Skip to content

System hardening

This release hardens saige across the agent loop, providers, safety controls, storage, retrieval, evaluation, and the CLI. The goal is that a run never silently drops data, never runs a tool on partial or invalid input, never hangs on an unanswered decision, and never lets a model widen its own permissions.

Summary

  • Agent loop: tool panics, malformed arguments, truncated turns, and runaway loops are contained in tool results or clear errors. One run holds a branch at a time, and new input joins an active run only at safe points (queue, steer, interrupt, continue).
  • Structured output: Structured[T] validates, repairs, and can escalate to a stronger model through an OutcomePolicy. Response schemas are never dropped silently.
  • Delegation: sub-agents can be delegated, spawned in the background, or handed off. Every decision (approval, clarification, budget escalation) is an addressable interrupt that a durable engine can suspend on and resume.
  • Providers: every adapter accepts request options and tool choice, errors are classified (auth, rate limit, transient, invalid request), and retry, fallback, routing, traffic splits, and caching compose without stacking SDK retries.
  • Safety: placeholder-based privacy at the tool and provider boundary, capability gates, per-tool quotas, deferred tool disclosure, skills that only narrow permissions, host-scoped memory, and opt-in tool packs (fs, exec, fetch) that are read-only or approval-gated by default.
  • Storage and wire format: tenant-scoped stores, versioned writes, persisted tree state, and a versioned delta wire format.
  • RAG and eval: source sync with replace-by-URI, scoped retrieval, rank fusion, and an eval stack with gates, Wilson intervals, regression comparison, trajectory scorers, a results store, and manifest validation.
  • CLI: saige serve (HTTP and SSE with approve and cancel), JSON output, strict approval prompts, saige update, and an MCP server that enforces approval for mutating tools.

Final local check: gofmt -l ., go build ./..., go vet ./... and golangci-lint run ./... were clean, govulncheck ./... found no vulnerabilities, and go test -race -count=1 -p 2 ./... passed (83 packages ok, 0 failed, 24 without tests).

Work by area

Status values: implemented (done and covered by tests), partial (done with a known limit listed under Remaining gaps). No planned item was skipped.

Agent loop

Item Status Key files Tests
Concurrent tool calls keep call ID and result paired implemented agent/agent.go, agent/aggregator.go concurrency_test.go, aggregator_test.go, aggregator_args_test.go
Tool, gate, and handoff panics become error results implemented agent/agent.go run_panic_test.go, failure_test.go
Malformed or schema-mismatched arguments rejected before the gate implemented agent/types/toolargs.go, agent/types/schema_validate.go toolargs_test.go, schema_validate_test.go, tool_safety_test.go
Truncated turns committed with TruncationPart, their tool calls never run partial agent/forcing.go, agent/types/delta.go forcing_test.go, stream_test.go
One-turn forced tool choice, stop-at-tools, MaxIterForceFinal implemented agent/forcing.go, agent/types/request_options.go forcing_test.go, toolchoice_test.go, limits_test.go
Run guards: one run per branch, consecutive errors, repeated calls implemented agent/runguard.go, agent/deadline.go run_guard_test.go, limits_test.go
Submissions: queue, steer, interrupt-replace, side, continue partial agent/submit.go, agent/stream.go submit_test.go, submit_safety_test.go
Stop: wraps the turn up with its partial text, closed-out tool calls and a recorded reason partial agent/stop.go stop_test.go
Compaction by input tokens, pair-safe boundary, context-length retry partial agent/overflow.go, agent/tree/compact.go, agent/types/compact_clear.go overflow_test.go, compact_test.go, compact_clear_test.go
Gate composition ranks deny over approval over allow implemented agent/types/gate.go gate_test.go

Structured output and escalation

Item Status Key files Tests
OutputMode native, tool (final_answer), and prompt implemented agent/output.go output_test.go, output_check_test.go, response_schema_test.go
Structured[T] with recursive validation, caller Validate, and Repair implemented agent/structured.go structured_test.go
OutputAuto respects capabilities withdrawn by the adapter partial agent/agent.go (checkStructuredOutput), agent/types/capabilities.go capabilities_test.go, output_check_test.go
OutcomePolicy and escalation ladders recorded in the tree implemented agent/outcome.go, agent/types/outcome.go outcome_test.go, router_pin_test.go

Delegation, interrupts, and durable runs

Item Status Key files Tests
Child approvals forwarded to the root stream with stable interrupt IDs implemented agent/interrupt.go, agent/types/interrupt.go interrupt_test.go, delegation_test.go, interrupt_capacity_test.go
ResolveMarkerErr and idempotent replies, ErrMarkerResolved implemented agent/stream.go, agent/interrupt_router.go interrupt_router_test.go, stream_test.go
Interrupt expiry: deny, fail, escalate implemented agent/interrupt.go interrupt_test.go
Spawn mode with await_subagent, search_subagent, read_subagent implemented agent/spawn.go, agent/subagent_tools.go spawn_test.go
Context modes: task-only, fork, filtered implemented agent/subagent_context.go subagent_context_test.go
Sub-agent timeouts and response schemas implemented agent/subagent.go, agent/subagent_result.go subagent_result_test.go, subagent_schema_test.go
Handoff round trips implemented agent/handoff.go, agent/handoff_context.go handoff_test.go
Local durable engine suspends on child approval and resumes after Decide implemented agent/durable/local/local.go, router.go, input.go local_test.go, router_test.go, lifecycle_test.go, input_test.go
anchor engine: replayed steps, parked approvals, idempotent replies, reconciliation implemented agent/durable/anchor/anchor.go, runner.go, interrupt.go anchor_test.go, runner_test.go, router_test.go, postgres_test.go
Clarification tool (ask_user) as an interrupt implemented agent/clarify.go interrupt_test.go

Providers and routing

Item Status Key files Tests
OptionsProvider and tool choice on every adapter (Ollama emulates named and none) implemented agent/provider/*/options.go, agent/provider/*/tools.go, agent/types/options_provider.go options_test.go, tools_test.go, agent_toolchoice_test.go
Error classification, including Gemini API_KEY_INVALID as auth and RetryInfo delay partial agent/types/errors.go, agent/provider/google/adapter.go errors_test.go, google/errors_test.go
Retry without stacked SDK retries, Retry-After honored implemented agent/provider/retry/retry.go retry_test.go, hardening_test.go
Classified fallback implemented agent/provider/fallback/fallback.go fallback_test.go, classified_test.go
Router with failover, sticky and affinity policies, route locks implemented agent/provider/router/router.go, state.go router_test.go, state_test.go
Traffic splits with sticky assignment, canaries, budgeted shadow arms implemented agent/provider/split/ split_test.go
Response cache with a persisted codec implemented agent/provider/cache/cache.go, codec.go, record.go hardening_test.go, options_test.go
Model catalog and live model listing partial agent/provider/catalog/ discovery_test.go
OpenAI Responses API adapter implemented agent/provider/openai/responses.go responses_test.go
Decorator conventions (Unwrap, As, Members, Innermost) implemented agent/provider/wrapper/ wrapper_test.go, wrapper_conformance_test.go
provider.Build factory implemented agent/provider/provider.go, generate.go provider_test.go
Prompt and context caching implemented anthropic/prompt_cache.go, google/context_cache.go prompt_cache_test.go, context_cache_test.go

Safety controls

Item Status Key files Tests
Placeholder privacy vault, ToolRedactor, privacy.Provider, stream restore implemented agent/privacy/ privacy_test.go, tool_redaction_test.go
Capability gate and per-tool quotas implemented agent/types/capability_gate.go, agent/types/tool_quota.go tool_quota_test.go
Deferred tool disclosure through tool_search implemented agent/selector/ selector_test.go, deferred_test.go
Skills: hashed snapshots, allowed-tools narrowing only implemented agent/skills/ skill_test.go, policy_test.go, tools_test.go, frontmatter_test.go
Host-scoped memory with approval-marked writes and sensitive-content check implemented agent/memory/ memory_test.go
Content-addressed workspace and Spill for large results implemented agent/workspace/ workspace_test.go
fs pack: root confinement, symlink checks, approval for writes implemented tools/fs/, tools/internal/ fs_test.go
exec pack: sandbox, command and network policy, always approval-marked implemented tools/exec/ exec_test.go, proc_unix_test.go
fetch pack: blocks private, loopback, and metadata addresses, including redirects implemented tools/fetch/ fetch_test.go
MCP client: pooled sessions, safe retries, catalog drift, capability gate implemented agent/mcp/ connection_test.go, e2e_test.go
saige-mcp approval for mutating tools (elicit, host, deny) implemented cmd/saige-mcp/bridge.go bridge_test.go
saige-mcp streamable HTTP: bearer tokens, per-token rate limit, TLS or loopback bind implemented cmd/saige-mcp/http.go http_test.go
saige-mcp agent tool: inner marked calls decided like direct calls implemented cmd/saige-mcp/agenttool.go agenttool_test.go
Research tools path safety implemented tools/research/safepath.go tools_test.go, longline_test.go

Storage and wire format

Item Status Key files Tests
Tree state persisted through the store (branches, feedback, active branch) implemented agent/tree/persist.go, agent/tree/tree.go persist_test.go, agent_store_test.go
Tenant scope and version conflicts in pgstore and memstore implemented agent/pgstore/, agent/store/memstore/ scope_test.go, conformance_test.go, agent/store/storetest/
WAL recovery implemented agent/store/walrecover/, agent/store/filewal/ walrecover_test.go, recover_internal_test.go
Versioned delta wire format and tree format version implemented agent/types/wire.go, agent/tree/serialize.go wire_test.go, wire_toolargs_test.go, serialize_test.go, codec_test.go
Remote stream consumption (NewRemoteStream) and AG-UI mapping implemented agent/stream.go, agent/agui/ stream_test.go, agui_test.go
Postgres migrations stop on first failure, no forced sslmode=disable implemented postgres/migrate.go, postgres/pool.go migrate_test.go, pool_test.go

RAG and knowledge graph

Item Status Key files Tests
SyncSource replaces content at the same source URI implemented rag/internal/pipeline/sync.go, rag/types/sync.go sync_test.go
Scoped retrieval across stores and retrievers implemented rag/types/filter.go, rag/*/scope_test.go scope_test.go, filter_test.go
Rank fusion (RRF, weighted) and BM25 contract implemented rag/fusion/, rag/bm25retriever/ fusion_test.go, contract_test.go
Parent-context expansion partial rag/parentretriever/retriever.go, rag/chunker/chunker.go retriever_test.go, chunker_property_test.go
Embedding purpose, batching, retry, validation implemented rag/embedderregistry/ purpose_test.go, retry_internal_test.go, validation_test.go
Source size limits and modified times implemented rag/source/limits.go, rag/source/http.go limits_test.go, modified_test.go
Filesystem source skips tool dirs, dot paths, secret names, and ignored paths; skipped URIs are never pruned implemented rag/source/filesystem.go, rag/source/ignore.go, rag/internal/pipeline/sync.go filter_test.go, sync_test.go
Extractor fuzzing implemented rag/extractor/ fuzz_test.go
Observability for the pipeline and graph implemented rag/otel/, rag/knowledge/observe.go, rag/internal/pipeline/observe.go otel_test.go, observe_test.go
Knowledge graph episode provenance and document deletion implemented rag/knowledge/pgstore/, rag/knowledge/internal/engine/ graph_test.go, ingest_test.go

Evaluation

Item Status Key files Tests
Thresholds and assertions set pass flags, Wilson intervals implemented eval/assert.go, eval/check.go, eval/group.go assert_test.go, check_test.go, group_test.go
Regression comparison with McNemar test implemented eval/compare.go, eval/analysis/ compare_test.go, analysis_test.go
Results store (memory and JSONL files) implemented eval/store/ eval/store/storetest/
Trajectory scorers from a real agent run implemented agent/eval/trajectory.go, agent/eval/collect.go trajectory_test.go, collect_test.go
Judge with structured output implemented eval/judge.go, agent/eval/generator.go judge_test.go, judge_structured_test.go
Text quality scorers (token F1, ROUGE-L) partial eval/textquality.go textquality_test.go
Manifest validation with JSON pointers implemented eval/harness/manifest.go, cmd/saige/eval_validate.go manifest_test.go, eval_run_test.go

Observability

Item Status Key files Tests
Run spans (invoke_agent), cache usage and outcome metrics, redaction implemented agent/otel/ option_test.go, provider_test.go, tool_test.go, metrics_test.go

CLI and servers

Item Status Key files Tests
saige serve: sessions, resumable SSE, approve, cancel, host check implemented cmd/saige/serve.go, serve_server.go serve_test.go
JSON output for ask and chat implemented agent/tui/output_json.go, cmd/saige/ask.go ask_test.go
Strict approval prompt (only y/yes or n/no, end of input denies) implemented agent/tui/approval.go, agent/tui/runner.go approval_test.go, runner_test.go
saige rag against pgvector partial cmd/saige/rag.go rag_pg_test.go
saige update with checksum verification, install.sh implemented cmd/saige/update.go, install.sh update_test.go
saige models catalog view implemented cmd/saige/models.go provider_test.go

Breaking changes and migration

Behavior changes are also listed in upgrade notes.

Change Migration
Anthropic and OpenAI chat adapters set SDK retries to 0. Wrap with retry.New(adapter, retry.DefaultConfig()), or pass WithMaxRetries(n) to a bare adapter.
fallback.New falls back on every error except cancellation, invalid model config, invalid request, and budget errors. Set FallbackOn: types.IsTransient to keep transient-only fallback.
Decorators reject a response schema or request options the inner provider cannot enforce (ErrSchemaUnsupported, ErrOptionsUnsupported). Use a provider that implements StructuredOutputProvider or OptionsProvider, or drop the schema or option.
A second Invoke or Continue on a busy branch fails with ErrRunActive. Use Agent.Submit or EventStream.Submit to queue or steer into the active run.
Required and named tool choices apply to one turn, then revert to auto. Set the choice again if a later turn must be forced.
Malformed tool arguments return an "invalid tool arguments" result instead of reaching the tool. A truncated tool call fails the turn with ErrResponseTruncated. Treat these as turn failures, not empty answers.
A repeated decision for an answered marker returns ErrMarkerResolved; saige serve answers 409. Prefer ResolveMarkerErr over ResolveMarker and treat 409 as already decided.
agent.Config.Store without a Tree builds a store-backed tree, and failed store writes fail the run. Build caller-supplied trees with tree.WithStore(store).
Compacted branch IDs no longer nest (compact-main-<id>). Update code that parses branch names.
WithTracing opens an invoke_agent span per run. Remove a manual StartAgent around Invoke to avoid duplicate spans.
pgstore reads are scoped to the conversation; SaveNode returns ErrVersionConflict and ErrConversationMismatch. Handle the errors instead of relying on silent skips.
postgres.NewPool no longer forces sslmode=disable. New migrations add columns and tables. Set sslmode explicitly and run migrations before deploying.
The DBOS backend is removed. Use the local engine or the anchor adapter.
eval.ToolCallRecord gains ID, ArgumentsError, and Exec; tool_success_rate counts argument errors and unfinished calls as failures. Re-baseline stored results that relied on the old rate.
Google embeddings send a retrieval task type from the embed purpose. Re-index if query and document vectors must match an older index, or fix the type with WithTaskType.
CLI: saige ask --raw is removed; install.sh runs under bash; ask denies approval-marked tools by default. Use --template minimal, pipe the installer to bash, and pass --approve=allow when unattended writes are intended.

Live validation

Scenarios were run against live providers on top of the unit and race tests. Where a scenario was re-run after fixes, the table shows the final result and notes the change.

Scenario Provider Result Evidence
Parallel tool-call pairing OpenAI gpt-4o-mini pass 2 calls in turn 1, each result matched its own call ID's arguments
Parallel tool-call pairing Ollama qwen3.5:4b pass Same pairs as OpenAI, 2 provider calls
Delta codec round trip OpenAI, Ollama pass 36 and 29 live deltas re-encoded byte-identical, 0 failures
Tool panic contained OpenAI, Ollama pass Wait()=nil, ToolExecEnd.Error="tool explode panicked: boom-from-tool"
Truncated JSON arguments rejected OpenAI, Ollama pass "invalid tool arguments: unexpected end of JSON input", tool not run; OpenAI retried and succeeded
Schema-mismatched arguments rejected OpenAI, Ollama pass "property \"city\" must be string, got number", 0 executions
Named tool choice forced for one turn OpenAI pass calls=[ping] on turn 1, auto on turn 2
Named tool choice Ollama qwen3.5:4b pass (after fix) Was rejected with "tool_choice: not declared supported". Now the first turn sends only the named tool, calls=[ping], err=nil. Emulation filters tools and cannot force a call
Required tool choice OpenAI pass Forced once, then reverted
Required tool choice Ollama pass Rejected before any request, as documented
Stop at tools OpenAI, Ollama pass StopToolCallID set, result "submitted:Oslo", 1 provider call
MaxIterForceFinal OpenAI, Ollama pass 1 tool run, final tool-free call; default policy returns "max iterations reached"
Structured[T] native OpenAI, Ollama pass response_format / format schema sent, 1 attempt, typed value decoded
Structured[T] with tools (final_answer) OpenAI, Ollama pass Auto chose tool mode, value used the tool result
Structured[T] repair after Validate rejects OpenAI, Ollama pass 2 attempts, second answer accepted
OutcomePolicy escalation weak to strong Router (Ollama, OpenAI) pass Switch on schema_invalid logged, RouteDelta emitted, OpenAI answer accepted
Escalation ladder exhausted Router (Ollama) pass errors.Is(err, ErrSchemaInvalid)=true after 2 attempts
OutputAuto on a model that rejects forced tool choice Anthropic claude-sonnet-5-5 unit tested Auto picks native output: the adapter sends the schema as output_config.format for a model that declares forced_tool_choice: false
Delegate with child approval via ResolveMarkerErr OpenAI, Ollama pass Child marker on parent stream at depth 0, approved, 1 write, no hang
Durable suspend, Decide, resume OpenAI, Ollama pass First run ErrSuspended with 1 interrupt; resume completed with 1 write
Spawn and await_subagent OpenAI, Ollama pass Handle returned at once, awaited result "Canberra"
search_subagent over child transcript OpenAI pass Returned the matching assistant line
Fork context vs task-only OpenAI pass Fork saw the earlier code word, task-only did not
Handoff round trip OpenAI pass triage to billing to triage, "TRIAGE DONE"
Handoff round trip Ollama qwen3.5:4b pass (2 of 3) One run repeated handoffs until MaxIter stopped it cleanly; model behavior
Queue mid-run OpenAI pass Held until the first answer finished, then answered "BANANA"
Steer between tool calls OpenAI pass Landed after the tool result, pairing intact, model followed it
Interrupt-replace OpenAI pass Partial text committed with TruncationPart, new message answered
Continue after output-limit truncation OpenAI pass Continued without repeating text
Token-based compaction OpenAI pass Branch moved to compact-main-*, summarized fact recalled; input stayed above the limit after one pass
Router: permanent error does not fail over OpenAI to Ollama pass 404 invalid request, 0 fallback requests
Router: transient 503 fails over OpenAI to Ollama pass Route reason failover, answer from Ollama
Retry count without SDK stacking OpenAI, Anthropic, Google (forced errors) pass HTTP requests equal MaxAttempts; Retry-After: 1 honored
Forced Google 429 Google adapter pass kind=rate_limit, retryAfter=7s from RetryInfo; header-only Retry-After reads 0
Invalid Google key classified as auth Google gemini-2.5-flash pass (after fix) Was InvalidRequest; now kind=auth, also through privacy.Provider
Split sticky assignment OpenAI, Ollama pass Same arm across turns and split instances, RouteDelta carries experiment and variant
Response cache across instances OpenAI pass 1 HTTP request for 2 calls, CacheHit=true, metadata preserved
Model listing Ollama, OpenAI pass 13 and 133 models listed
Privacy: ToolRedactor with stream restore OpenAI pass Requests held only <<EMAIL_1>>, tool received the real address, stream restored
Privacy: shared vault with privacy.Provider OpenAI pass No real address in any request body
Skills: load and allowed-tools narrowing OpenAI pass Tools narrowed after load; delete_note ran 0 times
Memory: approval-gated write and recall OpenAI pass Deny wrote nothing; approve stored the record; recall answered "Teal"; SSN content refused
Workspace spill Ollama pass 60 KB result stored, 2 KB preview sent, tail sentinel never reached the provider
fs path traversal OpenAI and direct calls pass .., absolute paths, and symlinks rejected; nothing written outside the root
exec approval and policy OpenAI pass Deny ran nothing; network blocked by default; sudo denied by policy
fetch metadata address blocking direct calls pass Decimal, hex, IPv6-mapped, DNS, and redirect forms all blocked
saige-mcp approval for store_knowledge MCP stdio client pass Decline, missing elicitation, and --approval=deny all refused; accept ran once
RAG ingest and hybrid search Ollama nomic-embed-text pass Vector and BM25 hits returned
Re-ingest at the same URI via SyncSource Ollama pass Document UUID kept, old text gone
saige rag ingest and search Ollama + pgvector pass (after re-run) First run had no Postgres; with pgvector, 2 hits ranked correctly
eval.Run gates and Wilson intervals Ollama pass 3/3 interval [0.439, 1.000]; failing case flagged
Regression comparison Ollama pass 2 regressions found, also after reload from the file store
Trajectory scorers Ollama pass All trajectory scores 1.0 on a real run
Judge with structured output OpenAI pass Schema path used, scores 1.0 and 0.0 as expected
saige eval validate offline pass Field-specific errors with JSON pointers, exit code 2
go build ./cmd/..., saige models offline pass Both binaries built; catalog, JSON, and unknown-model output correct
saige ask and chat (text, stdin, JSON) OpenAI pass JSONL envelopes on stdout, prompts on stderr
Approval prompt input validation OpenAI pass Empty and invalid input re-prompt; end of input denies
saige serve SSE, approve/deny, cancel, delete OpenAI pass Resume by Last-Event-ID, 409 on repeated decision, 415 and 403 checks
saige-mcp start and tool listing offline pass Read-only tools listed, invalid regex returned isError=true
Parallel tool-call pairing Anthropic claude-haiku-5-5 pass Each result paired with its own call ID
Named and required tool choice Anthropic claude-haiku-5-5 pass {"type":"tool"} and {"type":"any"} accepted with adaptive thinking on, forced for one turn, then absent
Stop at tools Anthropic claude-haiku-5-5 pass StopToolCallID set, 1 provider call
MaxIterForceFinal Anthropic claude-haiku-5-5 pass 1 tool run, final tool-free call
Structured[T] auto, with tools, and repair Anthropic claude-haiku-5-5 pass Auto without tools used the forced structured_output tool; with tools it chose final_answer with tool_choice absent; repair accepted the second answer
Delegate with child approval Anthropic claude-haiku-5-5 pass Child marker approved, 1 write, no hang
Durable suspend, Decide, resume Anthropic claude-haiku-5-5 pass First run ErrSuspended, resume completed
Spawn and await_subagent Anthropic claude-haiku-5-5 pass Handle returned at once, result awaited
Handoff round trip Anthropic claude-haiku-5-5 pass Control returned to the first agent
Continue after output-limit truncation Anthropic claude-haiku-5-5 pass Continued through the continue-prompt path, since the model accepts no prefill
Submit queue and InterruptReplace Anthropic claude-haiku-5-5 pass Queued message answered after the run; replace committed the partial turn
Prompt caching Anthropic claude-haiku-5-5 pass First call wrote 2643 cache tokens, the second read 2643
web_search server tool deltas Anthropic claude-haiku-5-5 pass server tool call and result parts streamed
Preset failover on 529 Anthropic to OpenAI pass Forced 529 failed over to OpenAI; route names named the vendor and model, not the decorators
ToolRedactor Anthropic claude-haiku-5-5 pass Request bodies held only placeholders
saige ask and --format json Anthropic claude-haiku-5-5 pass Text answer, and JSONL envelopes ending in usage and done
Empty system prompt Anthropic claude-haiku-5-5 pass (after fix) Was 400 "text content blocks must be non-empty"; system is now omitted when no text is left
Forced tool choice on an adaptive model Anthropic claude-haiku-5-5 pass (after fix) Was rejected locally because thinking is on by default; only a manual thinking budget, or a model that declares forced_tool_choice: false, now rejects it
Tool call on Chat Completions OpenAI gpt-6-luna pass The adapter sent reasoning_effort: "none" with the tools; parallel_tool_calls: false accepted
Tool call through the Responses API OpenAI gpt-6.1-sol pass provider.Build served the model on /v1/responses; tool ran once; parallel_tool_calls: false accepted
One-call smoke OpenAI gpt-6-luna, gpt-6.1-sol pass Answered "ok" on Chat Completions and Responses
saige ask default and openai, openai-quality presets Anthropic claude-haiku-5-5, OpenAI pass The default preset ran claude-haiku-5-5; both OpenAI presets answered
Text answer Google gemini-3.1-flash-lite, gemini-3.8-flash pass Answered "ok", 1 request each
Tool call Google gemini-3.1-flash-lite, gemini-3.8-flash pass (after fix) Was 400 "Function call is missing a thought_signature" on the second turn. The adapter now returns the signature Gemini puts on the function call part; tool ran once, 2 requests
Structured[T] auto Google gemini-3.1-flash-lite, gemini-3.8-flash pass Native responseSchema path, 1 request, typed value decoded
saige ask --preset google and google-quality Google pass gemini-3.1-flash-lite and gemini-3.8-flash answered
Vertex AI text, tool call and Structured[T] Google gemini-3.1-flash-lite on Vertex AI pass GOOGLE_GENAI_USE_VERTEXAI=true with the project and global location from the environment and no API key; Application Default Credentials attached; tool ran once; native schema path
Vertex AI embeddings Google gemini-embedding-001 on Vertex AI pass WithEmbedVertex, 3072 dimensions
saige --provider vertex ask Google gemini-3.1-flash-lite on Vertex AI pass Answered "ok" with no Gemini API key set
Live Google 429 Google not exercised No cheap way to trigger one; covered by the forced 429 scenario

Remaining gaps and known issues

Issue Where Effect
Plain Ingest of changed content at the same URI creates a second document. rag/internal/pipeline/pipeline.go Documented: use SyncSource or Update for replace-by-URI.
A reply that lands while its anchor run is still parking waits for the worker's janitor. agent/durable/anchor/interrupt.go The run resumes on the next janitor pass instead of at once; set worker.WithJanitorEvery to bound the delay.
Spawn mode is unavailable under durable runners. agent/spawn.go By design: returns ErrSpawnUnsupported.