Make agent tool handling schema-driven and recoverable #72

New Issue

hugo · 2026-08-29T17:40:10Z

hugo commented

2026-08-29 17:40:10 +00:00

Goal

Make DS4Server tool use reliable for local coding models through Rust-native executable schemas, precise failure feedback, and targeted loop recovery. Tool metadata must have one source of truth shared by prompt rendering and runtime validation.

Plan

Define each tool's name, description, parameter contract, validation, and execution behavior together in Rust; generate the model-visible schema from that contract instead of maintaining separate prompt-only JSON.
Validate required fields, primitive types, enums, numeric bounds, and cross-field rules before any side effect. Remove silent numeric/boolean coercion and resolve existing schema/runtime disagreements such as optional list.path and job-versus-pid lookup.
Return stable, precise failures containing the tool name, error code, field path, expected shape, and received type/value. Keep malformed transport syntax, unknown tools, policy denials, execution failures, and timeouts distinguishable.
Keep DeepSeek DSML and GLM tool syntax inside their Rust model adapters. Reduce the global prompt to transport rules and move invocation/recovery guidance into the owning tool descriptions or prompt sections.
Add a per-session exact-repeat detector over canonical tool name/arguments. Inject advisory reminders at escalating thresholds (initially 3, 5, and 8), count denied and failed calls, and reset after a direct user message. Do not rewrite calls or retry side effects automatically.
Add focused tests/evals for malformed syntax, missing/wrong arguments, enum/range violations, unknown tools, denied actions, execution errors, repeated calls, and successful model correction on the following step.

Boundaries

Rust only; no JavaScript runtime, workflow DSL, or external helper process.
This issue does not implement Ralph, subagents, or generic workflows.
Do not preserve current prompt or coercion behavior without evidence; retain only model-wire requirements demonstrated by DSML/GLM tests.

Acceptance criteria

Model-visible schemas and runtime validation are produced from the same Rust definitions.
Invalid arguments cannot reach tool bodies or cause side effects, and feedback identifies the exact tool and violation.
Existing schema/runtime disagreements and silent coercions are removed or explicitly justified by tests.
Tool-specific guidance covers failure recovery, shell exit handling, sandbox denial behavior, editing, and background jobs without duplicating a large global prompt.
Exact repeated calls trigger bounded, escalating notices without blocking legitimate execution.
DeepSeek V4 Flash and GLM tool-call parsing and correction paths have regression coverage.
cargo fmt --all -- --check, Clippy with -D warnings, make bundle, and cargo test --all-features pass.

## Goal Make DS4Server tool use reliable for local coding models through Rust-native executable schemas, precise failure feedback, and targeted loop recovery. Tool metadata must have one source of truth shared by prompt rendering and runtime validation. ## Plan - Define each tool's name, description, parameter contract, validation, and execution behavior together in Rust; generate the model-visible schema from that contract instead of maintaining separate prompt-only JSON. - Validate required fields, primitive types, enums, numeric bounds, and cross-field rules before any side effect. Remove silent numeric/boolean coercion and resolve existing schema/runtime disagreements such as optional `list.path` and job-versus-pid lookup. - Return stable, precise failures containing the tool name, error code, field path, expected shape, and received type/value. Keep malformed transport syntax, unknown tools, policy denials, execution failures, and timeouts distinguishable. - Keep DeepSeek DSML and GLM tool syntax inside their Rust model adapters. Reduce the global prompt to transport rules and move invocation/recovery guidance into the owning tool descriptions or prompt sections. - Add a per-session exact-repeat detector over canonical tool name/arguments. Inject advisory reminders at escalating thresholds (initially 3, 5, and 8), count denied and failed calls, and reset after a direct user message. Do not rewrite calls or retry side effects automatically. - Add focused tests/evals for malformed syntax, missing/wrong arguments, enum/range violations, unknown tools, denied actions, execution errors, repeated calls, and successful model correction on the following step. ## Boundaries - Rust only; no JavaScript runtime, workflow DSL, or external helper process. - This issue does not implement Ralph, subagents, or generic workflows. - Do not preserve current prompt or coercion behavior without evidence; retain only model-wire requirements demonstrated by DSML/GLM tests. ## Acceptance criteria - Model-visible schemas and runtime validation are produced from the same Rust definitions. - Invalid arguments cannot reach tool bodies or cause side effects, and feedback identifies the exact tool and violation. - Existing schema/runtime disagreements and silent coercions are removed or explicitly justified by tests. - Tool-specific guidance covers failure recovery, shell exit handling, sandbox denial behavior, editing, and background jobs without duplicating a large global prompt. - Exact repeated calls trigger bounded, escalating notices without blocking legitimate execution. - DeepSeek V4 Flash and GLM tool-call parsing and correction paths have regression coverage. - `cargo fmt --all -- --check`, Clippy with `-D warnings`, `make bundle`, and `cargo test --all-features` pass.

hugo added the enhancement critical labels 2026-08-29 17:40:10 +00:00

hugo referenced this issue

2026-08-29 17:40:53 +00:00

Add a Rust-native fresh-agent Ralph loop #73

hugo commented

2026-08-29 18:44:28 +00:00

Implemented in f12aafc. Tool names, descriptions, parameter kinds/bounds/enums/cross-field rules, schema generation, and execution dispatch now share one Rust ToolSpec catalog, including the Dev Brain tools. Runtime validation runs before approval or execution, rejects undeclared/missing/wrong/quoted/out-of-range arguments without coercion, and returns stable tool/code/field/expected/received failures. list.path is optional and bash_status/bash_stop accept either job or pid. DeepSeek DSML and GLM adapters preserve declared integer/boolean primitives, malformed transport and unknown/denied/execution/timeout outcomes remain distinct, recovery guidance lives with each tool, and per-session canonical repeat notices fire at 3/5/8 and reset on direct user input. Added isolated coverage for schema/runtime agreement, no-side-effect invalid writes followed by a successful correction, parser typing and malformed syntax, every validation class, policy denial, real execution errors/timeouts, and canonical repeat/reset behavior. Verification: cargo fmt --all -- --check; cargo clippy --all-targets --all-features -- -D warnings; make bundle; cargo test --all-features (176 passed, 14 hardware/browser fixtures ignored).

Implemented in f12aafc. Tool names, descriptions, parameter kinds/bounds/enums/cross-field rules, schema generation, and execution dispatch now share one Rust ToolSpec catalog, including the Dev Brain tools. Runtime validation runs before approval or execution, rejects undeclared/missing/wrong/quoted/out-of-range arguments without coercion, and returns stable tool/code/field/expected/received failures. list.path is optional and bash_status/bash_stop accept either job or pid. DeepSeek DSML and GLM adapters preserve declared integer/boolean primitives, malformed transport and unknown/denied/execution/timeout outcomes remain distinct, recovery guidance lives with each tool, and per-session canonical repeat notices fire at 3/5/8 and reset on direct user input. Added isolated coverage for schema/runtime agreement, no-side-effect invalid writes followed by a successful correction, parser typing and malformed syntax, every validation class, policy denial, real execution errors/timeouts, and canonical repeat/reset behavior. Verification: cargo fmt --all -- --check; cargo clippy --all-targets --all-features -- -D warnings; make bundle; cargo test --all-features (176 passed, 14 hardware/browser fixtures ignored).

hugo closed this issue

2026-08-29 18:44:33 +00:00

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: hugo/DS4Server#72