90 lines
3.6 KiB
Markdown
90 lines
3.6 KiB
Markdown
|
# Sprint 171 Plan: Tool-Chained Autonomous Remediation to Green
|
||
|
|
|
||
|
|
## Context
|
||
|
|
|
||
|
|
With improved generation and real gates, the final requirement is robust autonomous
|
||
|
|
remediation: when output fails, the system should chain generation + debug/fix tools
|
||
|
|
until gates are green (or a clear, auditable stop reason is reached).
|
||
|
|
|
||
|
|
This sprint connects coding tools to remediation/debug tooling so failed outputs
|
||
|
|
trigger deterministic repair workflows rather than manual intervention.
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## Goals
|
||
|
|
|
||
|
|
1. Chain generation failures into debug/fix tool workflows automatically
|
||
|
|
2. Use taskitem + remediation planning to drive next-best tool calls
|
||
|
|
3. Preserve full audit trace of each remediation step
|
||
|
|
4. Achieve repeated benchmark green completion without manual prompt edits
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## Steps
|
||
|
|
|
||
|
|
### Step 1899: Remediation router from gate diagnostics to tool actions (10 tests)
|
||
|
|
|
||
|
|
Map normalized diagnostics to repair actions/tool choices:
|
||
|
|
- signature/type errors -> AST mutation/codegen refinement
|
||
|
|
- missing methods/fields -> class/data-model completion actions
|
||
|
|
- test behavior failures -> targeted body patch plan
|
||
|
|
|
||
|
|
Tests (10): diagnostic->action mapping coverage, deterministic action ordering,
|
||
|
|
priority handling, unsupported-category fallback, no-op on green state, confidence
|
||
|
|
scoring for routes, stable serialization, conflict resolution, retry budget policy,
|
||
|
|
tool argument schema validation.
|
||
|
|
|
||
|
|
### Step 1900: Debug tool integration path for failed generations (10 tests)
|
||
|
|
|
||
|
|
Integrate existing debug-oriented tools into remediation loop where relevant:
|
||
|
|
- diagnostic fetch, quick-fix application, build/test output parsing
|
||
|
|
- replayable fix attempts with rollback metadata
|
||
|
|
|
||
|
|
Tests (10): debug tool invocation path, quick-fix roundtrip, build-parse integration,
|
||
|
|
rollback metadata capture, failure classification, deterministic trace format,
|
||
|
|
tool permission safety, no infinite retry loops, non-recoverable failure stop,
|
||
|
|
structured stop reason output.
|
||
|
|
|
||
|
|
### Step 1901: Autonomous loop controller with “green or explicit blocked” contract (12 tests)
|
||
|
|
|
||
|
|
Controller semantics:
|
||
|
|
- iterate generate -> gate -> remediate
|
||
|
|
- finish only on all-green OR explicit blocked status with complete evidence
|
||
|
|
|
||
|
|
Tests (12): green completion path, blocked path with reason, mixed-failure handling,
|
||
|
|
budget exhaustion semantics, state checkpointing, resumable runs, deterministic
|
||
|
|
iteration IDs, artifact retention, status schema validity, strict-mode compliance,
|
||
|
|
benchmark repetition stability, no silent partial success.
|
||
|
|
|
||
|
|
### Step 1902: Benchmark suite for production capability claim (8 tests)
|
||
|
|
|
||
|
|
Run repeated benchmark batch (minimum N runs) for:
|
||
|
|
- PriorityQueue
|
||
|
|
- one additional structured workload
|
||
|
|
|
||
|
|
Collect pass rate, average iterations, mean/95p latency, failure taxonomy.
|
||
|
|
|
||
|
|
Tests (8): batch runner works, repeated-run summary generated, pass-rate field
|
||
|
|
accurate, latency stats present, failure taxonomy non-empty when failures exist,
|
||
|
|
trace links preserved, deterministic IDs, report schema validity.
|
||
|
|
|
||
|
|
### Step 1903: Sprint 171 Integration Summary (8 tests)
|
||
|
|
|
||
|
|
Add `editor/src/Sprint171IntegrationSummary.h`.
|
||
|
|
Record: steps_completed=5 (1899-1903), remediation_router_active=true,
|
||
|
|
debug_tool_chaining_active=true, autonomous_green_or_blocked_contract=true,
|
||
|
|
production_capability_benchmarked=true, success=true.
|
||
|
|
|
||
|
|
Tests (8): constructable summary struct, steps_completed==5, router active,
|
||
|
|
debug chaining active, contract enforced, benchmark report present, regression
|
||
|
|
suite pass, success==true.
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## Architecture Gate
|
||
|
|
|
||
|
|
- No silent “success” without green gates
|
||
|
|
- Non-green exits must include explicit blocked reason + full evidence trail
|
||
|
|
- Remediation routing must stay deterministic and auditable
|
||
|
|
|