ultimate-pi 0.2.4 → 0.2.5

package/.pi/extensions/lib/harness-paths.ts

@@ -45,3 +45,11 @@ export function resolveHarnessAsset(
 ): string {
 	return join(getHarnessPackageRoot(moduleUrl), ...segments);
 }
+
+/** Harness CLI scripts shipped under `.pi/scripts/` in the npm package. */
+export function resolveHarnessScript(
+	moduleUrl: string,
+	scriptName: string,
+): string {
+	return resolveHarnessAsset(moduleUrl, ".pi", "scripts", scriptName);
+}

package/.pi/extensions/sentrux-rules-sync.ts

@@ -4,21 +4,15 @@
 
 import { spawn } from "node:child_process";
 import { existsSync } from "node:fs";
-import { join } from "node:path";
 import type { ExtensionAPI } from "@mariozechner/pi-coding-agent";
-import { resolveHarnessAsset } from "./lib/harness-paths.js";
+import { resolveHarnessScript } from "./lib/harness-paths.js";
 
 function resolveSyncScript(): string {
-	const packaged = resolveHarnessAsset(
+	return resolveHarnessScript(
 		// @ts-expect-error pi extensions run as ESM
 		import.meta.url,
-		"scripts",
 		"sentrux-rules-sync.mjs",
 	);
-	if (existsSync(packaged)) {
-		return packaged;
-	}
-	return join(process.cwd(), "scripts", "sentrux-rules-sync.mjs");
 }
 
 function runSync(args: string[]): Promise<{ code: number; output: string }> {

package/.pi/harness/browser.json

@@ -1 +1,5 @@
-{"headless": true, "timeout": 30000, "viewport": {"width": 1280, "height": 720}}
+{
+	"headless": true,
+	"timeout": 30000,
+	"viewport": { "width": 1280, "height": 720 }
+}

package/.pi/harness/debates/README.md

@@ -0,0 +1,9 @@
+# Harness Debates
+
+Store debate artifacts (`RoundResult`, `ConsensusPacket`, budget events) here.
+
+Locked defaults for aggressive budget profile:
+
+- `max_rounds=6`
+- `round_token_cap=2500`
+- `debate_global_cap=35000`

package/.pi/harness/docs/adrs/0006-sentrux-dual-layer.md

@@ -28,4 +28,4 @@ Evaluator trust requires both programmatic gates (policy, budget, integrity) and
 ## References
 
 - `.pi/harness/specs/observation.schema.json`
-- `scripts/harness-verify.mjs`
+- `.pi/scripts/harness-verify.mjs`

package/.pi/harness/docs/adrs/0009-sentrux-rules-lifecycle.md

@@ -11,7 +11,7 @@ Sentrux enforces architecture via [`.sentrux/rules.toml`](https://sentrux.dev/do
 
 1. **Canonical source:** [`.pi/harness/sentrux/architecture.manifest.json`](../../sentrux/architecture.manifest.json) — layers, boundaries, global constraints.
 2. **Generated artifact:** `.sentrux/rules.toml` — committed to git; managed block between `harness:managed:start/end` markers.
-3. **Sync command:** `npm run harness:sentrux-sync` (`scripts/sentrux-rules-sync.mjs`).
+3. **Sync command:** `npm run harness:sentrux-sync` (`.pi/scripts/sentrux-rules-sync.mjs`).
 4. **Pi command:** `/harness-sentrux-sync` via `sentrux-rules-sync.ts` extension.
 5. **When to sync:**
    - `/harness-setup` Step 2.8 (after sentrux install)
@@ -34,5 +34,5 @@ Sentrux enforces architecture via [`.sentrux/rules.toml`](https://sentrux.dev/do
 ## References
 
 - ADR 0006 (Sentrux dual layer)
-- `scripts/sentrux-rules-sync.mjs`
+- `.pi/scripts/sentrux-rules-sync.mjs`
 - `.pi/extensions/sentrux-rules-sync.ts`

package/.pi/harness/incidents/README.md

@@ -0,0 +1,6 @@
+# Harness Incidents
+
+Store `IncidentRecord` artifacts and any policy override justifications here.
+
+- Override policy: one human approver only.
+- Justification is mandatory for every override record.

package/.pi/harness/release-readiness-report.md

@@ -0,0 +1,128 @@
+# Release Readiness Report
+
+Date: 2026-05-14
+Repo root used: `/home/aryaniyaps/ai-projects/ultimate-pi` (active workspace root, treated as canonical)
+
+## Requested remaining work
+
+- `run-adversarial-canary-and-release`
+- `final-prompt-expert-feature-sweep`
+
+Plan file was not modified.
+
+## Final integration checks
+
+### 1) TypeScript compile check
+
+- Command: `npm run check:ts`
+- Result: PASS
+
+### 2) Full lint/format/test gate
+
+- Command: `npm run check:ts && npm run lint && npm run format:check && npm test`
+- Result: FAIL (expected in current tree state)
+- Notes:
+  - `biome check` reports existing lint/format issues (including `.pi/extensions/custom-footer.ts` and multiple `.pi/harness/specs/*.json` files).
+  - `npm test` fails before test execution due Node runtime flag incompatibility:
+    - `node: bad option: --experimental-strip-types`
+
+### 3) Release preflight checks
+
+- Command: `git rev-parse --is-inside-work-tree && git remote -v && git symbolic-ref -q HEAD && (git diff --quiet && git diff --cached --quiet && echo CLEAN || echo DIRTY)`
+- Result:
+  - inside git repo: yes
+  - branch: `refs/heads/main`
+  - remote `origin`: configured
+  - tree cleanliness: `DIRTY` (release/tag push should stay blocked until clean)
+
+## Targeted canary validations
+
+### 1) Prompt and policy canary assertions
+
+- Static canary suite executed against:
+  - harness prompt templates
+  - `policy-gate`
+  - `test-diff-integrity`
+  - `debate-orchestrator`
+- Result: PASS after prompt sweep updates
+  - locked clauses in `harness-auto` preserved
+  - prompt argument parsing + usage surfaces present across harness prompts
+  - completion behavior sections present for operator-facing harness prompts
+  - policy/test/debate lock signals present in extension code
+
+### 2) Router tuning canary (proposal-only)
+
+- Created synthetic canary evidence:
+  - `.pi/harness/runs/canary-evidence.json`
+- Candidate router for dry proposal:
+  - `.pi/harness/runs/canary-candidate-router.json`
+- Command:
+  - `node .pi/harness/router/propose-router-tuning.mjs --evidence ... --candidate ... --proposal-out .pi/harness/router/proposals/canary-proposal.json`
+- Result: PASS (proposal created, no live router write)
+
+### 3) Harness schema parse check
+
+- Command: Node JSON parse validation across `.pi/harness/specs/*.json`
+- Result: PASS (all 9 schema files parse successfully)
+
+## Lightweight adversarial drills
+
+### 1) Negative apply drill (guardrail validation)
+
+- Command:
+  - `node .pi/harness/router/apply-router-proposal.mjs --proposal ... --approve-by ... --justification ...`
+  - intentionally omitted `--write`
+- Result: PASS (guard correctly blocked apply)
+- Expected error:
+  - `missing --write (blind writes and implicit applies are disallowed)`
+
+### 2) Adversarial lock retention
+
+- Verified locked governance semantics remain stated in `harness-auto`:
+  - adversarial review always required
+  - severity-policy-engine remains merge-block authority
+  - strict pre-PR gates mandatory
+  - never auto-merge
+
+## Prompt expert feature sweep
+
+Using guidance from `.pi/agents/pi-pi/prompt-expert.md`, harness prompt templates were refined for:
+
+1. Argument handling:
+   - explicit `$ARGUMENTS` parse sections
+   - required/optional argument normalization
+   - deterministic usage fallback lines
+2. Completion behavior:
+   - explicit terminal output contracts for predictable downstream handoff
+3. UX consistency:
+   - harmonized command usage patterns and closure blocks across harness prompts
+4. Policy integrity:
+   - locked policy constraints intentionally kept intact
+
+## Files updated in this sweep
+
+- `.pi/prompts/harness-auto.md`
+- `.pi/prompts/harness-plan.md`
+- `.pi/prompts/harness-run.md`
+- `.pi/prompts/harness-review.md`
+- `.pi/prompts/harness-critic.md`
+- `.pi/prompts/harness-eval.md`
+- `.pi/prompts/harness-trace.md`
+- `.pi/prompts/harness-incident.md`
+- `.pi/prompts/harness-router-tune.md`
+- `.pi/prompts/harness-setup.md`
+- `.pi/harness/release-readiness-report.md` (this report)
+
+## New canary artifacts
+
+- `.pi/harness/runs/canary-evidence.json`
+- `.pi/harness/runs/canary-candidate-router.json`
+- `.pi/harness/router/proposals/canary-proposal.json`
+
+## Residual risks
+
+1. Full repo lint/format gate currently fails due pre-existing issues unrelated to this sweep.
+2. `npm test` is currently not runnable in this environment because the configured Node flag is unsupported.
+3. Release flow should remain blocked until working tree is clean and CI-equivalent checks pass.
+4. Router apply path was intentionally not executed with `--write` during this run (safety-preserving drill).
+

package/.pi/harness/router/proposals/canary-proposal.json

@@ -0,0 +1,96 @@
+{
+	"schema_version": "1.0.0",
+	"proposal_id": "router-tune-2026-05-14T15-44-44-399Z",
+	"created_at": "2026-05-14T15:44:44.399Z",
+	"router_path": ".pi/model-router.json",
+	"base_router_sha256": "2a96fba517cc5b5147f37428d7ed62961b1968c0e83c0e69f02524265449856b",
+	"candidate_router_sha256": "2a96fba517cc5b5147f37428d7ed62961b1968c0e83c0e69f02524265449856b",
+	"evidence": {
+		"sample_count": 24,
+		"min_sample_count": 12,
+		"success_rate_delta": 0.08,
+		"cost_per_task_delta": -0.04,
+		"regression_guard_passed": true,
+		"trace_refs": ["run-canary-001", "run-canary-002"],
+		"notes": "canary validation synthetic evidence"
+	},
+	"status": "proposed",
+	"approval": {
+		"required": true,
+		"approved_by": null,
+		"approved_at": null,
+		"justification": null
+	},
+	"candidate_router": {
+		"defaultProfile": "auto",
+		"debug": false,
+		"classifierModel": "opencode-go/qwen3.6-plus",
+		"phaseBias": 0.5,
+		"maxSessionBudget": 1,
+		"largeContextThreshold": 100000,
+		"rules": [
+			{
+				"matches": ["deploy", "production", "release"],
+				"tier": "high",
+				"reason": "Safety check for production tasks"
+			},
+			{
+				"matches": "changelog",
+				"tier": "low"
+			}
+		],
+		"profiles": {
+			"auto": {
+				"high": {
+					"model": "opencode-go/deepseek-v4-pro",
+					"thinking": "high",
+					"fallbacks": ["opencode-go/qwen3.6-plus", "opencode-go/kimi-k2.6"]
+				},
+				"medium": {
+					"model": "opencode-go/qwen3.6-plus",
+					"thinking": "medium",
+					"fallbacks": ["opencode-go/deepseek-v4-pro"]
+				},
+				"low": {
+					"model": "opencode-go/deepseek-v4-flash",
+					"thinking": "low",
+					"fallbacks": ["opencode-go/qwen3.5-plus"]
+				}
+			},
+			"cheap": {
+				"high": {
+					"model": "opencode-go/qwen3.6-plus",
+					"thinking": "low",
+					"fallbacks": ["opencode-go/qwen3.5-plus"]
+				},
+				"medium": {
+					"model": "opencode-go/qwen3.5-plus",
+					"thinking": "off",
+					"fallbacks": ["opencode-go/deepseek-v4-flash"]
+				},
+				"low": {
+					"model": "opencode-go/deepseek-v4-flash",
+					"thinking": "off",
+					"fallbacks": ["opencode-go/qwen3.5-plus"]
+				}
+			},
+			"deep": {
+				"high": {
+					"model": "opencode-go/deepseek-v4-pro",
+					"thinking": "xhigh",
+					"fallbacks": ["opencode-go/kimi-k2.6"]
+				},
+				"medium": {
+					"model": "opencode-go/kimi-k2.6",
+					"thinking": "medium",
+					"fallbacks": ["opencode-go/deepseek-v4-pro"]
+				},
+				"low": {
+					"model": "opencode-go/qwen3.6-plus",
+					"thinking": "low",
+					"fallbacks": ["opencode-go/deepseek-v4-flash"]
+				}
+			}
+		}
+	}
+}

package/.pi/harness/runs/019e272f-3eef-7107-9712-ce281de55707-1778773891854/events.jsonl

@@ -0,0 +1,2 @@
+{"timestamp":"2026-05-14T15:51:31.965Z","type":"run_start","run_id":"019e272f-3eef-7107-9712-ce281de55707-1778773891854","plan_id":"plan-unknown","phase":"plan"}
+{"timestamp":"2026-05-14T15:51:38.346Z","type":"run_end","run_id":"019e272f-3eef-7107-9712-ce281de55707-1778773891854","phase":"plan","tool_span_count":0,"artifact_ref_count":0}

package/.pi/harness/runs/019e272f-3eef-7107-9712-ce281de55707-1778773891854/trace.json

@@ -0,0 +1,17 @@
+{
+	"schema_version": "1.0.0",
+	"contract_version": "1.0.0",
+	"run_id": "019e272f-3eef-7107-9712-ce281de55707-1778773891854",
+	"plan_id": "plan-unknown",
+	"agent_id": "019e272f-3eef-7107-9712-ce281de55707",
+	"phase": "plan",
+	"model": "auto",
+	"thinking_level": "off",
+	"tool_spans": [],
+	"artifact_refs": [],
+	"cost": {
+		"input_tokens": 15381,
+		"output_tokens": 33,
+		"total_tokens": 15414
+	}
+}

package/.pi/harness/runs/019e272f-3eef-7107-9712-ce281de55707-1778773912057/events.jsonl

@@ -0,0 +1,2 @@
+{"timestamp":"2026-05-14T15:51:52.062Z","type":"run_start","run_id":"019e272f-3eef-7107-9712-ce281de55707-1778773912057","plan_id":"plan-unknown","phase":"plan"}
+{"timestamp":"2026-05-14T15:52:14.313Z","type":"run_end","run_id":"019e272f-3eef-7107-9712-ce281de55707-1778773912057","phase":"plan","tool_span_count":0,"artifact_ref_count":0}

package/.pi/harness/runs/019e272f-3eef-7107-9712-ce281de55707-1778773912057/trace.json

@@ -0,0 +1,17 @@
+{
+	"schema_version": "1.0.0",
+	"contract_version": "1.0.0",
+	"run_id": "019e272f-3eef-7107-9712-ce281de55707-1778773912057",
+	"plan_id": "plan-unknown",
+	"agent_id": "019e272f-3eef-7107-9712-ce281de55707",
+	"phase": "plan",
+	"model": "auto",
+	"thinking_level": "off",
+	"tool_spans": [],
+	"artifact_refs": [],
+	"cost": {
+		"input_tokens": 31337,
+		"output_tokens": 528,
+		"total_tokens": 31865
+	}
+}

package/.pi/harness/runs/019e2732-8651-74e5-9f5d-4d06c3105f25-1778774086096/events.jsonl

@@ -0,0 +1,6 @@
+{"timestamp":"2026-05-14T15:54:46.136Z","type":"run_start","run_id":"019e2732-8651-74e5-9f5d-4d06c3105f25-1778774086096","plan_id":"plan-unknown","phase":"plan"}
+{"timestamp":"2026-05-14T15:54:59.110Z","type":"tool_start","run_id":"019e2732-8651-74e5-9f5d-4d06c3105f25-1778774086096","tool_call_id":"call_00_7UHDcydTHJHVR2dT5xpb0903","tool_name":"bash"}
+{"timestamp":"2026-05-14T15:54:59.137Z","type":"tool_start","run_id":"019e2732-8651-74e5-9f5d-4d06c3105f25-1778774086096","tool_call_id":"call_01_aNsry1whTl5hRf5Ew91t3142","tool_name":"bash"}
+{"timestamp":"2026-05-14T15:54:59.139Z","type":"tool_start","run_id":"019e2732-8651-74e5-9f5d-4d06c3105f25-1778774086096","tool_call_id":"call_02_N2e56Q6vKr6cAYzd4Z9q7953","tool_name":"bash"}
+{"timestamp":"2026-05-14T15:55:11.546Z","type":"tool_start","run_id":"019e2732-8651-74e5-9f5d-4d06c3105f25-1778774086096","tool_call_id":"call_00_wG71Rv3SKrf6R9K03EeS0264","tool_name":"ctx_batch_execute"}
+{"timestamp":"2026-05-14T15:55:25.167Z","type":"run_end","run_id":"019e2732-8651-74e5-9f5d-4d06c3105f25-1778774086096","phase":"plan","tool_span_count":4,"artifact_ref_count":0}

package/.pi/harness/runs/019e2732-8651-74e5-9f5d-4d06c3105f25-1778774086096/trace.json

@@ -0,0 +1,42 @@
+{
+	"schema_version": "1.0.0",
+	"contract_version": "1.0.0",
+	"run_id": "019e2732-8651-74e5-9f5d-4d06c3105f25-1778774086096",
+	"plan_id": "plan-unknown",
+	"agent_id": "019e2732-8651-74e5-9f5d-4d06c3105f25",
+	"phase": "plan",
+	"model": "auto",
+	"thinking_level": "off",
+	"tool_spans": [
+		{
+			"tool_call_id": "call_00_7UHDcydTHJHVR2dT5xpb0903",
+			"tool_name": "bash",
+			"started_at": "2026-05-14T15:54:59.108Z",
+			"ended_at": "2026-05-14T15:54:59.108Z"
+		},
+		{
+			"tool_call_id": "call_01_aNsry1whTl5hRf5Ew91t3142",
+			"tool_name": "bash",
+			"started_at": "2026-05-14T15:54:59.136Z",
+			"ended_at": "2026-05-14T15:54:59.136Z"
+		},
+		{
+			"tool_call_id": "call_02_N2e56Q6vKr6cAYzd4Z9q7953",
+			"tool_name": "bash",
+			"started_at": "2026-05-14T15:54:59.139Z",
+			"ended_at": "2026-05-14T15:54:59.139Z"
+		},
+		{
+			"tool_call_id": "call_00_wG71Rv3SKrf6R9K03EeS0264",
+			"tool_name": "ctx_batch_execute",
+			"started_at": "2026-05-14T15:55:11.541Z",
+			"ended_at": "2026-05-14T15:55:11.541Z"
+		}
+	],
+	"artifact_refs": [],
+	"cost": {
+		"input_tokens": 16951,
+		"output_tokens": 1020,
+		"total_tokens": 17971
+	}
+}

package/.pi/harness/runs/019e2732-8651-74e5-9f5d-4d06c3105f25-1778774136101/events.jsonl

@@ -0,0 +1 @@
+{"timestamp":"2026-05-14T15:55:36.107Z","type":"run_start","run_id":"019e2732-8651-74e5-9f5d-4d06c3105f25-1778774136101","plan_id":"plan-unknown","phase":"plan"}

package/.pi/harness/runs/019e2758-b332-771b-ad6f-54d0d8478768-1778776600591/events.jsonl

@@ -0,0 +1,2 @@
+{"timestamp":"2026-05-14T16:36:40.660Z","type":"run_start","run_id":"019e2758-b332-771b-ad6f-54d0d8478768-1778776600591","plan_id":"plan-unknown","phase":"plan"}
+{"timestamp":"2026-05-14T16:36:47.570Z","type":"run_end","run_id":"019e2758-b332-771b-ad6f-54d0d8478768-1778776600591","phase":"plan","tool_span_count":0,"artifact_ref_count":0}

package/.pi/harness/runs/019e2758-b332-771b-ad6f-54d0d8478768-1778776600591/trace.json

@@ -0,0 +1,17 @@
+{
+	"schema_version": "1.0.0",
+	"contract_version": "1.0.0",
+	"run_id": "019e2758-b332-771b-ad6f-54d0d8478768-1778776600591",
+	"plan_id": "plan-unknown",
+	"agent_id": "019e2758-b332-771b-ad6f-54d0d8478768",
+	"phase": "plan",
+	"model": "auto",
+	"thinking_level": "off",
+	"tool_spans": [],
+	"artifact_refs": [],
+	"cost": {
+		"input_tokens": 21,
+		"output_tokens": 32,
+		"total_tokens": 53
+	}
+}

package/.pi/harness/runs/README.md

@@ -0,0 +1,6 @@
+# Harness Runs
+
+Store lightweight run metadata and trace indexes here.
+
+- Primary source of truth for full trace payloads remains external telemetry.
+- Local files should contain run IDs, pointers, and replay metadata only.

package/.pi/harness/runs/budget-events.jsonl

@@ -0,0 +1,4 @@
+{"timestamp":"2026-05-14T15:54:59.134Z","schema_version":"1.0.0","contract_version":"1.0.0","event_type":"budget_exhausted","run_id":"019e2732-8651-74e5-9f5d-4d06c3105f25","debate_id":"plan-budget-guard","round_count":1,"budget_used":16593,"exhaustion_reason":"debate_global_cap_exceeded","caps":{"max_rounds":6,"round_token_cap":2500,"debate_global_cap":35000},"minimum_evidence_confidence":0.6,"default_policy_outcome":"block","human_override_allowed":true}
+{"timestamp":"2026-05-14T15:54:59.138Z","schema_version":"1.0.0","contract_version":"1.0.0","event_type":"budget_exhausted","run_id":"019e2732-8651-74e5-9f5d-4d06c3105f25","debate_id":"plan-budget-guard","round_count":1,"budget_used":16593,"exhaustion_reason":"debate_global_cap_exceeded","caps":{"max_rounds":6,"round_token_cap":2500,"debate_global_cap":35000},"minimum_evidence_confidence":0.6,"default_policy_outcome":"block","human_override_allowed":true}
+{"timestamp":"2026-05-14T15:54:59.140Z","schema_version":"1.0.0","contract_version":"1.0.0","event_type":"budget_exhausted","run_id":"019e2732-8651-74e5-9f5d-4d06c3105f25","debate_id":"plan-budget-guard","round_count":1,"budget_used":16593,"exhaustion_reason":"debate_global_cap_exceeded","caps":{"max_rounds":6,"round_token_cap":2500,"debate_global_cap":35000},"minimum_evidence_confidence":0.6,"default_policy_outcome":"block","human_override_allowed":true}
+{"timestamp":"2026-05-14T15:55:11.581Z","schema_version":"1.0.0","contract_version":"1.0.0","event_type":"budget_exhausted","run_id":"019e2732-8651-74e5-9f5d-4d06c3105f25","debate_id":"plan-budget-guard","round_count":1,"budget_used":17161,"exhaustion_reason":"debate_global_cap_exceeded","caps":{"max_rounds":6,"round_token_cap":2500,"debate_global_cap":35000},"minimum_evidence_confidence":0.6,"default_policy_outcome":"block","human_override_allowed":true}

package/.pi/harness/runs/canary-candidate-router.json

@@ -0,0 +1,72 @@
+{
+	"defaultProfile": "auto",
+	"debug": false,
+	"classifierModel": "opencode-go/qwen3.6-plus",
+	"phaseBias": 0.5,
+	"maxSessionBudget": 1.0,
+	"largeContextThreshold": 100000,
+	"rules": [
+		{
+			"matches": ["deploy", "production", "release"],
+			"tier": "high",
+			"reason": "Safety check for production tasks"
+		},
+		{
+			"matches": "changelog",
+			"tier": "low"
+		}
+	],
+	"profiles": {
+		"auto": {
+			"high": {
+				"model": "opencode-go/deepseek-v4-pro",
+				"thinking": "high",
+				"fallbacks": ["opencode-go/qwen3.6-plus", "opencode-go/kimi-k2.6"]
+			},
+			"medium": {
+				"model": "opencode-go/qwen3.6-plus",
+				"thinking": "medium",
+				"fallbacks": ["opencode-go/deepseek-v4-pro"]
+			},
+			"low": {
+				"model": "opencode-go/deepseek-v4-flash",
+				"thinking": "low",
+				"fallbacks": ["opencode-go/qwen3.5-plus"]
+			}
+		},
+		"cheap": {
+			"high": {
+				"model": "opencode-go/qwen3.6-plus",
+				"thinking": "low",
+				"fallbacks": ["opencode-go/qwen3.5-plus"]
+			},
+			"medium": {
+				"model": "opencode-go/qwen3.5-plus",
+				"thinking": "off",
+				"fallbacks": ["opencode-go/deepseek-v4-flash"]
+			},
+			"low": {
+				"model": "opencode-go/deepseek-v4-flash",
+				"thinking": "off",
+				"fallbacks": ["opencode-go/qwen3.5-plus"]
+			}
+		},
+		"deep": {
+			"high": {
+				"model": "opencode-go/deepseek-v4-pro",
+				"thinking": "xhigh",
+				"fallbacks": ["opencode-go/kimi-k2.6"]
+			},
+			"medium": {
+				"model": "opencode-go/kimi-k2.6",
+				"thinking": "medium",
+				"fallbacks": ["opencode-go/deepseek-v4-pro"]
+			},
+			"low": {
+				"model": "opencode-go/qwen3.6-plus",
+				"thinking": "low",
+				"fallbacks": ["opencode-go/deepseek-v4-flash"]
+			}
+		}
+	}
+}

package/.pi/harness/runs/canary-evidence.json

@@ -0,0 +1,9 @@
+{
+	"sample_count": 24,
+	"min_sample_count": 12,
+	"success_rate_delta": 0.08,
+	"cost_per_task_delta": -0.04,
+	"regression_guard_passed": true,
+	"trace_refs": ["run-canary-001", "run-canary-002"],
+	"notes": "canary validation synthetic evidence"
+}

package/.pi/harness/runs/index.jsonl

@@ -0,0 +1,4 @@
+{"timestamp":"2026-05-14T15:51:38.345Z","run_id":"019e272f-3eef-7107-9712-ce281de55707-1778773891854","plan_id":"plan-unknown","phase":"plan","trace_file":"/home/aryaniyaps/ai-projects/ultimate-pi/.pi/harness/runs/019e272f-3eef-7107-9712-ce281de55707-1778773891854/trace.json"}
+{"timestamp":"2026-05-14T15:52:14.312Z","run_id":"019e272f-3eef-7107-9712-ce281de55707-1778773912057","plan_id":"plan-unknown","phase":"plan","trace_file":"/home/aryaniyaps/ai-projects/ultimate-pi/.pi/harness/runs/019e272f-3eef-7107-9712-ce281de55707-1778773912057/trace.json"}
+{"timestamp":"2026-05-14T15:55:25.166Z","run_id":"019e2732-8651-74e5-9f5d-4d06c3105f25-1778774086096","plan_id":"plan-unknown","phase":"plan","trace_file":"/home/aryaniyaps/ai-projects/ultimate-pi/.pi/harness/runs/019e2732-8651-74e5-9f5d-4d06c3105f25-1778774086096/trace.json"}
+{"timestamp":"2026-05-14T16:36:47.569Z","run_id":"019e2758-b332-771b-ad6f-54d0d8478768-1778776600591","plan_id":"plan-unknown","phase":"plan","trace_file":"/home/aryaniyaps/ai-projects/ultimate-pi/.pi/harness/runs/019e2758-b332-771b-ad6f-54d0d8478768-1778776600591/trace.json"}

package/.pi/harness/sentrux/architecture.manifest.json

@@ -34,9 +34,9 @@
 		},
 		{
 			"name": "tooling",
-			"paths": ["scripts/*", "test/*"],
+			"paths": [".pi/scripts/*", "test/*"],
 			"order": 4,
-			"description": "Deterministic scripts and tests"
+			"description": "Harness CLI scripts and tests"
 		}
 	],
 	"boundaries": [
@@ -61,7 +61,7 @@
 			"reason": "Contracts are data-only JSON schemas; extensions implement behavior"
 		},
 		{
-			"from": "scripts/*",
+			"from": ".pi/scripts/*",
 			"to": ".agents/skills/*",
 			"reason": "CLI scripts stay independent of skill markdown"
 		}

package/.pi/model-router.json

@@ -0,0 +1,95 @@
+{
+    "defaultProfile": "auto",
+    "debug": false,
+    "classifierModel": "opencode-go/qwen3.6-plus",
+    "phaseBias": 0.5,
+    "maxSessionBudget": 1.0,
+    "largeContextThreshold": 100000,
+    "rules": [
+        {
+            "matches": [
+                "deploy",
+                "production",
+                "release"
+            ],
+            "tier": "high",
+            "reason": "Safety check for production tasks"
+        },
+        {
+            "matches": "changelog",
+            "tier": "low"
+        }
+    ],
+    "profiles": {
+        "auto": {
+            "high": {
+                "model": "opencode-go/deepseek-v4-pro",
+                "thinking": "high",
+                "fallbacks": [
+                    "opencode-go/qwen3.6-plus",
+                    "opencode-go/kimi-k2.6"
+                ]
+            },
+            "medium": {
+                "model": "opencode-go/qwen3.6-plus",
+                "thinking": "medium",
+                "fallbacks": [
+                    "opencode-go/deepseek-v4-pro"
+                ]
+            },
+            "low": {
+                "model": "opencode-go/deepseek-v4-flash",
+                "thinking": "low",
+                "fallbacks": [
+                    "opencode-go/qwen3.5-plus"
+                ]
+            }
+        },
+        "cheap": {
+            "high": {
+                "model": "opencode-go/qwen3.6-plus",
+                "thinking": "low",
+                "fallbacks": [
+                    "opencode-go/qwen3.5-plus"
+                ]
+            },
+            "medium": {
+                "model": "opencode-go/qwen3.5-plus",
+                "thinking": "off",
+                "fallbacks": [
+                    "opencode-go/deepseek-v4-flash"
+                ]
+            },
+            "low": {
+                "model": "opencode-go/deepseek-v4-flash",
+                "thinking": "off",
+                "fallbacks": [
+                    "opencode-go/qwen3.5-plus"
+                ]
+            }
+        },
+        "deep": {
+            "high": {
+                "model": "opencode-go/deepseek-v4-pro",
+                "thinking": "xhigh",
+                "fallbacks": [
+                    "opencode-go/kimi-k2.6"
+                ]
+            },
+            "medium": {
+                "model": "opencode-go/kimi-k2.6",
+                "thinking": "medium",
+                "fallbacks": [
+                    "opencode-go/deepseek-v4-pro"
+                ]
+            },
+            "low": {
+                "model": "opencode-go/qwen3.6-plus",
+                "thinking": "low",
+                "fallbacks": [
+                    "opencode-go/deepseek-v4-flash"
+                ]
+            }
+        }
+    }
+}