atris 3.68.2 → 3.68.3

package/atris/skills/engines/SKILL.md

@@ -93,7 +93,7 @@ Smoke: `reply with the word OK`, run from `$HOME`, 120s cap, one at a time on a
 | Claude | `claude -p "<prompt>" --model claude-opus-5-5 --output-format json --no-session-persistence` | `claude-opus-5-5` | 2026-09-23 | 20.3 | `--model opus` also resolves to `claude-opus-5-5` today; pin the id for reproducibility. |
 | Codex | `codex exec --dangerously-bypass-approvals-and-sandbox --ephemeral -m gpt-6-sol -c model_reasoning_effort=high -o <result-file> "<prompt>" </dev/null` (run from the target repo/worktree; read-only: swap the bypass flag for `--sandbox read-only`) | `gpt-6-sol`, high | 2026-09-23 | 13.8 | Keep the `</dev/null`: with stdin open codex prints "Reading additional input from stdin..." and waits forever (observed 2026-09-23). Outside a git repo add `--skip-git-repo-check` or it exits 1 (re-checked 2026-09-23). Final answer lands in the `-o` file. Run as a tracked background Bash task; in one-shot `claude -p` workers run it in the FOREGROUND, a one-shot session never wakes again (2026-08-11). `codex-companion.mjs task --background` is deprecated: its job store never notifies the session (2026-08-10). |
 | Cursor | `cursor-agent --trust --model composer-2.5 -p "<prompt>"` (run from the target repo) | `composer-2.5` | 2026-09-23 | 37.5 | `--trust` required for non-interactive. |
-| Fable | `atris engine fable --timeout 120 --json "<question>"` | `claude-fable-5` | 2026-09-23 | 21.4 | Canonical read-only FABLE ask with receipt; scale `--timeout` to the work. |
+| Fable | `atris engine fable --timeout 120 --json "<question>"` | `claude-fable-5-1` | not re-run on 5.1 (2026-09-23 on `claude-fable-5`) | 21.4 | Canonical read-only FABLE ask with receipt; scale `--timeout` to the work. |
 | Composer | `cursor-agent --trust --model composer-2.5 -p "<prompt>"` | `composer-2.5` via Cursor | 2026-09-23 | 37.5 | Same binary as Cursor. |
 | Haiku | `claude -p "<prompt>" --model claude-haiku-5-5 --output-format json --no-session-persistence` | `claude-haiku-5-5` | 2026-10-07 | 9.5 | Fast validation and bounded read-only checks. |
 | Devin (build) | `devin -p --permission-mode dangerous --model swe-2-max --prompt-file <brief>` (run from an isolated worktree) | `swe-2-max` (Free tag in `devin models list`, 2026-09-23) | 2026-09-23 | 14.0 | In `-p` mode the default permission mode rejects shell commands too, so even a read-only audit dies after one step (2026-09-23). Keep the brief outside the worktree. Long runs (5+ min) can print nothing even on success: judge by the worktree diff. `devin cloud` outlives this machine. |

package/atris/team/customer-lead/MEMBER.md

@@ -2,7 +2,7 @@
 name: customer-lead
 role: Customer Lead
 description: Keeps customer commitments visible, spots retention risk, and turns evidence into the next useful follow-up.
-version: 1.0.0
+version: 2.0.0
 
 skills:
   - customer-commitments

package/atris/team/executor/MEMBER.md

@@ -2,9 +2,10 @@
 name: executor
 role: Builder
 description: Execute from build specs, one step at a time
-version: 1.0.0
+version: 2.0.0
 
 skills:
+  - isolated-slice-builder
   - code-writer
   - test-runner
 
@@ -40,16 +41,16 @@ tools: []
 
 ## Your Job
 
-When navigator hands you build.md:
+When navigator or the orchestrator hands you a build spec or idea:
 
-1. **Read build.md** — Exact files, steps, error cases
-2. **Execute one step at a time** — Never batch multiple steps
-3. **Show ASCII progress** — After each step, show what happened
-4. **Wait for confirmation** — Human approves before next step
-5. **Final summary** — When done, show ASCII completion status
-6. **Keep wiki memory honest** — If the task changes durable project knowledge, update the relevant `atris/wiki/` page or note the source for immediate ingest
+1. **Read the spec and MAP.md**: Exact files, steps, and edge cases
+2. **Execute one bounded slice per loop**: Stay in the declared file scope
+3. **Prove with bare checks**: Run the real test command (`node --test ...`) and `git diff --check` before finishing
+4. **Pass cross-model review**: Fix any referee feedback before marking work ready for human review
+5. **Final summary**: Report exact files changed, check command run, and plain-language outcome
+6. **Keep wiki and member memory honest**: Append the lesson to `logs/<today>.md` and update `atris/MAP.md` when routes change
 
-**DO NOT skip steps. DO NOT batch. One shot at a time.**
+**Never claim progress without a passing check. Stop for human approval only before destructive actions or out-of-scope refactors.**
 
 ---
 
@@ -114,15 +115,15 @@ If new evidence changes the task, stop and add a task note instead of forcing co
 
 ## Rules
 
-1. **Read build.md first** — Never guess, always follow the spec
-2. **One step at a time** — Show ASCII after each, wait for confirmation
-3. **Check MAP.md** — Verify file paths exist, update if structure changed
-4. **Run tests after changes** — Catch issues immediately
-5. **Check wiki context** — Read `atris/wiki/STATUS.md` if present before executing in a known area
-6. **No shortcuts** — Follow the build.md steps exactly
-7. **Anti-slop aware** — Read `atris/policies/ANTISLOP.md` before writing. No sparkles, no filler, no purple prose.
-8. **Stay in scope** — Only touch files listed in the task. If you need to change something outside scope, stop and flag it. That's a new task.
-9. **If no exit condition, stop** — A task without a clear "done" definition is not ready for execution. Send it back to navigator.
+1. **Read build.md first**: Never guess, always follow the spec
+2. **One step at a time**: Show ASCII after each, wait for confirmation
+3. **Check MAP.md**: Verify file paths exist, update if structure changed
+4. **Run tests after changes**: Catch issues immediately
+5. **Check wiki context**: Read `atris/wiki/STATUS.md` if present before executing in a known area
+6. **No shortcuts**: Follow the build.md steps exactly
+7. **Anti-slop aware**: Read `atris/policies/ANTISLOP.md` before writing. No sparkles, no filler, no purple prose.
+8. **Stay in scope**: Only touch files listed in the task. If you need to change something outside scope, stop and flag it. That's a new task.
+9. **If no exit condition, stop**: A task without a clear "done" definition is not ready for execution. Send it back to navigator.
 
 ---
 

package/atris/team/executor/skills/isolated-slice-builder/SKILL.md

@@ -0,0 +1,21 @@
+---
+name: isolated-slice-builder
+description: Build one bounded code slice in an isolated checkout, prove it with a bare test command and git diff --check, and update MAP.md before review.
+version: 1.0.0
+tags:
+  - build
+  - verification
+  - worktree
+---
+
+# Isolated Slice Builder
+
+Use this skill when building a feature slice or dispatched task in an isolated checkout.
+
+## Workflow
+
+1. Read `atris/MAP.md` to jump directly to the target `file:line` locations.
+2. Edit only the files required for the current slice, using Node built-ins only.
+3. Add or update a regression test that exercises the real runtime behavior.
+4. Run the bare test command (`node --test <file>`) and `git diff --check`.
+5. Update `atris/MAP.md` if any exported function or command route changed.

package/atris/team/navigator/MEMBER.md

@@ -2,8 +2,10 @@
 name: navigator
 role: System Navigator
 description: Transform messy human intent into precise execution plans
-version: 1.0.0
+version: 2.0.0
 agent-id: navigator-3
+skills:
+  - loophole-proof-planner
 permissions:
   can-read: true
   can-plan: true

package/atris/team/navigator/skills/loophole-proof-planner/SKILL.md

@@ -0,0 +1,20 @@
+---
+name: loophole-proof-planner
+description: Turn messy operator intent into single-slice tasks with exact file:line pointers, falsifiable bare test commands, and explicit rollback plans.
+version: 1.0.0
+tags:
+  - planning
+  - navigation
+  - confidence-gate
+---
+
+# Loophole Proof Planner
+
+Use this skill when turning a broad idea or goal into an executable roadmap.
+
+## Workflow
+
+1. Read `atris/MAP.md` first to locate exact `file:line` references without blind searching.
+2. Split work touching 4 or more files into ordered single-slice tasks.
+3. Attach a bare runnable check command (`node --test <file>`) and a rollback step to every task.
+4. Close every loophole (missing owner, vague exit condition, unrunnable check) before handing off to `executor`.

package/atris/team/researcher/MEMBER.md

@@ -2,9 +2,10 @@
 name: researcher
 role: Deep Researcher
 description: Find ground truth on any topic — competitors, standards, technologies, markets
-version: 1.0.0
+version: 2.0.0
 
 skills:
+  - grounded-evidence-brief
   - deep-search
   - source-verification
 

package/atris/team/researcher/skills/grounded-evidence-brief/SKILL.md

@@ -0,0 +1,19 @@
+---
+name: grounded-evidence-brief
+description: Answer open technical or product questions using primary sources, exact file:line citations, and one actionable recommendation.
+version: 1.0.0
+tags:
+  - research
+  - evidence
+  - decision
+---
+
+# Grounded Evidence Brief
+
+Use this skill when investigating an unknown, comparing approaches, or benchmarking options before building.
+
+## Workflow
+
+1. Check local repo code and receipts first before external sources.
+2. Cite exact `file:line` paths or primary URLs for every factual claim.
+3. End with one concrete recommendation and the member that should execute it.

package/atris/team/validator/MEMBER.md

@@ -2,9 +2,10 @@
 name: validator
 role: Reviewer
 description: Validate execution, run tests, ensure quality before shipping
-version: 1.0.0
+version: 2.0.0
 
 skills:
+  - cross-model-referee
   - test-runner
   - doc-updater
 
@@ -30,12 +31,6 @@ tools: []
 
 **Validation:** Run `npm test` to verify changes work correctly.
 
-## Project Context
-
-**Project Type:** nodejs (nodejs)
-
-**Validation:** Run `npm test` to verify changes work correctly.
-
 ---
 
 ## MAPFIRST (Before ANY Validation)

package/atris/team/validator/skills/cross-model-referee/SKILL.md

@@ -0,0 +1,20 @@
+---
+name: cross-model-referee
+description: Audit worker diffs from an independent model perspective, run bare test commands, block slop or unverified claims, and emit actionable fix instructions.
+version: 1.0.0
+tags:
+  - review
+  - verification
+  - referee
+---
+
+# Cross Model Referee
+
+Use this skill when validating a plan before execution or auditing a worker diff before merge.
+
+## Workflow
+
+1. Inspect `git status` and `git diff` directly rather than trusting the worker's summary.
+2. Run the declared test check bare (`node --test <file>`) and verify zero new dependencies in `package.json`.
+3. Check that `atris/MAP.md` was updated if any file or command route changed.
+4. Emit a concrete pass verdict or bounce with the exact line and failure that must be fixed.

package/commands/autoland.js

@@ -379,6 +379,7 @@ function sweepLanding(root, { ttlDays, staleHours, now = Date.now() } = {}) {
     remote: false,
     includeDetached: false,
     now,
+    by: 'atris autoland',
   });
   const afterBoard = collectBoard(root, { ttlDays, staleHours, now, light: true });
   const human = [];
@@ -926,7 +927,7 @@ async function runTickBody(root, { json, policy, receipt, engineValidationDeps =
     }
     try {
       const { cleanupWorktrees } = require('./worktree');
-      const cleaned = cleanupWorktrees({ root, apply: true });
+      const cleaned = cleanupWorktrees({ root, apply: true, by: 'atris autoland' });
       receipt.worktrees_reaped = cleaned.removed.length;
       if (cleaned.removed.length > 0) receipt.worktrees_reaped_paths = cleaned.removed.map((w) => w.path);
     } catch (err) {

package/commands/engine.js

@@ -38,7 +38,7 @@ const { teamRosterView } = require('../lib/member-engine');
 const { engineModelText, engineRunsView, availableModels, modelLabel } = require('../lib/roster-models');
 const { renderTable } = require('../lib/text-table');
 const { TEAM_FOOTER } = require('../lib/team-lineup');
-const { readRosterRuns, workerRuns, summarizeRuns, recentRuns, renderRunLine, DEFAULT_DAYS: RUN_DAYS } = require('../lib/roster-runs');
+const { readRosterRuns, workerRuns, summarizeRuns, recentRuns, renderRunLine, runView, DEFAULT_DAYS: RUN_DAYS } = require('../lib/roster-runs');
 const { attachSuggestions } = require('../lib/roster-suggest');
 const { PIN_MAX_AGE_DAYS: PIN_STALE_DAYS, attachStaleModels, modelCatalog } = require('../lib/roster-stale');
 const {
@@ -1507,7 +1507,7 @@ const RUNS_LIST_LIMIT = 20;
 function runRosterRunsCommand(job, json, root, now) {
   const runs = recentRuns(readRosterRuns(root, { now, days: RUNS_LIST_DAYS }), { job, limit: RUNS_LIST_LIMIT });
   if (json) {
-    console.log(JSON.stringify({ job: job || null, days: RUNS_LIST_DAYS, runs }, null, 2));
+    console.log(JSON.stringify({ job: job || null, days: RUNS_LIST_DAYS, runs: runs.map(runView) }, null, 2));
     return 0;
   }
   if (!runs.length) {

package/commands/goal.js

@@ -168,9 +168,65 @@ function renderGoalView(view) {
   return lines.join('\n');
 }
 
+// A goal that never becomes a mission run is a wish. `goal set` therefore also
+// files the sentence through the normal wish capture path (capture only: no
+// mission, no driver), so the goal -> wish -> mission -> run chain starts here.
+// Wish capture is best-effort: it must never fail or undo the goal itself.
+function wishNextStep(wish, goal) {
+  return wish.status === 'needs_input'
+    ? 'atris wish answer "your words"'
+    : `atris wish again ${wish.id} "${goal}"`;
+}
+
+function captureGoalWish(sentence, root) {
+  const goal = String(sentence || '').replace(/\s+/g, ' ').trim();
+  const { isFreshWorkspace } = require('../lib/first-minute');
+  if (isFreshWorkspace(root)) {
+    console.error('note: wish not filed, no atris/ workspace here. run: atris init');
+    return null;
+  }
+  const previous = process.cwd();
+  let wish = null;
+  try {
+    if (previous !== root) process.chdir(root);
+    // Re-setting the same goal must not stack duplicate wishes.
+    const { readWishes } = require('../lib/wish-store');
+    const open = readWishes(root).find((w) => (
+      normalizeText(w.text) === normalizeText(goal)
+      && ['captured', 'captured_no_mission', 'needs_input'].includes(w.status)
+    ));
+    if (open) {
+      console.log(`wish already filed: ${open.id}`);
+      console.log(`next      ${wishNextStep(open, goal)}`);
+      return open;
+    }
+    const { runCapturedWish } = require('../lib/wish-delegate');
+    runCapturedWish(goal, root, {
+      noMission: true,
+      afterCapture: (captured) => { wish = captured; },
+    });
+  } catch (error) {
+    console.error(`warning: goal set, but wish capture failed: ${normalizeText(error?.message || error)}`);
+    return null;
+  } finally {
+    if (process.cwd() !== previous) process.chdir(previous);
+  }
+  if (!wish) return null;
+  let latest = wish;
+  try {
+    const { readWishes } = require('../lib/wish-store');
+    latest = readWishes(root).find((w) => w.id === wish.id) || wish;
+  } catch {
+    latest = wish;
+  }
+  console.log(`wish id   ${wish.id}`);
+  console.log(`next      ${wishNextStep(latest, goal)}`);
+  return latest;
+}
+
 function printHelp() {
   console.log('usage: atris goal [--json]');
-  console.log('       atris goal set "<sentence>"');
+  console.log('       atris goal set "<sentence>" [--no-wish]');
   console.log('       atris goal metric <name> <value>');
   console.log('       atris goal metric <name> --rm');
 }
@@ -187,7 +243,9 @@ function run(args = [], context = {}) {
     }
 
     if (command === 'set') {
-      const goal = normalizeText(args.slice(1).join(' '));
+      const noWish = args.includes('--no-wish');
+      const sentence = args.slice(1).filter((arg) => arg !== '--no-wish').join(' ');
+      const goal = normalizeText(sentence);
       if (!goal) {
         console.error('error: goal sentence is required');
         return 2;
@@ -195,6 +253,7 @@ function run(args = [], context = {}) {
       const scoreboard = readScoreboard(root);
       writeScoreboard({ ...scoreboard, goal }, root);
       console.log('goal set.');
+      if (!noWish) captureGoalWish(sentence, root);
       return 0;
     }
 

package/commands/land.js

@@ -2,7 +2,7 @@ const fs = require('fs');
 const path = require('path');
 const gitSpawn = require('../lib/git-spawn');
 const { compactErrorPayload, printCliJson } = require('../lib/cli-json');
-const { listWorktrees, statusCounts } = require('./worktree');
+const { listWorktrees, statusCounts, defaultMainlineBase, listActiveProcessCwds, worktreeHasActiveProcess, recordWorktreeRemoval } = require('./worktree');
 
 const DEFAULT_TTL_DAYS = 7;
 const DEFAULT_STALE_HOURS = 48;
@@ -38,10 +38,11 @@ function refExists(root, ref) {
 
 function baseBranch(root, override = '') {
   if (override && refExists(root, override)) return override;
-  if (refExists(root, 'origin/master')) return 'origin/master';
-  if (refExists(root, 'origin/main')) return 'origin/main';
-  if (refExists(root, 'master')) return 'master';
-  if (refExists(root, 'main')) return 'main';
+  const mainline = defaultMainlineBase(root);
+  if (mainline !== 'HEAD') return mainline;
+  for (const candidate of ['main', 'master']) {
+    if (refExists(root, candidate)) return candidate;
+  }
   return '';
 }
 
@@ -309,7 +310,7 @@ function collectBoard(root, { ttlDays = DEFAULT_TTL_DAYS, staleHours = DEFAULT_S
   const cherryCache = loadCherryCache(root);
   const baseSha = resolveSha(root, base);
   for (const branch of listBranches(root, base)) {
-    if (PROTECTED_BRANCHES.has(branch.name)) continue;
+    if (PROTECTED_BRANCHES.has(branch.name) || branch.name === base.replace(/^origin\//, '')) continue;
     const lastCommitMs = Number(branch.lastCommitUnix) * 1000;
     if (Number.isFinite(lastCommitMs)) lastCommitMsByBranch.set(branch.name, lastCommitMs);
     const ahead = branch.ahead !== undefined ? branch.ahead : aheadCount(root, base, branch.name);
@@ -527,6 +528,15 @@ function salvageWorktree(w, dir, receipt) {
   }
 }
 
+function reapReason(w, board, ttlDays, force) {
+  const state = (board.branches.find((b) => b.name === w.branch) || {}).state;
+  if (state === 'landed') return 'branch_landed';
+  if (state === 'due') return 'branch_past_ttl';
+  if (w.state === 'detached') return 'detached_head';
+  if (typeof w.ageDays === 'number' && w.ageDays > ttlDays) return 'worktree_past_ttl';
+  return force ? 'forced' : 'unlanded_commits';
+}
+
 // Reap: salvage then delete everything landed (residue) or past TTL.
 // Salvage first, always: unlanded commits go into a git bundle, dirty
 // worktrees into patch files + untracked-file copies, so a reap is never
@@ -534,7 +544,9 @@ function salvageWorktree(w, dir, receipt) {
 // includeDetached: detached-HEAD worktrees have no branch, so the salvage
 // bundle cannot cover their commits, unattended reaps (autoland) pass false
 // and leave them for a human reap.
-function reap(root, { ttlDays = DEFAULT_TTL_DAYS, staleHours = DEFAULT_STALE_HOURS, base: baseOverride = '', dryRun = false, remote = true, includeDetached = true, protectCurrent = true, force = false, now = Date.now() } = {}) {
+function reap(root, { ttlDays = DEFAULT_TTL_DAYS, staleHours = DEFAULT_STALE_HOURS, base: baseOverride = '', dryRun = false, remote = true, includeDetached = true, protectCurrent = true, force = false, now = Date.now(), activeCwds: activeCwdsOverride, by = 'atris land --reap' } = {}) {
+  const readActiveCwds = () => (Array.isArray(activeCwdsOverride) ? activeCwdsOverride : listActiveProcessCwds());
+  let activeCwds = null;
   const board = collectBoard(root, { ttlDays, staleHours, base: baseOverride, now });
   const targets = board.branches.filter((b) => b.state === 'landed' || b.state === 'due');
   const targetNames = new Set(targets.map((b) => b.name));
@@ -554,6 +566,13 @@ function reap(root, { ttlDays = DEFAULT_TTL_DAYS, staleHours = DEFAULT_STALE_HOU
       protectedWorktrees.push({ ...w, reason: 'current_checkout' });
       continue;
     }
+    // A run working inside the folder outranks --force: removing it pulls
+    // the floor out from under a live engine.
+    if (!activeCwds) activeCwds = readActiveCwds();
+    if (worktreeHasActiveProcess(w.path, activeCwds)) {
+      protectedWorktrees.push({ ...w, reason: 'active_process' });
+      continue;
+    }
     if (worktreeWithinReapGrace(w.path, now)) {
       protectedWorktrees.push({ ...w, reason: 'fresh_worktree_grace' });
       continue;
@@ -642,7 +661,14 @@ function reap(root, { ttlDays = DEFAULT_TTL_DAYS, staleHours = DEFAULT_STALE_HOU
   const survivingWorktreeTargets = worktreeTargets.filter(
     (w) => targetNames.has(w.branch) || (includeDetached && w.state === 'detached')
   );
+  const activeAtRemoval = survivingWorktreeTargets.length ? readActiveCwds() : [];
   for (const w of survivingWorktreeTargets) {
+    // A run can start in the folder while the bundle is built; look again.
+    if (worktreeHasActiveProcess(w.path, activeAtRemoval)) {
+      receipt.keptWorktrees.push(`${w.path} (active_process)`);
+      if (w.branch) targetNames.delete(w.branch);
+      continue;
+    }
     // The grace check ran at scan time, and a bundle build can sit between it
     // and removal — an engine that booted mid-sweep refreshes the directory
     // mtime in that gap. Stat once more at removal time, the cheapest place
@@ -665,6 +691,7 @@ function reap(root, { ttlDays = DEFAULT_TTL_DAYS, staleHours = DEFAULT_STALE_HOU
     const removed = runGit(['worktree', 'remove', '--force', w.path], { cwd: root, check: false });
     if (removed.status === 0) {
       receipt.removedWorktrees.push(w.path);
+      recordWorktreeRemoval(root, { path: w.path, branch: w.branch, by, reason: reapReason(w, board, ttlDays, force) });
     } else {
       // A locked worktree fails here and, unreported, reap re-salvages the
       // same patches every pass forever. Say what blocked it so a human can