feat(tinyagents): Phase 0/1 port cutover — pin v1.7.1, crate-back SchemaCleanr / model-context / observed stream#4539
Conversation
…ools/agent_orchestration Claude-Session: https://claude.ai/code/session_017389bP8GsLiFUCmqEd4ReL
Use the crate-backed schema cleaner and generic model context hints, adopt context-preserving invoke_stream for observed runs, and record the remaining Phase 1 host wrapper boundaries.
Audit of the plan against current code surfaced two inaccuracies: - §5: the 12-positional-arg `#[allow(too_many_arguments)]` is on `all_tools_with_runtime`, not `all_tools` (the thin `all_tools` shim forwards 9). Re-attribute so the "builder overdue" note points at the real offender. - §1.1: the reasoning-content row still said the channel is "today smuggled via ProviderExtension", contradicting the CLOSED status in §3 Phase 1.4 / ledger P1-5. The host now writes ContentBlock::Thinking for new messages and only reads legacy ProviderExtension from persisted transcripts; mark the motion as largely done. Claude-Session: https://claude.ai/code/session_01LHbWHHV8MAFsv8aLi9XSbB
Reconcile the TinyAgents v1.7.1 port branch with main so the PR is mergeable and CI can run. Conflicts resolved: - Cargo.toml: union of both dependency sets — keep the branch's tinyagents 1.7.1 bump AND main's new tinyjuice/tinycortex deps and tokenjuice-treesitter refactor. Restored the `repl` feature main added to tinyagents (powers main's new `rlm` language-workflow tool) on top of the 1.7.1 pin: features = ["sqlite", "repl"]. Verified v1.7.1 still exposes `repl = ["dep:rhai"]`. - docs/tinyagents-port-plan.md: add/add conflict — kept the branch's evolved plan (main only had the original draft from squash-merged tinyhumansai#4525). - vendor/tinyagents submodule: kept the branch's v1.7.1 / 3e81e493 pin. - vendor/tinycortex, vendor/tinyjuice: adopted main's pins; submodules initialized. - Both Cargo.lock files reconciled via `cargo metadata` (resolution succeeds on both Cargo worlds; rhai now present in the tauri lockfile). Claude-Session: https://claude.ai/code/session_01LHbWHHV8MAFsv8aLi9XSbB
|
Warning Review limit reachedYou’ve reached a temporary PR review limit under our Fair Usage Limits Policy. Next review available in: 29 minutes Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available. How can I continue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews. How do review limits work?CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability. For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window. Please refer docs for additional details. Review details⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Run ID: 📒 Files selected for processing (8)
Comment |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d968709ac6
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| AgentStreamItem::Failed(error) => { | ||
| terminal = Some(Err(tinyagents::TinyAgentsError::Model(error))); |
There was a problem hiding this comment.
When on_progress is set, this wrapped value is now the only terminal error that reaches the existing error mapper; converting every AgentStreamItem::Failed into TinyAgentsError::Model discards harness variants such as LimitExceeded and SubAgentDepth/RecursionLimit. The handler just below only translates those concrete variants into MaxIterationsExceeded or the spawn-depth error when the variant is preserved, so an interactive streamed run that fails for one of those harness conditions now surfaces as a generic harness/model failure instead of the typed error callers downcast and render specially.
Useful? React with 👍 / 👎.
Audit against current code found the arm is not globally dead: the agent-tool running_subagents::steer path enforces parent ownership and returns NotOwned (with a passing test). Only the trusted RPC steer_control path skips the check, making the arm unreachable there. Reword §5 to reflect this and point at the Phase 4 SteeringRegistry consolidation. The matching code comment is clarified in fix/tinyagents-audit-fixes (PR tinyhumansai#4542). Claude-Session: https://claude.ai/code/session_01LHbWHHV8MAFsv8aLi9XSbB
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 634ded9b01
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| ); | ||
| return Some(*window); | ||
| } | ||
| if let Some(window) = tinyagents::harness::model::context_window_for_model_id(normalized) { |
There was a problem hiding this comment.
For raw model ids where o1/o3 is a delimiter-bounded segment but not the first segment after the provider prefix, such as ollama/mistral-for-o1-benchmark (the regression guard in this file still expects this to resolve), this delegation drops the match when there is no cost-catalog row. TinyAgents v1.7.1 only compares the first -/_/. segment of the final model name, so these ids now return None and skip the 200k pre-dispatch budget (or fall back to a much smaller local profile), letting long histories reach providers without the intended trimming and fail with context-window errors.
Useful? React with 👍 / 👎.
The Rust Core Coverage job (which compiles `cargo test`, unlike the `cargo check --lib` used for the earlier local validation) surfaced three test-only E0063 errors against TinyAgents 1.7.x: - middleware.rs test ToolRuntime literals missing `timeout` → set `ToolTimeout::Inherit` (matches the source-side seam mapping). - observability.rs ToolCompleted test event missing `started_at_ms` → `None` (mirrors the sibling test at line 1273 that already had it). Source (non-test) constructors already carried these fields, so `cargo check --lib` passed while the test build did not. Verified by a `cargo check --tests` run that clears the E0063s. Claude-Session: https://claude.ai/code/session_01LHbWHHV8MAFsv8aLi9XSbB
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 5a3b74cdd3
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| #[cfg(test)] | ||
| #[path = "schema_tests.rs"] | ||
| mod tests; | ||
| pub use tinyagents::harness::tool::{CleaningStrategy, SchemaCleanr, GEMINI_UNSUPPORTED_KEYWORDS}; |
There was a problem hiding this comment.
With this replacement the old #[path = "schema_tests.rs"] mod tests; hook is gone, and rg "schema_tests" shows no remaining include for src/openhuman/tools/schema_tests.rs. That means the host seam tests for Gemini ref resolution, union flattening, null stripping, and validation no longer run in cargo test; a mismatch in the new TinyAgents re-export can ship unnoticed even though the test file is still present. Please reattach these tests or remove/port them explicitly.
Useful? React with 👍 / 👎.
The port delegated generic model-context fallback to the crate's context_window_for_model_id, which resolves canonical o1/o3 ids but does not match an o1/o3 token embedded mid-name (e.g. ollama/mistral-for-o1-benchmark). That regressed o1_o3_segment_match_does_not_overmatch (surfaced by the full-suite coverage run, which the port's Cargo.toml/lock change triggers). Restore a minimal host-side segment heuristic that runs only when the crate returns None: an o1/o3 token bounded by non-alphanumeric characters maps to the 200K window, while substrings (solo1-7b, proto3-chat, octo3thing) do not over-match. Preserves the pre-port behavior (PR tinyhumansai#2100 regression guard) until the crate matcher covers internal segments. Claude-Session: https://claude.ai/code/session_01LHbWHHV8MAFsv8aLi9XSbB
Main merged tinyhumansai#4552 (tinyflows integration) which independently bumped tinyagents to v1.7.1 and updated the seam. Reconcile: - Cargo.toml: take main's `tinyagents = "1.7"` line (semver-equivalent to the branch's "1.7.1"; the [patch] pins the vendored submodule anyway). - tinyagents/observability.rs, tinyagents/tools.rs: take main's version — its tinyhumansai#4552 seam update is the authoritative 1.7.1 adaptation (ignores the new ToolCompleted fields via `..` in favour of its own capture side channel). The branch's redundant test-constructor field fills are dropped in favour of main's. - Kept this branch's unique work (no conflict): tools/schema.rs re-export of the crate SchemaCleanr, model_context.rs delegation to the crate + the host o1/o3 segment fallback, reasoning ContentBlock::Thinking, and the port-plan/drift-ledger docs. - Both Cargo.lock files reconciled via `cargo metadata`. Claude-Session: https://claude.ai/code/session_01LHbWHHV8MAFsv8aLi9XSbB
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 92e3bb95f7
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| [[package]] | ||
| name = "tinyflows" | ||
| version = "0.3.2" | ||
| version = "0.3.0" |
There was a problem hiding this comment.
This lockfile-only downgrade selects tinyflows 0.3.0 even though the checked-in tree already uses the observed journal APIs added after 0.3.0 (for example run_with_checkpointer_journaled_observed / resume_with_checkpointer_journaled_observed in src/openhuman/flows/ops.rs). Since neither Cargo.toml nor the vendor/tinyflows submodule changes in this commit, keeping the older lock entry will either make cargo --locked reject the stale path-package version or compile against a crate version that lacks those flow-run APIs; please keep both lockfiles on the version from the parent lock.
Useful? React with 👍 / 👎.
Main's tinyhumansai#4552 already bumped tinyagents to v1.7.1 in both lockfiles, so this branch no longer needs to touch Cargo.toml/Cargo.lock at all. Reset both lockfiles to main's so the PR diff is source + docs only — which keeps CI on the domain-scoped lane (inference/tinyagents/tools) instead of the config-level full-suite fallback. The full suite was failing to compile the pre-existing-broken tests/subconscious_e2e.rs (stale SubconsciousEngine::new / with_connection call sites that are also broken on main but never compiled by main's scoped CI) — unrelated to this port. Claude-Session: https://claude.ai/code/session_01LHbWHHV8MAFsv8aLi9XSbB
1 participant
Summary
Cargo.locks +vendor/tinyagentssubmodule) and route real host paths through crate primitives.tools/schema.rsnow re-exportstinyagents::harness::tool::{SchemaCleanr, CleaningStrategy, GEMINI_UNSUPPORTED_KEYWORDS}(the in-treeSchemaCleanris deleted; import path preserved).inference/model_context.rsdelegates the generic context-window fallback totinyagents::harness::model::context_window_for_model_idafter checking OpenHuman tier aliases + the cost catalog first.invoke_stream_in_context(theSend-stream fix landed in tinyagents#28 / v1.7.1); reasoning is written to typedContentBlock::Thinkinginstead of theProviderExtensionsmuggle.docs/tinyagents-port-plan.md(audit + phased plan) anddocs/tinyagents-drift-ledger.md(per-row upstream/host/delete tracking).Problem
The host carries a large parallel implementation of things TinyAgents already ships (provider wire client, schema cleaning, retry, tool/subagent model). The audit (see the plan doc) establishes that most of the "bulky" code is duplication, not portable logic, and lays out a phased consolidation so each deletion is gated behind a released crate API + a crate-backed seam test. This PR lands Phase 0 (version alignment) and the Phase 1 partial cutover — the pieces that are safe to switch on now — without deleting any host code that still lacks a released crate replacement.
Solution
vendor/tinyagents, released to crates.io, then pinned here in standalonechore(vendor)commits (v1.6.0 → v1.7.1); host PRs contain no engine source edits.ToolCompletedoutcome projection, SHA-256 prompt-cache fingerprinting over the fullToolSchemalist, and the redaction-layer audit.SchemaCleanr, generic model-context lookup, typed reasoning channel, and the context-preserving observed stream. Worktree / time / error-classification / display-metadata APIs are released but host-wrapper/call-site retained (documented per drift-ledger row) until their Phase 2 model reconciliation.all_tools_with_runtimeowns the 12-argtoo_many_arguments; the reasoning-channel row no longer claims "smuggled via ProviderExtension").upstream/maininto the branch to keep it mergeable: unionedCargo.toml(kept the branch's tinyagents 1.7.1 bump + restored thereplfeature main added for its newrlmtool + main'stinyjuice/tinycortexdeps); both lockfiles reconciled viacargo metadata.Submission Checklist
schema_,context_window,tinyagents/*conversion) continue to guard the seam.SchemaCleanr), docs, and lockfile/manifest bumps; the remaining changed host lines are exercised by existing seam tests. Full validation deferred to CI runners per request.N/A: no new user-facing feature rows(internal engine consolidation; wire surface unchanged).## Related—N/A: behaviour-preserving consolidation.tinyagentsbump, plus main'stinyjuice/tinycortex) are vendored submodules patched to local paths.N/A: no release-cut surface change.## Related(tracking issue Improve agent harness with LangGraph-style state machine or Polaris-like architecture #4249 workstream; not auto-closed by this partial phase).Impact
ContentBlock::Thinkingfor new messages while legacyProviderExtensionreasoning is still read from existing transcripts.SecurityPolicy/approval gating; upstreamed code is generic (no product policy or backend phrases moved).vendor/tinyagentspinned to the v1.7.1 tag (never floated).Related
compatible*.rsdeletion), Phase 4 (orchestration ontograph::orchestration) — all enumerated indocs/tinyagents-port-plan.md§3 and the drift-ledger phase gates.AI Authored PR Metadata (required for Codex/Linear PRs)
Linear Issue
Commit & Branch
docs/tinyagents-port-pland968709ac63066179c97f0fbd9a201c740554555Validation Run
pnpm --filter openhuman-app format:check— N/A (noapp/srcchanges)pnpm typecheck— N/A (no TypeScript changes)cargo fmt --checkpassed; hostcargo check --librun locally on the merged tree (see Validation Blocked for the environment caveat); full suites deferred to CI.cargo fmt --checkclean; dependency resolution succeeds on both Cargo worlds (cargo metadataexit 0 for root +app/src-tauri).Validation Blocked
command:cargo check --lib --manifest-path Cargo.tomlerror:earlier local runs hit a transientcouldn't create a temp dirintarget/debug/depsunder concurrent-build contention (not a code error); full compile/test validation intentionally deferred to GitHub runners per the requester.impact:CI (CI Lite on the PR, CI Full on the release lane) is the authoritative gate before merge.Behavior Changes
Parity Contract
Duplicate / Superseded PR Handling