@oriro/orirocli 0.1.11 → 0.3.1

package/skills/craft/become-an-ai-engineer-26/phases/phase-5-production/README.md

@@ -1,16 +1,16 @@
-# Phase 5 — Production Hardening (Forever)
-
-This phase never ends. See `SKILL.md` → _Phase 5_ for the code patterns.
-
-The five fronts:
-
-1. **Cost discipline** — prompt caching (up to 90% savings), route by difficulty
-   (Haiku → Sonnet → Opus), Batch API for non-real-time work (50% off).
-2. **Latency** — parallel tool calls always; sub-agent fan-out turns a 60-step
-   sequential run into a 10-step lead + parallel sub-agents.
-3. **Safety & sandboxing** — never `exec()` model output; run it in an isolated
-   container. Never put credentials in the model's context — broker them.
-4. **Monitoring & drift** — alert on cost/request, tool-failure rate, judge
-   score, p95 latency. Re-baseline after every model upgrade.
-5. **Resilience** — durable execution (PostgresSaver) for anything running >60s,
-   so a process kill is a resume, not a restart.
+# Phase 5 — Production Hardening (Forever)
+
+This phase never ends. See `SKILL.md` → _Phase 5_ for the code patterns.
+
+The five fronts:
+
+1. **Cost discipline** — prompt caching (up to 90% savings), route by difficulty
+   (Haiku → Sonnet → Opus), Batch API for non-real-time work (50% off).
+2. **Latency** — parallel tool calls always; sub-agent fan-out turns a 60-step
+   sequential run into a 10-step lead + parallel sub-agents.
+3. **Safety & sandboxing** — never `exec()` model output; run it in an isolated
+   container. Never put credentials in the model's context — broker them.
+4. **Monitoring & drift** — alert on cost/request, tool-failure rate, judge
+   score, p95 latency. Re-baseline after every model upgrade.
+5. **Resilience** — durable execution (PostgresSaver) for anything running >60s,
+   so a process kill is a resume, not a restart.

package/skills/craft/become-an-ai-engineer-26/projects/1-mobile-app-slm/README.md

@@ -1,11 +1,11 @@
-# Project 1 — Mobile App + SLM (Beginner)
-
-**Proves:** Edge AI, resource constraints.
-
-Offline-first mobile app using a small language model. Zero API costs, complete
-privacy. See `SKILL.md` → _The 5 Portfolio Projects_.
-
-- Lazy-load models on demand; unload under memory pressure
-- Sliding context window with semantic chunking
-- 4-bit quantisation for older devices, 8-bit for newer
-- Batch inference to reduce battery wake cycles
+# Project 1 — Mobile App + SLM (Beginner)
+
+**Proves:** Edge AI, resource constraints.
+
+Offline-first mobile app using a small language model. Zero API costs, complete
+privacy. See `SKILL.md` → _The 5 Portfolio Projects_.
+
+- Lazy-load models on demand; unload under memory pressure
+- Sliding context window with semantic chunking
+- 4-bit quantisation for older devices, 8-bit for newer
+- Batch inference to reduce battery wake cycles

package/skills/craft/become-an-ai-engineer-26/projects/2-self-improving-coder/README.md

@@ -1,11 +1,11 @@
-# Project 2 — Self-Improving Coding Agent (Intermediate)
-
-**Proves:** Agentic loops, debugging.
-
-Agent writes code → runs tests → learns from failures → loops until functional.
-See `SKILL.md` → _The 5 Portfolio Projects_.
-
-- Plan → Execute → Test → Reflect loop with a max-iteration limit
-- Isolated execution per task with resource limits
-- Memory: short-term (last 5) + long-term (successes) + failure memory (error signatures)
-- Static analysis before execution — detect dangerous operations
+# Project 2 — Self-Improving Coding Agent (Intermediate)
+
+**Proves:** Agentic loops, debugging.
+
+Agent writes code → runs tests → learns from failures → loops until functional.
+See `SKILL.md` → _The 5 Portfolio Projects_.
+
+- Plan → Execute → Test → Reflect loop with a max-iteration limit
+- Isolated execution per task with resource limits
+- Memory: short-term (last 5) + long-term (successes) + failure memory (error signatures)
+- Static analysis before execution — detect dangerous operations

package/skills/craft/become-an-ai-engineer-26/projects/3-video-editor-agent/README.md

@@ -1,11 +1,11 @@
-# Project 3 — Video Editor Agent (Advanced)
-
-**Proves:** Multimodal AI, tool integration.
-
-Fork an open-source editor. User says "make this cinematic"; the agent handles
-cuts, transitions, colour. See `SKILL.md` → _The 5 Portfolio Projects_.
-
-- Vision model analyses frames + audio model analyses dialogue
-- Intent translation: "cinematic" → concrete parameters
-- Scene detection via frame-difference analysis
-- Incremental preview — only re-render affected sections
+# Project 3 — Video Editor Agent (Advanced)
+
+**Proves:** Multimodal AI, tool integration.
+
+Fork an open-source editor. User says "make this cinematic"; the agent handles
+cuts, transitions, colour. See `SKILL.md` → _The 5 Portfolio Projects_.
+
+- Vision model analyses frames + audio model analyses dialogue
+- Intent translation: "cinematic" → concrete parameters
+- Scene detection via frame-difference analysis
+- Incremental preview — only re-render affected sections

package/skills/craft/become-an-ai-engineer-26/projects/4-personal-life-os/README.md

@@ -1,12 +1,12 @@
-# Project 4 — Personal Life OS Agent (Expert)
-
-**Proves:** Deep context management, privacy-first design.
-
-Manages calendar, finances, and health. Plans months ahead. Detects burnout.
-See `SKILL.md` → _The 5 Portfolio Projects_.
-
-- Real-time ingestion: calendar, finance, health, communications
-- Personal knowledge graph of entities and relationships
-- Background thread every 6 hours checking for anomalies
-- Value alignment: user states priorities — recommendations validated against them
-- All data encrypted at rest with user-controlled keys
+# Project 4 — Personal Life OS Agent (Expert)
+
+**Proves:** Deep context management, privacy-first design.
+
+Manages calendar, finances, and health. Plans months ahead. Detects burnout.
+See `SKILL.md` → _The 5 Portfolio Projects_.
+
+- Real-time ingestion: calendar, finance, health, communications
+- Personal knowledge graph of entities and relationships
+- Background thread every 6 hours checking for anomalies
+- Value alignment: user states priorities — recommendations validated against them
+- All data encrypted at rest with user-controlled keys

package/skills/craft/become-an-ai-engineer-26/projects/5-enterprise-workflow/README.md

@@ -1,12 +1,12 @@
-# Project 5 — Enterprise Workflow Agent (Master)
-
-**Proves:** Production orchestration.
-
-Monitors Slack/Jira → plans execution → delegates → reports with audit logs.
-See `SKILL.md` → _The 5 Portfolio Projects_.
-
-- Event-driven: Slack, Jira, email, monitoring systems
-- Multi-agent delegation: orchestrator → communication, data, analysis, documentation agents
-- Self-healing: exponential backoff, circuit breakers, retry decisions
-- Immutable audit log: every action, authorisation, outcome
-- Human-in-the-loop: agent proposes a plan before executing critical workflows
+# Project 5 — Enterprise Workflow Agent (Master)
+
+**Proves:** Production orchestration.
+
+Monitors Slack/Jira → plans execution → delegates → reports with audit logs.
+See `SKILL.md` → _The 5 Portfolio Projects_.
+
+- Event-driven: Slack, Jira, email, monitoring systems
+- Multi-agent delegation: orchestrator → communication, data, analysis, documentation agents
+- Self-healing: exponential backoff, circuit breakers, retry decisions
+- Immutable audit log: every action, authorisation, outcome
+- Human-in-the-loop: agent proposes a plan before executing critical workflows

package/skills/craft/become-an-ai-engineer-26/references/benchmark-numbers.md

@@ -1,41 +1,41 @@
-# Benchmark Numbers (May 2026)
-
-The leaderboard changes weekly. What does **not** change: the same model on a
-different harness swings 10–36 points on any benchmark. The harness writes the
-score.
-
-## The one number
-
-```
-Same model. Two harnesses. CORE benchmark:
-
-  Claude Code harness:  78%
-  Smolagents harness:   42%
-                        ───
-                   36-point gap
-```
-
-## Current scores
-
-| Benchmark          | What it measures        | Leader (May 2026)                       |
-| ------------------ | ----------------------- | --------------------------------------- |
-| SWE-bench Verified | Real-world coding fixes | GPT-5.5 ~88.7% · Claude Opus 4.7 ~87.6% |
-| GAIA               | General agent tasks     | Claude Sonnet 4.5 — 74.6%               |
-| τ-bench            | Customer-service agents | Claude Mythos Preview — 89.2%           |
-
-## How to read a benchmark
-
-- **A single number is meaningless without the harness.** Always ask: which
-  scaffold produced this score? Tool set, context strategy, retry policy and
-  sub-agent structure move the number more than the base model does.
-- **Re-baseline after every model upgrade.** Harnesses encode assumptions about
-  model limitations; those assumptions go stale when the model changes.
-- **Your own golden dataset beats any public leaderboard** for deciding whether
-  _your_ system got better. Public benchmarks tell you the field is moving;
-  your evals tell you _you_ are.
-
-> Update this file as new scores publish — cite the primary source in the PR.
-
----
-
-_Reference for "Become An AI Engineer 26"._
+# Benchmark Numbers (May 2026)
+
+The leaderboard changes weekly. What does **not** change: the same model on a
+different harness swings 10–36 points on any benchmark. The harness writes the
+score.
+
+## The one number
+
+```
+Same model. Two harnesses. CORE benchmark:
+
+  Claude Code harness:  78%
+  Smolagents harness:   42%
+                        ───
+                   36-point gap
+```
+
+## Current scores
+
+| Benchmark          | What it measures        | Leader (May 2026)                       |
+| ------------------ | ----------------------- | --------------------------------------- |
+| SWE-bench Verified | Real-world coding fixes | GPT-5.5 ~88.7% · Claude Opus 4.7 ~87.6% |
+| GAIA               | General agent tasks     | Claude Sonnet 4.5 — 74.6%               |
+| τ-bench            | Customer-service agents | Claude Mythos Preview — 89.2%           |
+
+## How to read a benchmark
+
+- **A single number is meaningless without the harness.** Always ask: which
+  scaffold produced this score? Tool set, context strategy, retry policy and
+  sub-agent structure move the number more than the base model does.
+- **Re-baseline after every model upgrade.** Harnesses encode assumptions about
+  model limitations; those assumptions go stale when the model changes.
+- **Your own golden dataset beats any public leaderboard** for deciding whether
+  _your_ system got better. Public benchmarks tell you the field is moving;
+  your evals tell you _you_ are.
+
+> Update this file as new scores publish — cite the primary source in the PR.
+
+---
+
+_Reference for "Become An AI Engineer 26"._

package/skills/craft/become-an-ai-engineer-26/references/mhc-stable-training.md

@@ -1,73 +1,73 @@
-# mHC — Stable Training at Scale
-
-> How to keep training stable at 27B+ parameters using manifold-constrained
-> residual streams. DeepSeek · arxiv:2512.24880 (Dec 2025)
-
----
-
-## The problem
-
-Standard **residual connections** preserve signal by adding the input back to
-the output: `x_out = x_in + f(x_in)`. The identity term (`x_in`) is what keeps
-gradients flowing through very deep stacks.
-
-**Hyper-Connections (HC)** generalised this by adding several parallel residual
-streams and a learned matrix that mixes them. More capacity — but the learned
-mixing matrix can drift away from preserving identity, and when it does you get
-**loss spikes** (in the reported runs, around step 12,000). One spike can undo
-hours of training or diverge the run entirely.
-
-## The fix: constrain the mixing matrix to a manifold
-
-**mHC (manifold-constrained Hyper-Connections)** projects the stream-mixing
-matrix onto the **Sinkhorn manifold** — the set of doubly-stochastic matrices
-where every row and every column sums to 1. On that manifold, identity mapping
-is preserved automatically, so the residual stream can't blow up or collapse.
-
-```python
-import torch
-
-def sinkhorn(M: torch.Tensor, iters: int = 8, eps: float = 1e-6) -> torch.Tensor:
-    """Project a positive matrix toward doubly-stochastic (rows & cols sum to 1)."""
-    M = M.clamp_min(eps)
-    for _ in range(iters):
-        M = M / M.sum(dim=1, keepdim=True)   # normalise rows
-        M = M / M.sum(dim=0, keepdim=True)   # normalise columns
-    return M
-
-class ManifoldHyperConnection(torch.nn.Module):
-    def __init__(self, n_streams: int):
-        super().__init__()
-        self.logits = torch.nn.Parameter(torch.zeros(n_streams, n_streams))
-
-    def mixing_matrix(self) -> torch.Tensor:
-        # exp keeps entries positive; Sinkhorn makes it doubly-stochastic
-        return sinkhorn(self.logits.exp())
-
-    def forward(self, streams: torch.Tensor) -> torch.Tensor:
-        # streams: (n_streams, batch, dim) -> mixed streams, identity preserved
-        return torch.einsum("ij,jbd->ibd", self.mixing_matrix(), streams)
-```
-
-## Results
-
-- **6.7% compute overhead** vs standard residuals.
-- **Zero loss spikes** across the reported runs.
-- Evaluated at **3B / 9B / 27B** parameter scales.
-
-## The harness parallel
-
-The lesson generalises beyond model internals. **Constrain the system at the
-right point and the whole thing becomes stable under real load.** In harness
-design the equivalent constraints are:
-
-- **Context compaction at ~85%** of the window — never let context overflow.
-- **Credential brokering** — the model never sees raw secrets.
-- **Sandboxed execution** — model output runs in an isolated container.
-
-Same principle: pick the one invariant that must always hold, enforce it
-structurally, and stop firefighting the symptoms.
-
----
-
-_Reference for "Become An AI Engineer 26". See SKILL.md → Advanced Science._
+# mHC — Stable Training at Scale
+
+> How to keep training stable at 27B+ parameters using manifold-constrained
+> residual streams. DeepSeek · arxiv:2512.24880 (Dec 2025)
+
+---
+
+## The problem
+
+Standard **residual connections** preserve signal by adding the input back to
+the output: `x_out = x_in + f(x_in)`. The identity term (`x_in`) is what keeps
+gradients flowing through very deep stacks.
+
+**Hyper-Connections (HC)** generalised this by adding several parallel residual
+streams and a learned matrix that mixes them. More capacity — but the learned
+mixing matrix can drift away from preserving identity, and when it does you get
+**loss spikes** (in the reported runs, around step 12,000). One spike can undo
+hours of training or diverge the run entirely.
+
+## The fix: constrain the mixing matrix to a manifold
+
+**mHC (manifold-constrained Hyper-Connections)** projects the stream-mixing
+matrix onto the **Sinkhorn manifold** — the set of doubly-stochastic matrices
+where every row and every column sums to 1. On that manifold, identity mapping
+is preserved automatically, so the residual stream can't blow up or collapse.
+
+```python
+import torch
+
+def sinkhorn(M: torch.Tensor, iters: int = 8, eps: float = 1e-6) -> torch.Tensor:
+    """Project a positive matrix toward doubly-stochastic (rows & cols sum to 1)."""
+    M = M.clamp_min(eps)
+    for _ in range(iters):
+        M = M / M.sum(dim=1, keepdim=True)   # normalise rows
+        M = M / M.sum(dim=0, keepdim=True)   # normalise columns
+    return M
+
+class ManifoldHyperConnection(torch.nn.Module):
+    def __init__(self, n_streams: int):
+        super().__init__()
+        self.logits = torch.nn.Parameter(torch.zeros(n_streams, n_streams))
+
+    def mixing_matrix(self) -> torch.Tensor:
+        # exp keeps entries positive; Sinkhorn makes it doubly-stochastic
+        return sinkhorn(self.logits.exp())
+
+    def forward(self, streams: torch.Tensor) -> torch.Tensor:
+        # streams: (n_streams, batch, dim) -> mixed streams, identity preserved
+        return torch.einsum("ij,jbd->ibd", self.mixing_matrix(), streams)
+```
+
+## Results
+
+- **6.7% compute overhead** vs standard residuals.
+- **Zero loss spikes** across the reported runs.
+- Evaluated at **3B / 9B / 27B** parameter scales.
+
+## The harness parallel
+
+The lesson generalises beyond model internals. **Constrain the system at the
+right point and the whole thing becomes stable under real load.** In harness
+design the equivalent constraints are:
+
+- **Context compaction at ~85%** of the window — never let context overflow.
+- **Credential brokering** — the model never sees raw secrets.
+- **Sandboxed execution** — model output runs in an isolated container.
+
+Same principle: pick the one invariant that must always hold, enforce it
+structurally, and stop firefighting the symptoms.
+
+---
+
+_Reference for "Become An AI Engineer 26". See SKILL.md → Advanced Science._

package/skills/craft/become-an-ai-engineer-26/references/stack-decisions.md

@@ -1,37 +1,37 @@
-# Stack Decisions
-
-The recommended 2026 production stack, with the reasoning for each call.
-
-## Use
-
-| Layer             | Choice                                                                                                                                 | Why                                                                                             |
-| ----------------- | -------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------- |
-| Framework         | **LangGraph 1.0 + Deep Agents**                                                                                                        | State machine + durable checkpointing + observability. Survives process kills, resumes mid-run. |
-| Harness reference | **Claude Agent SDK**                                                                                                                   | Same harness as Claude Code: CLAUDE.md + Skills + sub-agents + hooks + filesystem-as-memory.    |
-| Observability     | **LangSmith** (LangGraph-native) / **Braintrust** (framework-agnostic, CI gating) / **Arize Phoenix** (open-source, OTEL-native, free) | Pick one. You cannot improve what you cannot trace.                                             |
-
-## Avoid in production
-
-| Avoid                            | Reason                                                 |
-| -------------------------------- | ------------------------------------------------------ |
-| CrewAI in production             | Fast demo, fragile at scale. Fine for hackathons only. |
-| OpenAI Swarm                     | Their own README says "not production-ready".          |
-| OpenAI Assistants API            | Sunsetting mid-2026.                                   |
-| No-code agent platforms          | Throwaway prototypes only.                             |
-| Custom vector store (by default) | Measure recall first — most apps don't need one.       |
-
-## The decision rule
-
-1. What is the **simplest** thing that solves this problem?
-2. What does it look like at **10×** current scale (not 100×)?
-3. What is the **cost of being wrong** — easy to change, or locked in?
-
-Start with a monolith. Start serverless. Buy everything that isn't your core
-product (auth, email, payments). Split into services only when parts genuinely
-need to scale independently.
-
-> Update with benchmark-backed comparisons only — see CONTRIBUTING.md.
-
----
-
-_Reference for "Become An AI Engineer 26". See SKILL.md → The Stack._
+# Stack Decisions
+
+The recommended 2026 production stack, with the reasoning for each call.
+
+## Use
+
+| Layer             | Choice                                                                                                                                 | Why                                                                                             |
+| ----------------- | -------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------- |
+| Framework         | **LangGraph 1.0 + Deep Agents**                                                                                                        | State machine + durable checkpointing + observability. Survives process kills, resumes mid-run. |
+| Harness reference | **Claude Agent SDK**                                                                                                                   | Same harness as Claude Code: CLAUDE.md + Skills + sub-agents + hooks + filesystem-as-memory.    |
+| Observability     | **LangSmith** (LangGraph-native) / **Braintrust** (framework-agnostic, CI gating) / **Arize Phoenix** (open-source, OTEL-native, free) | Pick one. You cannot improve what you cannot trace.                                             |
+
+## Avoid in production
+
+| Avoid                            | Reason                                                 |
+| -------------------------------- | ------------------------------------------------------ |
+| CrewAI in production             | Fast demo, fragile at scale. Fine for hackathons only. |
+| OpenAI Swarm                     | Their own README says "not production-ready".          |
+| OpenAI Assistants API            | Sunsetting mid-2026.                                   |
+| No-code agent platforms          | Throwaway prototypes only.                             |
+| Custom vector store (by default) | Measure recall first — most apps don't need one.       |
+
+## The decision rule
+
+1. What is the **simplest** thing that solves this problem?
+2. What does it look like at **10×** current scale (not 100×)?
+3. What is the **cost of being wrong** — easy to change, or locked in?
+
+Start with a monolith. Start serverless. Buy everything that isn't your core
+product (auth, email, payments). Split into services only when parts genuinely
+need to scale independently.
+
+> Update with benchmark-backed comparisons only — see CONTRIBUTING.md.
+
+---
+
+_Reference for "Become An AI Engineer 26". See SKILL.md → The Stack._