From b0268cceb4265a4158ecac4af3aac163a361f3f2 Mon Sep 17 00:00:00 2001 From: Baptiste Grenier Date: Mon, 16 Nov 2026 16:10:31 +0100 Subject: [PATCH] Stop pointing the model at Grep and Glob tools native builds lack Native Claude Code builds on macOS and Linux replaced the Glob and Grep tools with embedded search reached through the Bash tool (2.1.117). The Tools preference in the principal identity template, two context strings in settings.enhancements.json and one architecture decision still told the model to use them, so on a native build the model is steered toward tools it does not have. The wording now names Read and Bash search (rg, fd), which holds on every build. Both copies of settings.enhancements.json are changed so they stay identical. --- .../install/LIFEOS/DOCUMENTATION/LifeosSystemArchitecture.md | 2 +- LifeOS/install/USER/PRINCIPAL/PRINCIPAL_IDENTITY.md | 2 +- LifeOS/install/settings.enhancements.json | 4 ++-- .../install/skills/LifeOS/install/settings.enhancements.json | 4 ++-- 4 files changed, 6 insertions(+), 6 deletions(-) diff --git a/LifeOS/install/LIFEOS/DOCUMENTATION/LifeosSystemArchitecture.md b/LifeOS/install/LIFEOS/DOCUMENTATION/LifeosSystemArchitecture.md index 5c6ff65671..ba75b225a2 100644 --- a/LifeOS/install/LIFEOS/DOCUMENTATION/LifeosSystemArchitecture.md +++ b/LifeOS/install/LIFEOS/DOCUMENTATION/LifeosSystemArchitecture.md @@ -700,7 +700,7 @@ The running log of the load-bearing choices that shape LifeOS — what we decide ### AD-3: Fast non-standard CLIs — fd, rg, bat, eza - **Decided:** longstanding; codified 2026-06-14 -- **Decision:** In shell work prefer `rg` over `grep`, `fd` over `find`, `bat` over `cat`, `eza` over `ls`. In-tool work uses the built-in Grep/Glob/Read (Grep already wraps ripgrep). Portable skill code prefers Bun fs APIs; when shelling out, prefer `find` for guaranteed availability. +- **Decision:** In shell work prefer `rg` over `grep`, `fd` over `find`, `bat` over `cat`, `eza` over `ls`. In-tool work uses the built-in Read; search runs through Bash, since native Claude Code builds no longer ship the Grep and Glob tools. Portable skill code prefers Bun fs APIs; when shelling out, prefer `find` for guaranteed availability. - **Why:** The modern tools are faster, respect `.gitignore` by default, and have better defaults and output. Describing LifeOS as "built on fast, modern tooling" is a real architectural stance, not cosmetic. - **Replaced:** Default `grep`/`find`/`cat`/`ls`. - **Source:** manual (backfill 2026-06-14) diff --git a/LifeOS/install/USER/PRINCIPAL/PRINCIPAL_IDENTITY.md b/LifeOS/install/USER/PRINCIPAL/PRINCIPAL_IDENTITY.md index f9c13f0075..aa05dbe737 100644 --- a/LifeOS/install/USER/PRINCIPAL/PRINCIPAL_IDENTITY.md +++ b/LifeOS/install/USER/PRINCIPAL/PRINCIPAL_IDENTITY.md @@ -48,7 +48,7 @@ Loaded at every session startup via CLAUDE.md `@` import. Gives the DA a concise ## Preferences - **Units:** Metric (Celsius, meters, kilograms) -- **Tools:** Built-in Claude tools (Grep, Glob, Read) for in-tool work — Grep already wraps ripgrep. In Bash: `rg` over `grep`, `fd` over `find`, `bat` over `cat`. In portable skill code: prefer language-native fs APIs; if shelling out, prefer `find` for guaranteed availability. +- **Tools:** `Read` for file contents; search through Bash, since native Claude Code builds no longer ship the Grep and Glob tools: `rg` over `grep`, `fd` over `find`, `bat` over `cat`. In portable skill code: prefer language-native fs APIs; if shelling out, prefer `find` for guaranteed availability. - **Decision style:** (interview — build vs ask, simple vs clever) --- diff --git a/LifeOS/install/settings.enhancements.json b/LifeOS/install/settings.enhancements.json index 6d2bc8fd63..21b1024c88 100644 --- a/LifeOS/install/settings.enhancements.json +++ b/LifeOS/install/settings.enhancements.json @@ -89,7 +89,7 @@ "There are no effort tiers or levels — one loop serves a one-line answer and a week-long build alike; only the depth adapts.", "Spend is matched to the task, not a default tier — both overspending and underspending miss the euphoric-surprise target.", "Scaffolding creates the ISA and writes ISC criteria; prerequisite probing is one of the run's own criteria.", - "Context recovery uses Grep/Glob/Read directly — never agents. Direct tools for directed lookups.", + "Context recovery uses Read and Bash search (rg, fd) directly — never agents. Direct tools for directed lookups.", "Plan mode (EnterPlanMode) surfaces a plan for approval before the Algorithm builds high-blast-radius work.", "ISAs persist ISC to disk. They survive across sessions. ISA on disk wins conflicts.", "The Algorithm runs in the main loop against an ISA — no separate CLI to invoke.", @@ -158,7 +158,7 @@ "Scaffolding reverse-engineers the ask: explicit wants, explicit not-wanted, implied not-wanted, speed/urgency.", "Anti-criteria (ISC-A prefix) capture what must NOT happen. Every ISA should have at least one.", "ISA phase brackets progress from marking through climbing to complete — one table in ascent.ts owns the vocabulary.", - "The 2-second rule: if Grep/Glob/Read can do it in 2s, never spawn an agent.", + "The 2-second rule: if Read or a Bash search (rg, fd) can do it in 2s, never spawn an agent.", "No ISA criterion closes without tool evidence of the right modality — file→Read, code→Grep, HTTP→curl -i, deploy→live probe; \"should work\" is forbidden.", "The Workflows action CLI (lifeos.ts) runs actions and pipelines from the terminal.", "Actions compose via UNIX pipes: output of one action becomes input of the next.", diff --git a/LifeOS/install/skills/LifeOS/install/settings.enhancements.json b/LifeOS/install/skills/LifeOS/install/settings.enhancements.json index 6d2bc8fd63..21b1024c88 100644 --- a/LifeOS/install/skills/LifeOS/install/settings.enhancements.json +++ b/LifeOS/install/skills/LifeOS/install/settings.enhancements.json @@ -89,7 +89,7 @@ "There are no effort tiers or levels — one loop serves a one-line answer and a week-long build alike; only the depth adapts.", "Spend is matched to the task, not a default tier — both overspending and underspending miss the euphoric-surprise target.", "Scaffolding creates the ISA and writes ISC criteria; prerequisite probing is one of the run's own criteria.", - "Context recovery uses Grep/Glob/Read directly — never agents. Direct tools for directed lookups.", + "Context recovery uses Read and Bash search (rg, fd) directly — never agents. Direct tools for directed lookups.", "Plan mode (EnterPlanMode) surfaces a plan for approval before the Algorithm builds high-blast-radius work.", "ISAs persist ISC to disk. They survive across sessions. ISA on disk wins conflicts.", "The Algorithm runs in the main loop against an ISA — no separate CLI to invoke.", @@ -158,7 +158,7 @@ "Scaffolding reverse-engineers the ask: explicit wants, explicit not-wanted, implied not-wanted, speed/urgency.", "Anti-criteria (ISC-A prefix) capture what must NOT happen. Every ISA should have at least one.", "ISA phase brackets progress from marking through climbing to complete — one table in ascent.ts owns the vocabulary.", - "The 2-second rule: if Grep/Glob/Read can do it in 2s, never spawn an agent.", + "The 2-second rule: if Read or a Bash search (rg, fd) can do it in 2s, never spawn an agent.", "No ISA criterion closes without tool evidence of the right modality — file→Read, code→Grep, HTTP→curl -i, deploy→live probe; \"should work\" is forbidden.", "The Workflows action CLI (lifeos.ts) runs actions and pipelines from the terminal.", "Actions compose via UNIX pipes: output of one action becomes input of the next.",