refactor(errors): one failure-copy table and one budget matcher

[senamakel]

Design note

Current state. User-facing failure copy was spread over web_chat/web_errors/{classify,budget,timeout,backend_error_code,retry,response_predicates}.rs, the loop-guard halt text (agent/tinyagents/middleware/{loop_guards,repeated_failure}.rs), and the budget-phrase matcher existed three times with deliberately different semantics: a regex list in web_chat, a whole-word list in the triage evaluator (agent/triage/evaluator/arm.rs), and the billing list in tinyinference_providers used by the loop guards. The error_type wire tokens were written inline at ~35 construction sites.

Target contract.

1. Provider-text matchers move upstream (tinyinference, nested under tinyagents; PRs feat(llm,providers): one budget-phrase matcher with strictness modes; status-word failure predicates tinyinference#48, chore(deps): bump tinyinference for the shared budget matcher and failure predicates tinyagents#258): BudgetMatch::{Billing, Managed, Strict} + is_budget_message in tinyinference-providers, and the status-word predicates (is_rate_limit_text, is_timeout_text, is_auth_error_text, is_payment_required_text, is_server_error_text, is_context_length_text, is_model_unavailable_text, is_vision_unsupported_text, is_codex_token_expired_text, is_recoverable_failure_text) in tinyinference-llm::failure.
2. ONE host copy table, inference/failure_copy/table.rs: FailureClass -> FailureCopy { error_type, source, retryable, copy }. web_errors/classify.rs and backend_error_code.rs are now just the ladder (which class?) and build the envelope from the table. Loop-guard halt summaries and the user-actionable escalation live in inference/failure_copy/halt.rs.
3. ONE budget matcher. The three lists are three modes of the same upstream function, because they are not equivalent and a union would change classifications:

difference	web_chat (Managed)	triage (Strict)	loop guard (Billing)
normalisation	lowercase, -/_/whitespace runs -> one space	lowercase, every non-alphanumeric -> space	ASCII lowercase
matching	regex/substring, budget.*exceed, add.*credits gap patterns	whole-word sequences only	substring
top up	substring (so stop updating matches)	whole word (does not match)	no
out of credits	yes	yes	no
budget exceeds	yes via budget.*exceed	yes	no
insufficient budget / insufficient balance / credit balance is too low	yes (OR Billing)	no, deliberate: false positive hides a real fatal from Sentry	yes
add credits, budget exceeded, no remaining credits	yes	yes	yes

Each consumer keeps its own mode, so no classification changes. A corpus test compares each mode against a verbatim copy of its legacy implementation.

Stays host-side. The ladder order and which class an error maps to, the OpenHuman-own action-budget strings (SecurityPolicy), the turn-timeout marker, the harness-generated delegated-inference envelope markers, thresholds, and the Outcome/event plumbing.

Compatibility. Every user-facing string and every error_type / source / retryable value is byte-identical. Proof: web_chat/web_tests_error_copy_fixture_tests.rs has literal rows (input -> error_type, source, retryable, retry_after_ms, provider, fallback_available, message) for 37 inputs covering every class and every managed errorCode, captured from the pre-refactor binary before any code moved, plus literal halt-summary strings and the legacy-oracle budget corpus test. No persisted format is touched.

Frontend. app/src does not translate this copy: ChatRuntimeProvider renders chat_error.message verbatim and only branches on error_type (session_expired, cancelled, guardrail). No i18n added here. Follow-up: have the core emit the stable error_type plus params and let the frontend own localized text; the table in failure_copy/table.rs is the single place to key that from.

Changes

• New inference/failure_copy/{mod,table,halt}.rs; classify.rs/backend_error_code.rs rebuilt on the table; budget.rs is a thin accessor over the shared matcher; response_predicates.rs deleted.
• Triage is_inference_budget_exceeded and loop-guard budget detection call is_budget_message; loop_guards.rs keeps only thresholds and the repeat-call exemption.
• vendor/tinyagents gitlink -> cc33ca71 (tinyagents#258, which bumps tinyinference#48).
• Core src non-test lines: +723 / -853 vs base (the table is moved, not removed; net -130), plus ~300 lines of fixture tests.

Verification

• cargo test -p openhuman --lib --features "$FEAT": 9093 passed, 3 failed, all pre-existing: credential_scrub::browser_confirmation_token_survives_without_exempting_page_credentials, composio::...connected_integrations_cache_is_keyed_per_credential, and flows::tinyflows::caps::tools::composio::...backend_dispatch_forwards_the_workflow_connection_id (flaky shared composio state; passes alone).
• cargo check --tests -p openhuman -p openhuman-cli --features "$FEAT" and cargo check -p openhuman-embed -p openhuman-tinyhumans --tests: clean.
• clippy on touched files: clean (one pre-existing unused import in flows/tinyflows/caps/mod.rs remains on base).
• check-submodule-monotonic upstream/main HEAD: OK. rust:layout, check-agent-runtime-boundary: only the pre-existing entries. check-feature-forwarding, gated-test allowlist: OK.
• cargo check --manifest-path crates/openhuman-app/Cargo.toml fails in dependency resolution (plist via starship-battery/tinymemory-gate), unrelated to this change. check-module-pins could not run (shallow submodule tags).
• Merge order: tinyinference#48, then tinyagents#258, then this.

Summary by CodeRabbit

• Improvements
  • Standardized inference error messages and classifications across chat responses, including clearer retry guidance for rate limits and billing issues.
  • Improved recognition of budget-related errors and clearer explanations when an agent cannot continue.
  • Added consistent summaries for repeated or stalled agent actions and prompts for errors that may require a connection.
• Tests
  • Added coverage for error messages, retry details, budget detection, and agent halt summaries.

[tinysweeper]

Tiny Sweeper review

⚠️ Review failed for fbdc3ab648cf. pull request exceeds review limits: 387 changed files (limit 500), 59011 changed lines (limit 50000)