@piagent/platform 1.2.10 → 1.2.12

package/evals/fixtures/task-contract.valid.json

@@ -1,5 +1,13 @@
 {
-  "taskId": "TASK-0001",
+  "schemaVersion": 2,
+  "taskRunId": "task-0001-20260731120000-a1b2c3d4e5",
+  "taskId": "task-0001",
+  "sessionId": "pi-session-example",
+  "sessionName": "TASK-0001 Example task",
+  "changeMode": "source-change",
+  "attempt": 1,
+  "maxAttempts": 3,
+  "previousAttempts": [],
   "summary": "Replace with one-sentence task outcome.",
   "riskLane": "normal",
   "expectedOutput": "Replace with exact artifact or behavior expected.",
@@ -17,6 +25,7 @@
   "contextManifest": [],
   "memoryCitations": [],
   "mcpCapabilities": [],
+  "verifyGroup": "source",
   "verifyCommands": [
     "Replace with verify command."
   ],
@@ -66,9 +75,16 @@
       "watchdog": "Use a strong model only for final risk review, security, release, or high-impact changes."
     }
   },
+  "baselineChangedFiles": [],
+  "baselineFileDigests": {},
+  "observedChangedFiles": [],
+  "finalWorkingTreeFiles": [],
+  "finalFileDigests": {},
   "changedFiles": [],
   "verifyEvidence": [],
   "trace": {
     "outcome": "pending"
-  }
+  },
+  "createdAt": "2026-07-31T12:00:00.000Z",
+  "updatedAt": "2026-07-31T12:00:00.000Z"
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@piagent/platform",
-  "version": "1.2.10",
+  "version": "1.2.12",
   "license": "MIT",
   "publishConfig": {
     "access": "public",
@@ -30,6 +30,7 @@
     "SECURITY.md",
     ".gitattributes",
     "adapters/",
+    "benchmarks/",
     "catalog/",
     "docs/",
     "!docs/journals/",
@@ -113,7 +114,7 @@
     "install-global": "bash scripts/install-global.sh",
     "init-project": "bash scripts/init-project.sh",
     "doctor": "bash scripts/team-doctor.sh",
-    "benchmark": "bash scripts/quality-benchmark.sh",
+    "benchmark": "node scripts/benchmark-runner.mjs",
     "benchmark:redaction": "node scripts/redaction-benchmark.mjs",
     "audit:runtime": "bash scripts/audit-runtime-host.sh",
     "release:identity": "node scripts/verify-release-identity.mjs",

package/packages/piagent-core/README.md

@@ -6,13 +6,23 @@ Shared Pi package for reusable project workflows.
 
 - `extensions/piagent-guard.ts`: runtime guard tools and policy hooks.
 - `prompts/*.md`: workflow aliases that intentionally launch an agent turn.
-- `skills/piagent-ops/SKILL.md`: operating guidance for implementation tasks.
+- `skills/piagent-ops/SKILL.md`: operator-invoked reference for manual intake, recovery, and high-risk controls; routine runtime-managed tasks do not advertise or load it.
 - `skills/piagent-source-cache/`: local cache for user-provided external source repositories.
 - `subagents/*.md`: piagent roles for `pi-subagents`.
+- `benchmark/benchmark-core.js`: deterministic suite validation, evidence scoring, usage comparison, and report rendering for `piagent-benchmark`.
 - `policies/base-policy.json`: default runtime policy, including protected path and shell protected path defaults.
 - input hook support for local screenshot/image paths pasted into chat; supported images are attached as `[image1]`, `[image2]`, ...
 - compact tool-result rendering: oversized redacted output is previewed in Pi and captured under `.pi/piagent-state/tool-results/` for offline audit/reporting.
-- local Context Engine v2: incremental FTS5/symbol/import index, hybrid retrieval, token-budgeted packs, test-impact mapping, dynamic Piagent tools, and Agent Watch compatible telemetry.
+- local Context Engine v2: incremental FTS5/symbol/import index, hybrid retrieval, token-budgeted packs, bounded current-turn source/test snapshots, test-impact mapping, dynamic Piagent tools, and Agent Watch compatible telemetry.
+
+The root package exposes `piagent-benchmark` for an automatic paired Piagent
+steady-state benchmark against Raw Pi or the `codex-cli` surface. Treatment
+onboarding/context preparation happens before measured model usage;
+post-baseline mutations remain scope violations. Codex runs through a strict
+streaming JSONL adapter and defaults to controlled ephemeral execution with an
+isolated temporary Codex home. Run
+`piagent-benchmark --dry-run` before starting billed model sessions; see
+`docs/quality-benchmark.md` for score and verdict rules.
 
 ## Trusted run wrapper
 
@@ -70,12 +80,20 @@ Legacy aliases still work: `/permission-status`, `/read-only`, `/workspace-write
 - `piagent_profile_tech_context_record`
 - `piagent_project_onboarding_record`
 - `piagent_task_start`
+- `piagent_task_progress`
 - `piagent_source_checkout`
 - `piagent_document_read` — reads `.md`/`.txt`/`.csv`/`.json`/`.yaml`/`.pdf`/`.docx` from the project or a granted `additionalReadRoots` directory; read-only, and `protectedPaths` still wins
 - `piagent_context_record`
 - `piagent_verify_record` — records verify evidence only after matching an observed bash tool result after task start
 - `piagent_trace_record`
 
+These tools are registered capabilities, not a per-task call sequence. Runtime
+starts bounded source tasks directly and exposes no Piagent management schema
+to the model. Broad, high-risk, or ambiguous intake activates
+`piagent_task_start`; lifecycle hooks still collect context, changes,
+current-tree verification, trace and final-gate evidence. Diagnostic/recovery
+groups load only when requested.
+
 ## Runtime commands and workflow recipes
 
 - `/commands`: runtime menu/help for terminal, Pi, MCP, model, memory, session, context, permission, and subagent commands.
@@ -132,6 +150,7 @@ If no trusted profile is available, the extension still applies baseline secret
 Runtime task tools write local state to:
 
 - `.pi/piagent-state/tasks/*.json`
+- `.pi/piagent-state/session-tasks/*.json`
 - `.pi/project-context.md`
 - `.pi/tech-stack.json`
 - `.pi/tech-context/*.json`
@@ -147,12 +166,27 @@ Runtime task tools write local state to:
 
 Project-local state belongs in `.pi/.gitignore`.
 
+Task files use schema v2 and bind one unique `taskRunId` to one Pi `sessionId`.
+They record attempt/max-attempt history, dependency-safe work-plan progress,
+Git baseline/final digests, observed changes, all exact verify results and final
+trace. A terminal task is immutable; retry starts in a fresh session and carries
+the prior failure/ruled-out evidence. Legacy v1 state is migrated and archived
+without requiring project onboarding again.
+
+Local state directories are owner-only. Task state, context telemetry, trace,
+bash evidence and captures refuse ancestor symlinks out of the project. JSONL
+files rotate under a cross-process lock; compacted tool captures are pruned by
+age, count and aggregate bytes.
+
 The Context Engine SQLite index contains allowed source bodies. Its directory is
 kept at mode `0700`, with database/WAL/SHM at `0600`; core and FTS5 secure-delete
 cover normal refreshes, while exclusion-policy changes trigger a retryable FTS
 rebuild and database vacuum before the new policy is considered clean.
 
-Passing final gates require an observed exit `0` command that exactly matches one entry in `task.verifyCommands`. Other observed commands are traceable but advisory.
+Passing source final gates require an observed exit `0` result for every
+meaningful entry in `task.verifyCommands`. Other observed commands are traceable
+but advisory. Declared changed files must also differ from the task baseline,
+match observed tool results and remain inside `task.scope`.
 
 Raw path-like tool access to protected paths is blocked before execution. This includes Pi built-ins (`read`, `write`, `edit`, `grep`, `find`, `ls`) and custom/MCP tools with nested path-like strings, arrays, or `file://` URIs. Path-like strings are percent-decoded once, and input nesting above `MAX_TOOL_INPUT_INSPECTION_DEPTH=32` fails closed. Known content fields such as `content`, `query`, `pattern`, `text`, and `command` are excluded from generic extraction. `grep.glob` and `find.pattern` are checked when they explicitly target protected paths, while broad `grep`, `find`, and `ls` results are filtered so protected content lines or path metadata are redacted before reaching the model.