@redue/cli 0.1.0-alpha.14 → 0.1.0-alpha.15

package/CHANGELOG.md

@@ -1,5 +1,22 @@
 # Changelog
 
+## 0.1.0-alpha.15 — 2026-10-22
+
+No change to observation, status or `redue run` since alpha.14; the changes
+are to the Claude Code integration.
+
+- The CLAUDE.md block now tells agents to run the typecheck or test command
+  before saying it passes, even after an edit that seems irrelevant, and to
+  leave skipping to REDUE's hook. In the alpha.14 eval, two of six mixed-task
+  sessions skipped their final test run on their own after a README edit;
+  with the new block, none of six did (twilio and SwitchHosts, CI
+  39702429765), at the same token cost. Run `redue agent setup claude --apply`
+  again to update the block; the earlier wording is recognized and replaced.
+- The hook also matches a typecheck or test command after a leading
+  `cd <project root>;` or `cd <project root> &&`. A skip denies the whole
+  command; a rewrite keeps the `cd`. A `cd` to any other directory matches
+  nothing.
+
 ## 0.1.0-alpha.14 — 2026-10-21
 
 Bench numbers below are from one 800-file project on macOS arm64, Node 22.13.0.

package/README.md

@@ -37,7 +37,7 @@ it; relevant source edits make it stale. [Run the reproducible demo](docs/contri
 
 ## Install the alpha
 
-The opt-in pilot uses **0.1.0-alpha.14**. Follow the
+The opt-in pilot uses **0.1.0-alpha.15**. Follow the
 [fixed-version pilot guide](docs/user/pilot.md) for supported projects, setup,
 updates and voluntary feedback. No external usage results are claimed yet.
 
@@ -156,7 +156,9 @@ Medians per task, summed over the three projects:
   source-change task skipped nothing; its time difference is session variance.
   In two mixed-task sessions the agent left out its final test run on its own
   after a README-only edit, without asking REDUE, and said so; the eval report
-  lists such sessions. Cost was the same ($0.54 without, $0.56 with).
+  lists such sessions. Since alpha.15 the agent instructions ask agents to run
+  the check and leave skipping to REDUE; in a 12-session rerun of that task,
+  no agent left it out. Cost was the same ($0.54 without, $0.56 with).
 
 ## What the states mean
 

package/docs/user/pilot.md

@@ -1,4 +1,4 @@
-# Opt-in development pilot: alpha.14
+# Opt-in development pilot: alpha.15
 
 Use one fixed build for several ordinary development sessions. The owner will
 invite two or three willing testers and approve their projects; nothing is
@@ -35,7 +35,7 @@ are **NOT RUN**. The recording shortcut is macOS/direct-recipe only.
 For a new default npm installation:
 
 ```sh
-npm install --global @redue/cli@0.1.0-alpha.14
+npm install --global @redue/cli@0.1.0-alpha.15
 redue --version
 cd my-project
 redue --config redue.config.json init --recipe typescript-direct --check typecheck --dry-run
@@ -89,7 +89,7 @@ using it; keep the same npm prefix, config and any explicit state selection.
 
 ```sh
 redue --config redue.config.json stop
-npm install --global @redue/cli@0.1.0-alpha.14
+npm install --global @redue/cli@0.1.0-alpha.15
 redue --config redue.config.json agent setup codex --dry-run
 redue --config redue.config.json agent setup codex --apply
 redue --config redue.config.json start

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@redue/cli",
   "private": false,
-  "version": "0.1.0-alpha.14",
+  "version": "0.1.0-alpha.15",
   "publishConfig": {
     "access": "public",
     "tag": "alpha"

package/src/agent-integration.mjs

@@ -88,23 +88,39 @@ function saved(root,host){const text=read(root,marker(host));if(text===null)retu
   return value;
 }
 // The managed block is always in the agent's context, so it stays short. For
-// Claude Code it names the typecheck and test commands the hook matches and
-// what a skip means. The hook does the rest, so the block does not point to
-// the skill; the skill stays for agents that look for it. A typecheck-only
-// block keeps its earlier wording.
+// Claude Code it names the typecheck and test commands the hook matches, asks
+// the agent to run them rather than judge for itself that a rerun is not
+// needed, and says what a skip means. The hook does the rest, so the block
+// does not point to the skill; the skill stays for agents that look for it.
 const quoted=forms=>forms.map(form=>`\`${form}\``).join(' and ');
+const commandsOf=rows=>[rows.filter(row=>row.kind!=='test').map(row=>row.form),rows.filter(row=>row.kind==='test').map(row=>row.form)];
+const fence=(host,body)=>`\n\n<!-- REDUE:${host}:begin -->\n${body}<!-- REDUE:${host}:end -->\n`;
+const skillPointer=host=>`Before you run, skip or report a typecheck or other configured check, read\n${hosts[host].skill}.\n`;
 function blockFor(host,rows=[]){
-  const typechecks=rows.filter(row=>row.kind!=='test').map(row=>row.form),tests=rows.filter(row=>row.kind==='test').map(row=>row.form);
-  return `\n\n<!-- REDUE:${host}:begin -->\n`+
-  (!rows.length?`Before you run, skip or report a typecheck or other configured check, read\n`+
-    `${hosts[host].skill}.\n`:!tests.length?`Typecheck with ${quoted(typechecks)}. If its last pass still applies to\n`+
+  const [typechecks,tests]=commandsOf(rows);
+  if(!rows.length)return fence(host,skillPointer(host));
+  if(!tests.length)return fence(host,`Typecheck with ${quoted(typechecks)}. Run it before you say it passes, even after an\n`+
+    `edit you think cannot matter; don't skip it yourself. If its last pass still\n`+
+    `applies to these exact files, a REDUE hook skips the run and says so. That\n`+
+    `skip counts as a pass.\n`);
+  return fence(host,(typechecks.length?`Typecheck with ${quoted(typechecks)}.\n`:'')+
+    `Run the unit tests with ${quoted(tests)}.\n`+
+    `Run a check before you say it passes, even after an edit you think cannot\n`+
+    `matter; don't skip it yourself. If its last pass still applies to these exact\n`+
+    `files, a REDUE hook skips the run and says so. That skip counts as a pass.\n`);
+}
+// The alpha.12-alpha.14 wording, recognized so setup can replace it.
+function earlierBlock(host,rows=[]){
+  const [typechecks,tests]=commandsOf(rows);
+  if(!rows.length)return fence(host,skillPointer(host));
+  if(!tests.length)return fence(host,`Typecheck with ${quoted(typechecks)}. If its last pass still applies to\n`+
     `these exact files, a REDUE hook skips the run and says so. The skip counts as\n`+
-    `a pass; don't check it again.\n`:
-    (typechecks.length?`Typecheck with ${quoted(typechecks)}.\n`:'')+
+    `a pass; don't check it again.\n`);
+  return fence(host,(typechecks.length?`Typecheck with ${quoted(typechecks)}.\n`:'')+
     `Run the unit tests with ${quoted(tests)}.\n`+
     `If a check's last pass still applies to these exact files, a REDUE hook skips\n`+
-    `the run and says so. The skip counts as a pass; don't check it again.\n`)+
-  `<!-- REDUE:${host}:end -->\n`;}
+    `the run and says so. The skip counts as a pass; don't check it again.\n`);
+}
 // One form per check: the npm script when there is one, since that is what
 // the project's own instructions usually name.
 const blockForms=(host,config,relative)=>host!=='claude'?[]:hookForms(config,relative).map(row=>({kind:row.kind,
@@ -131,7 +147,7 @@ function knownBlock(host,relative,block){
   const forms=pattern=>{const named=new RegExp(pattern).exec(block);return named?[...named[1].matchAll(/`([^`]+)`/g)].map(m=>m[1]):[];};
   const rows=[...forms(`^\\n\\n<!-- REDUE:[a-z]+:begin -->\\nTypecheck with ${list}\\.`).map(form=>({kind:'typecheck',form})),
     ...forms(`Run the unit tests with ${list}\\.\\n`).map(form=>({kind:'test',form}))];
-  return block===blockFor(host,rows)||priorBlocks(host,relative).includes(block);
+  return block===blockFor(host,rows)||block===earlierBlock(host,rows)||priorBlocks(host,relative).includes(block);
 }
 // The policy with the exact commands for this config, so an agent never builds
 // a --config form the hook would not recognize. Claude Code gets the typecheck

package/src/claude-hook.mjs

@@ -7,7 +7,7 @@ import fs from 'node:fs';
 import path from 'node:path';
 import {spawn} from 'node:child_process';
 import {projectState} from './project-location.mjs';
-import {realObservedPath} from './path-identity.mjs';
+import {realObservedPath,samePath} from './path-identity.mjs';
 
 export const hookCommand='redue agent hook claude';
 const safe=/^[A-Za-z0-9_./:=,+-]+$/;
@@ -73,6 +73,21 @@ export function recordedCommand(command,name,configFile=defaultConfig){
   return parts.base===run?null:run+parts.suffix;
 }
 
+// A leading `cd DIR;` or `cd DIR &&` into the project root, which agents
+// often add before a check: {prefix, command} with the rest of the command,
+// the whole command when there is no such prefix, or null when it enters
+// another directory. DIR may be quoted; it may not use variables, a leading
+// `~` or escapes. Inside double quotes a backslash before anything but
+// $ ` " \ is literal, so a quoted Windows path matches; unquoted it is an escape.
+const cdPrefix=/^[ \t]*cd[ \t]+(?:"((?:[^"$`\\]|\\(?![$`"\\\n]))*)"|'([^']*)'|((?!~)[^\s;&|'"$`\\]+))[ \t]*(?:;|&&)[ \t]*/;
+export function rootCd(command,cwd,projectRoot){
+  const prefix=typeof command==='string'?cdPrefix.exec(command):null;
+  if(!prefix)return {prefix:'',command};
+  const dir=prefix[1]??prefix[2]??prefix[3];
+  try{if(!samePath(realObservedPath(path.resolve(cwd,dir)),projectRoot))return null;}catch{return null;}
+  return {prefix:prefix[0],command:command.slice(prefix[0].length)};
+}
+
 export function verdict(status,name,rest=null){
   const row=status?.schema===1&&status.observation?.healthy===true&&Array.isArray(status.checks)?
     status.checks.find(item=>item?.name===name):null;
@@ -129,7 +144,9 @@ export async function claudeHook(input,{entry,timeout}){
   if(record?.schema!==1||record.host!=='claude'||typeof record.config!=='string'||
     !/^[^/\\]+\.json$/.test(record.config))return null;
   const config=path.join(root,record.config);
-  const parsed=JSON.parse(fs.readFileSync(config,'utf8')),match=matched(command,parsed,record.config);
+  const parsed=JSON.parse(fs.readFileSync(config,'utf8'));
+  const entered=rootCd(command,root,realObservedPath(path.resolve(path.dirname(config),parsed.root||'.')));
+  const match=entered&&matched(entered.command,parsed,record.config);
   if(!match)return null;
   const name=match.check,status=await synchronizedStatus(entry,config,root,timeout??lastRunTimeout(config,parsed,name)),
     reason=verdict(status,name,match.rest);
@@ -140,12 +157,12 @@ export async function claudeHook(input,{entry,timeout}){
   // stays too: the rewrite returns "allow", which would also skip the prompt
   // for the rest of the chain.
   const rewritten=status?.schema===1&&status.observation?.healthy===true&&Array.isArray(status.checks)&&
-    status.checks.some(item=>item?.name===name)?recordedCommand(command,name,record.config):null;
+    status.checks.some(item=>item?.name===name)?recordedCommand(entered.command,name,record.config):null;
   // updatedInput replaces the whole tool input. Deny and ask rules still apply
   // to the rewritten command; "allow" skips only the prompt.
   return rewritten?{hookSpecificOutput:{hookEventName:'PreToolUse',permissionDecision:'allow',
     permissionDecisionReason:`REDUE: check ${JSON.stringify(name)} is not CURRENT; running it through redue run so the result is recorded.`,
-    updatedInput:{...input.tool_input,command:rewritten}}}:null;
+    updatedInput:{...input.tool_input,command:entered.prefix+rewritten}}}:null;
 }
 
 export async function runClaudeHook(entry){