.com` (mixed socket/TCP list). Replace with `is_local_postgres_config` that walks the parsed `Config::get_hosts()` and `Config::get_hostaddrs()` so all libpq normalisations land in the same code path. 3. MEDIUM — JSONL append fsynced file contents but not the parent directory entry. On POSIX the new file's directory entry must also be fsynced for crash durability; without it the first append can vanish after a power loss even though `append()` returned success. Detect first-create via `path.exists()` pre-open and fsync the parent dir after `sync_data()`. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(reborn_event_store): address PR #3171 round-3 review findings Closes the remaining libSQL/Postgres/JSONL gaps surfaced in the latest serrrfirat review pass on PR #3171. libSQL production gate (#34, #36, #41): - Reject `:memory:` for Production via the InMemory error variant. - Case-insensitive scheme matching for `HTTP://` / `HTTPS://` / `LibSQL://`, so a mixed-case URL no longer falls through to Builder::new_local and silently creates a node-local SQLite path. - Reject bare hostname-like values (e.g. `db.example.com`) for Production via a new ProductionLibsqlAmbiguousTarget error; LocalDev still accepts ergonomic forms like `events.db`. libSQL backend hardening (#45, #49, #50): - Skip `create_dir_all("")` for `events.db` in cwd. - Enable `journal_mode=WAL` + `synchronous=NORMAL` once at build time for file-backed local stores so a long replay reader can no longer block writer commits past `busy_timeout`. - Mirror the v1 retry pattern (`src/db/libsql/mod.rs::connect`): three attempts with exponential backoff so concurrent transient "unable to open database file" errors do not surface as durable-log failures. Postgres backend hardening (#46): - Honor `sslmode=require` for loopback configs so a TLS-only local Postgres / loopback proxy is accepted instead of forced to NoTls. JSONL hardening (#38, #40, #42, #43, #44): - Make JSONL constructors crate-private so production composition cannot bypass the single-node-durable acceptance gate. - Hash path components (SHA-256 / 16-hex-char prefix + 32-byte URL- encoded hint) so case-distinct IDs (`Alice` vs `alice`) cannot collide on case-insensitive filesystems and 256-byte scope IDs no longer overflow the 255-byte filename limit. - Snapshot file length before write/flush/sync; truncate back on any error so a partial write never leaves a torn JSON tail that wedges every subsequent append. - Create directories with `0o700` and stream files with `0o600` on Unix so durable history is not world-readable under the typical `umask 022`. SQL replay correctness (#37): - After fetching filtered rows, run a small unfiltered COUNT over the scanned cursor window in both libSQL and Postgres backends; mismatch surfaces a missing entry row as `EventError::ReplayGap`. JSONL already detects this via line-by-line cursor sequencing. Regression tests: - libSQL/Postgres production gate (case-insensitive scheme, in-memory, bare hostname). - Hashed-path-component case distinctness and length boundedness. - Atomic JSONL append: failed serialise leaves file at pre-append length and the next append still advances cursor cleanly. - Unix `0o700` JSONL root permissions. - libpq quoted single-quote socket path is local; whitespace-around- `=` keyword strings classify remote correctly. - libSQL replay surfaces a deleted entry row as `ReplayGap`. Deferred with explicit acknowledgement: - #39 (`ReadScope.invocation_id`) — module-doc note; needs a cross-crate change to `ironclaw_events` plus every replay caller. - #48 (replay holds writer-blocking locks) — inline note at the JSONL read path; needs a stream-bytes-snapshot redesign coordinated with the durable-log contract. Test plan - cargo test -p ironclaw_reborn_event_store - cargo test -p ironclaw_reborn_event_store --features "libsql postgres" - cargo test -p ironclaw_host_runtime - cargo test -p ironclaw_events - cargo test -p ironclaw_architecture reborn_crate_dependency_boundaries_hold - cargo clippy -p ironclaw_reborn_event_store --all-targets -- -D warnings - cargo clippy -p ironclaw_reborn_event_store --features "libsql postgres" --all-targets -- -D warnings - cargo clippy -p ironclaw_reborn_event_store --no-default-features --all-targets -- -D warnings - cargo clippy -p ironclaw_host_runtime --all-targets -- -D warnings - cargo fmt --all -- --check Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> Co-authored-by: serrrfirat <f@nuff.tech>

* Add Reborn event store backends * Address Reborn event store review feedback * Address Reborn event store review feedback (round 2) Postgres TLS / fail-closed (comment 3178548634): the Postgres event-store client now uses tokio-postgres-rustls for any non-loopback URL, mirroring src/db/tls.rs (native certs with webpki-roots fallback). Local sockets and loopback hosts continue to use NoTls; unparseable URLs with a scheme are treated as remote so a typo cannot silently downgrade to plaintext. next_cursor advancement (comment 3178548646): JSONL, libSQL, and Postgres read paths now track the highest scanned cursor independently from the last matched cursor, and return max(last_matched, last_scanned). Without this, a matched record followed by a filtered-out record would leave next_cursor pinned to the matched cursor and the filtered record would be rescanned on every subsequent replay. JSONL bounded replay (comment 3178548670): replays now stream the JSONL file line-by-line via BufReader, decoding only the cursor envelope until a line crosses `after`, and stop as soon as `limit` matches are collected. A `limit = 1` request against a multi-gigabyte stream no longer pays full-file allocation or full-stream parse latency. JSONL cross-process locking (comment 3178548701): JSONL appends now take an OS-level exclusive advisory lock (std::fs::File::lock) for the entire read-tail-cursor + write window. Two IronClaw processes pointing at the same JSONL root will block on this lock and emit monotonically-sequenced cursors instead of corrupting the stream with duplicates. Readers take a shared lock to avoid observing partially-written tail lines. New deps on the event-store crate (postgres feature only): tokio-postgres-rustls, rustls, rustls-native-certs, webpki-roots, url. All four were already used elsewhere in the workspace; no new external crates are introduced. File locking uses stdlib (Rust 1.89+ stable). Regression tests added: - jsonl_replay_advances_next_cursor_past_trailing_filtered_records - jsonl_concurrent_appenders_emit_monotonic_cursors_through_file_lock - jsonl_bounded_replay_does_not_parse_the_whole_file - postgres_store::tests::{local_postgres_urls_are_recognised, remote_postgres_urls_require_tls, unparseable_postgres_url_with_scheme_falls_closed_to_remote} Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(reborn_event_store): close TLS bypass for libpq keyword strings and http:// libsql in production Two High-severity findings from PR nearai#3171 review: 1. is_local_postgres_url() previously returned true for any string without "://", so libpq keyword form like "host=db.example.com user=..." took the NoTls branch and connected over plaintext. Walk the keyword list, look at the actual host, and only treat localhost / socket paths as local. Tests cover remote keyword strings, socket paths, and missing host=. 2. Production profile previously accepted http:// libSQL URLs and forwarded the auth token in cleartext. Reject http:// for RebornProfile::Production with a typed error before the build call. LocalDev / Test still accept http:// for running against local sqld instances. Tests cover both. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(reborn_event_store): use deadpool-postgres pool for transparent reconnection Replaces the single Arc<Client> in PostgresStore with a deadpool_postgres::Pool. The previous design built one tokio_postgres::Client at startup; if the underlying connection task exited (idle timeout, server restart, failover, transient network drop), all subsequent append/read operations would fail forever because nothing rebuilt the connection. deadpool's Manager replaces broken Clients on the next pool.get() call, so each call site sees a live connection without per-site reconnect logic. This mirrors v1's connection-management story (src/db/postgres.rs, src/db/tls.rs both use deadpool-postgres) and avoids two divergent reconnection paths in one binary. Addresses serrrfirat's Medium-severity finding on PR nearai#3171. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * chore(tests): fix post-merge clippy warning * fix(reborn_event_store): close TLS-mode + libpq-config + JSONL fsync gaps Three follow-ups from PR nearai#3171 round-3 review: 1. CRITICAL — remote Postgres TLS not actually enforced. Passing a rustls connector is not enough on its own: tokio-postgres only consults the connector when Config::ssl_mode is Prefer or Require, and `sslmode=disable` in the connection string returns a plaintext stream before TLS is attempted. Add `enforce_remote_ssl_mode`: reject Disable for non-local configs, and force Prefer (the default) up to Require so the server cannot decline TLS without failing the connection. 2. HIGH — local/remote detector missed valid libpq forms. Re-parsing the raw connection string failed on `hostaddr=10.0.0.5` (numeric-IP keyword, no `host=` entry), `postgresql:///db?host=db.example.com` (URL with empty authority + host in query), and `host=/var/run/postgresql, db.example.com` (mixed socket/TCP list). Replace with `is_local_postgres_config` that walks the parsed `Config::get_hosts()` and `Config::get_hostaddrs()` so all libpq normalisations land in the same code path. 3. MEDIUM — JSONL append fsynced file contents but not the parent directory entry. On POSIX the new file's directory entry must also be fsynced for crash durability; without it the first append can vanish after a power loss even though `append()` returned success. Detect first-create via `path.exists()` pre-open and fsync the parent dir after `sync_data()`. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(reborn_event_store): address PR nearai#3171 round-3 review findings Closes the remaining libSQL/Postgres/JSONL gaps surfaced in the latest serrrfirat review pass on PR nearai#3171. libSQL production gate (#34, #36, #41): - Reject `:memory:` for Production via the InMemory error variant. - Case-insensitive scheme matching for `HTTP://` / `HTTPS://` / `LibSQL://`, so a mixed-case URL no longer falls through to Builder::new_local and silently creates a node-local SQLite path. - Reject bare hostname-like values (e.g. `db.example.com`) for Production via a new ProductionLibsqlAmbiguousTarget error; LocalDev still accepts ergonomic forms like `events.db`. libSQL backend hardening (#45, #49, #50): - Skip `create_dir_all("")` for `events.db` in cwd. - Enable `journal_mode=WAL` + `synchronous=NORMAL` once at build time for file-backed local stores so a long replay reader can no longer block writer commits past `busy_timeout`. - Mirror the v1 retry pattern (`src/db/libsql/mod.rs::connect`): three attempts with exponential backoff so concurrent transient "unable to open database file" errors do not surface as durable-log failures. Postgres backend hardening (#46): - Honor `sslmode=require` for loopback configs so a TLS-only local Postgres / loopback proxy is accepted instead of forced to NoTls. JSONL hardening (#38, #40, #42, #43, #44): - Make JSONL constructors crate-private so production composition cannot bypass the single-node-durable acceptance gate. - Hash path components (SHA-256 / 16-hex-char prefix + 32-byte URL- encoded hint) so case-distinct IDs (`Alice` vs `alice`) cannot collide on case-insensitive filesystems and 256-byte scope IDs no longer overflow the 255-byte filename limit. - Snapshot file length before write/flush/sync; truncate back on any error so a partial write never leaves a torn JSON tail that wedges every subsequent append. - Create directories with `0o700` and stream files with `0o600` on Unix so durable history is not world-readable under the typical `umask 022`. SQL replay correctness (#37): - After fetching filtered rows, run a small unfiltered COUNT over the scanned cursor window in both libSQL and Postgres backends; mismatch surfaces a missing entry row as `EventError::ReplayGap`. JSONL already detects this via line-by-line cursor sequencing. Regression tests: - libSQL/Postgres production gate (case-insensitive scheme, in-memory, bare hostname). - Hashed-path-component case distinctness and length boundedness. - Atomic JSONL append: failed serialise leaves file at pre-append length and the next append still advances cursor cleanly. - Unix `0o700` JSONL root permissions. - libpq quoted single-quote socket path is local; whitespace-around- `=` keyword strings classify remote correctly. - libSQL replay surfaces a deleted entry row as `ReplayGap`. Deferred with explicit acknowledgement: - #39 (`ReadScope.invocation_id`) — module-doc note; needs a cross-crate change to `ironclaw_events` plus every replay caller. - #48 (replay holds writer-blocking locks) — inline note at the JSONL read path; needs a stream-bytes-snapshot redesign coordinated with the durable-log contract. Test plan - cargo test -p ironclaw_reborn_event_store - cargo test -p ironclaw_reborn_event_store --features "libsql postgres" - cargo test -p ironclaw_host_runtime - cargo test -p ironclaw_events - cargo test -p ironclaw_architecture reborn_crate_dependency_boundaries_hold - cargo clippy -p ironclaw_reborn_event_store --all-targets -- -D warnings - cargo clippy -p ironclaw_reborn_event_store --features "libsql postgres" --all-targets -- -D warnings - cargo clippy -p ironclaw_reborn_event_store --no-default-features --all-targets -- -D warnings - cargo clippy -p ironclaw_host_runtime --all-targets -- -D warnings - cargo fmt --all -- --check Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> Co-authored-by: serrrfirat <f@nuff.tech>

This was referenced Feb 11, 2026

If it's so local why do I need a NEAR AI account? #12

Closed

why do I need nearai integration? #37

Closed

ilblackdragon approved these changes Feb 12, 2026

View reviewed changes

ilblackdragon merged commit bb228f6 into nearai:main Feb 12, 2026

github-actions Bot mentioned this pull request Feb 12, 2026

chore: release v0.1.0 #46

Merged

gagdiez added a commit to gagdiez/ironclaw that referenced this pull request Apr 9, 2026

feat: add tutorial on hosting ironclaw through droplet (nearai#36)

fd210c6

Co-authored-by: Guillermo Alejandro Gallardo Diez <gagdiez@iR2.local>

zmanian mentioned this pull request May 5, 2026

Add Reborn event store backends #3171

Merged

henrypark133 mentioned this pull request Jun 26, 2026

fix(webui-v2): upgrade react-router 7.9.1 → 7.15.1 and rebuild bundle #5357

Closed

Navigation Menu

feat: Add multi-provider LLM support via rig-core - #36

feat: Add multi-provider LLM support via rig-core#36
ilblackdragon merged 1 commit into
nearai:mainfrom
zmanian:feat/multi-provider-llm

zmanian commented Feb 11, 2026 •

edited

Loading

Uh oh!

matthiasdebernardini commented Feb 11, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

zmanian commented Feb 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Why rig-core?

Test plan

Uh oh!

matthiasdebernardini commented Feb 11, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

zmanian commented Feb 11, 2026 •

edited

Loading