loadout-ai 0.7.0 → 0.8.0

package/docs/FEATURE_TEST_MATRIX.md

@@ -98,7 +98,6 @@ Confirm isolation before any applied command:
 
 ```bash
 loadout status --json
-loadout capabilities
 ```
 
 Expected: paths, if shown, are below `TEST_ROOT`; Codex and Claude Code are detected;
@@ -158,8 +157,6 @@ loadout library --json
 loadout health --explain --json
 loadout report --json
 loadout card --json
-loadout outcomes --json
-loadout capabilities --inspect --json
 ```
 
 Expected:
@@ -184,7 +181,6 @@ Network variants must be run deliberately:
 loadout catalog --refresh --json
 loadout health --updates --json
 loadout scan --agents codex,claude-code --refresh-provenance --json
-loadout compare brainstorming --offline --json
 loadout discover --source mcp-registry --limit 10 --json
 loadout discover --source skills-sh --limit 10 --json
 ```
@@ -201,79 +197,7 @@ skills.sh path needs its request-scoped `VERCEL_OIDC_TOKEN`; without one it must
 previous complete cache or return an attributed `unavailable` result without making
 an unauthenticated request. Neither source installs or promotes a lead.
 
-## 3. Package, manifest, lock, portability, and registry track (S/A)
-
-Create a package _inside_ the test project so its manifest can be exported portably:
-
-```bash
-mkdir -p "$TEST_PROJECT/packages"
-loadout create "$TEST_PROJECT/packages/matrix-demo" \
-  --name matrix-demo --description "Disposable matrix package"
-loadout pack "$TEST_PROJECT/packages/matrix-demo" --json
-loadout publish "$TEST_PROJECT/packages/matrix-demo" --local
-loadout search matrix-demo --json
-
-loadout init --path "$TEST_PROJECT/loadout.json" --name matrix \
-  --agents codex,claude-code --scope project
-(
-  cd "$TEST_PROJECT"
-  loadout add local-demo --manifest loadout.json --local \
-    --path packages/matrix-demo --agents codex,claude-code
-  loadout sync --manifest loadout.json --lock loadout.lock
-  loadout sync --manifest loadout.json --lock loadout.lock --yes
-  loadout audit --manifest loadout.json --lock loadout.lock --json
-)
-```
-
-Expected: `pack` returns a deterministic digest; local publication is immutable;
-the first `sync` is a dry run; applied sync prints one snapshot; `audit` returns
-`"valid": true`; skills appear only below the disposable profile.
-
-Exercise desired-state editing and portability:
-
-```bash
-(
-  cd "$TEST_PROJECT"
-  loadout lock --manifest loadout.json --output loadout.lock
-  loadout export portable.json --manifest loadout.json --lock loadout.lock
-  loadout import portable.json --manifest imported.json --lock imported.lock
-  loadout import portable.json --manifest imported.json --lock imported.lock --yes
-  loadout unadd local-demo --manifest imported.json
-)
-```
-
-Expected: import previews before writing; applied import snapshots destinations;
-`unadd` changes desired state only and does not delete installed files.
-
-The full authenticated remote-registry protocol, including wrong-token rejection,
-immutable-version conflict rejection, exact digest download, risk approval, and the
-HTTPS/non-loopback boundary, is reproducibly covered by:
-
-```bash
-npx vitest run tests/registry-api.test.ts tests/package.test.ts
-```
-
-Manual `registry-serve`/remote `publish` is an X test. After completing the throwaway
-credential setup in track 9, start this in one terminal:
-
-```bash
-loadout registry-serve --port 7331 \
-  --credential-keychain loadout-registry-test --credential-account tester
-```
-
-Then publish from another terminal:
-
-```bash
-loadout publish "$TEST_PROJECT/packages/matrix-demo" \
-  --registry-url http://127.0.0.1:7331 \
-  --credential-keychain loadout-registry-test --credential-account tester
-```
-
-Never place the token on the command line. An identical version/content publish may
-be accepted idempotently; changed content at the same version must be rejected. Stop
-the server with Ctrl-C.
-
-## 4. Install, active-set, outcome, and rollback track (A; Maximum is N)
+## 3. Install, active-set, and rollback track (A; Maximum is N)
 
 Exercise the new-user golden path before its constituent commands:
 
@@ -321,9 +245,6 @@ loadout disable direct-demo --agents codex
 loadout disable direct-demo --agents codex --yes --json
 loadout enable direct-demo --agents codex
 loadout enable direct-demo --agents codex --yes --json
-loadout outcome direct-demo/direct-demo --agent codex --task testing \
-  --result success
-loadout outcomes --json
 loadout share "$TEST_PROJECT/share.json"
 loadout remove direct-demo
 loadout remove direct-demo --yes
@@ -355,7 +276,7 @@ and preserves matching active Stable units at the same reviewed commit; MCP-only
 packages remain explicit setup items. `--approve-risk` acknowledges displayed static
 findings but does not execute third-party repository scripts.
 
-## 5. Existing-skill provenance, adoption, comparison, and freshness (R/S/A)
+## 4. Existing-skill provenance, adoption, and freshness (R/S/A)
 
 Copy one harmless skill into the disposable unmanaged profile, then inspect it:
 
@@ -366,7 +287,6 @@ cp "$TEST_PROJECT/packages/matrix-demo/skills/matrix-demo/SKILL.md" \
 loadout scan --agents codex --json
 loadout adopt unmanaged-demo --agent codex --json
 loadout adopt unmanaged-demo --agent codex --yes --json
-loadout compare unmanaged-demo --agent codex --offline --json
 ```
 
 Expected: scan labels the copy unmanaged; adoption preview does not change its bytes;
@@ -401,13 +321,12 @@ Only run the second command when the first reports a real reviewed update. Expec
 exact diff/safety plan, a snapshot on success, and refusal when new risky findings are
 not acknowledged with `--approve-risk`.
 
-## 6. Discovery and human review queue (N/S)
+## 5. Discovery and human review queue (N/S)
 
 ```bash
 loadout discover --source hacker-news --limit 20 --min-score 20 --json
 loadout discover --source github --limit 20 --queue --json
 loadout discover --source all --limit 20 --queue --json
-loadout review-queue --decision pending --json
 ```
 
 Expected: results contain source evidence and public repository identifiers; queueing
@@ -415,12 +334,6 @@ deduplicates leads; nothing is promoted, cloned into an agent, or installed.
 
 For a repository printed by the queue:
 
-```bash
-loadout review owner/repository --decision shortlisted
-loadout review-queue --decision shortlisted --json
-loadout review owner/repository --decision ignored
-```
-
 Private GitHub discovery is opt-in and reads `GITHUB_TOKEN` only when `--private` is
 present. Prefer a native credential reference:
 
@@ -431,7 +344,7 @@ loadout discover --source github --private \
 
 Use a low-scope test token. The output and state must never contain its value.
 
-## 7. Static inspection, MCP, conversion, canary, and sandbox (R/S/X)
+## 6. Static inspection, MCP, and conversion (R/S)
 
 Static package analysis never executes package content:
 
@@ -439,8 +352,6 @@ Static package analysis never executes package content:
 loadout inspect --source "$TEST_PROJECT/packages/matrix-demo" --json
 loadout evaluate --source "$TEST_PROJECT/packages/matrix-demo" --json
 loadout mcp --source "$TEST_PROJECT/packages/matrix-demo" --json
-loadout canary --source "$TEST_PROJECT/packages/matrix-demo" \
-  --package matrix-demo --json
 ```
 
 Repeat `inspect`, `evaluate`, or `mcp` with `--repository owner/repository` for the
@@ -469,10 +380,6 @@ loadout mcp-config --config "$TEST_PROJECT/mcp.json" --name local-example \
   --command node --arg server.js
 loadout mcp-config --config "$TEST_PROJECT/mcp.json" --name local-example \
   --command node --arg server.js --yes
-loadout codex-mcp-config --config "$TEST_PROJECT/config.toml" \
-  --name remote-example --url https://example.com/mcp
-loadout codex-mcp-config --config "$TEST_PROJECT/config.toml" \
-  --name remote-example --url https://example.com/mcp --yes
 loadout mcp-recipe --json
 ```
 
@@ -486,148 +393,11 @@ requirements you have reviewed.
 
 Docker sandbox execution is intentionally separate:
 
-```bash
-loadout sandbox-run --source "$TEST_PROJECT/packages/matrix-demo" \
-  --image '<reviewed-image>@sha256:<digest>' \
-  --command node --command --version --json
-loadout sandbox-run --source "$TEST_PROJECT/packages/matrix-demo" \
-  --image '<reviewed-image>@sha256:<digest>' \
-  --command node --command --version \
-  --approve-risk --timeout 30000 --json
-```
-
 Expected: the first invocation refuses/only plans without approval; the approved
 container has a read-only source mount, no inherited secrets, no Docker socket, no
 network, and a time bound. The image may need to be pulled beforehand.
 
-## 8. Signing and head-to-head evidence (S)
-
-Validate the model-free benchmark campaign and card/compare surfaces with their
-deterministic automated contracts:
-
-```bash
-npx vitest run tests/benchmark-campaign.test.ts tests/benchmark-cli.test.ts \
-  tests/loadout-card.test.ts tests/share-report.test.ts
-```
-
-Expected: campaign hashes and paired order are deterministic, every retry is included
-in the worst-case budget, over-budget plans are blocked, resumable metadata contains
-no prompt/output/credential bytes, and aggregate comparison never invents a quality
-delta. See `docs/EVALUATION_PROTOCOL_V1.md` for the campaign JSON contract. These
-tests do not call a model provider and consume no provider credit.
-
-```bash
-loadout keygen --private-key "$TEST_ROOT/private.pem" \
-  --public-key "$TEST_ROOT/public.pem"
-loadout catalog-sign --catalog "$LOADOUT_ROOT/catalog/packages.json" \
-  --private-key "$TEST_ROOT/private.pem" --output "$TEST_ROOT/catalog.signed.json"
-loadout catalog-verify --snapshot "$TEST_ROOT/catalog.signed.json" \
-  --public-key "$TEST_ROOT/public.pem"
-```
-
-Expected: the private key is owner-only and outside the repository; verification
-succeeds; changing any byte in the signed payload makes verification fail.
-
-Preview and apply the same signed catalog inside the disposable profile:
-
-```bash
-loadout catalog-update --source "$TEST_ROOT/catalog.signed.json" \
-  --public-key "$TEST_ROOT/public.pem"
-loadout catalog-update --source "$TEST_ROOT/catalog.signed.json" \
-  --public-key "$TEST_ROOT/public.pem" --yes
-loadout catalog --coverage --json
-```
-
-Expected: preview prints an exact signed diff without mutation; apply creates a
-snapshot and trusted state; the effective catalog re-verifies the stored envelope.
-Repeating `--yes` refuses a replay. Test removal only in the disposable profile and
-only with the separate `--allow-removals` acknowledgement.
-
-The repository's generated feed can be triaged without network access:
-
-```bash
-loadout candidate list --limit 5 --json
-loadout candidate list --query "codex skills"
-loadout capabilities --gaps --json
-loadout recommend --project "$TEST_PROJECT" --agent codex --json
-```
-
-`candidate inspect owner/repository --output ./candidate-dossier.json` is a networked
-test: it performs a real public Git clone and writes a static immutable dossier to
-disposable Loadout state. Review that output before exercising `candidate propose`;
-proposal preview and approved proposal output never mutate the catalog.
-
-Graphify is an explicit executable recipe rather than a broad-setup component. With
-`uv` installed, exercise it only inside the disposable profile:
-
-```bash
-loadout tool
-loadout tool graphify --agents codex
-loadout tool graphify --agents codex --yes --approve-risk
-"$LOADOUT_HOME/runtime/graphify/bin/graphify" --version
-test -f "$LOADOUT_USER_HOME/.codex/skills/graphify/SKILL.md"
-loadout tool graphify --remove
-loadout tool graphify --remove --yes --approve-risk
-test ! -e "$LOADOUT_USER_HOME/.codex/skills/graphify"
-test ! -e "$LOADOUT_HOME/runtime/graphify"
-```
-
-Expected: preview identifies the exact wheel hash and all commands; apply reports
-Graphify 0.9.17, writes only the disposable target and isolated runtime, and removal
-restores the original target. The installer subprocess must not inherit API keys.
-
-Create a deterministic workflow fixture and five declared trials per candidate. This
-is harness input, not model-generated evidence, and it executes no candidate content:
-
-```bash
-node --input-type=module <<'NODE'
-import { writeFile } from "node:fs/promises";
-const root = process.env.TEST_PROJECT;
-const fixture = {
-  id: "matrix-workflow",
-  version: "1",
-  category: "workflow-adherence",
-  requiredActions: ["inspect", "edit", "verify"],
-  forbiddenActions: ["delete-unrelated"]
-};
-const trials = Array.from({ length: 5 }, () => [
-  {
-    candidateId: "baseline",
-    fixtureId: fixture.id,
-    observations: ["inspect", "edit", "verify"],
-    durationMs: 10
-  },
-  {
-    candidateId: "improved",
-    fixtureId: fixture.id,
-    observations: ["inspect", "edit", "verify", "report-uncertainty"],
-    durationMs: 10
-  }
-]).flat();
-await writeFile(`${root}/fixture.json`, JSON.stringify(fixture, null, 2));
-await writeFile(`${root}/trials.json`, JSON.stringify(trials, null, 2));
-NODE
-```
-
-Sign and inspect the resulting evidence:
-
-```bash
-loadout head-to-head --fixture "$TEST_PROJECT/fixture.json" \
-  --trials "$TEST_PROJECT/trials.json" --private-key "$TEST_ROOT/private.pem" \
-  --output "$TEST_ROOT/evidence.json" --json
-loadout alerts --evidence "$TEST_ROOT/evidence.json" \
-  --public-key "$TEST_ROOT/public.pem" --json
-```
-
-The harness scores declared observations only. It never executes candidate content.
-The authoritative schema, safety-failure, minimum-trial, tamper, and practical-delta
-tests are also directly runnable:
-
-```bash
-npx vitest run tests/head-to-head.test.ts tests/signing.test.ts
-```
-
-## 9. Credentials and model-provider verification (H/$)
+## 7. Credentials and model-provider verification (H/$)
 
 This track touches the real operating-system credential store even when
 `LOADOUT_HOME` is disposable. Use a unique throwaway service name and delete it.
@@ -668,7 +438,7 @@ unset OPENROUTER_API_KEY
 `models verify` makes one minimal request and may consume provider credit ($). Inspect
 provider billing before and after; do not run it in a loop.
 
-## 10. Watchers, native scheduling, completions, and loopback UI/API (X/H)
+## 8. Watchers, native scheduling, completions, and loopback UI/API (X/H)
 
 One-shot update watching is safe and networked:
 
@@ -707,33 +477,10 @@ and `unschedule` commands remain available for job-specific control.
 
 The optional read-only loopback API can be checked separately:
 
-```bash
-loadout serve --port 0
-```
-
 Confirm that it binds only to `127.0.0.1`, inspect the API response, and stop it with
 Ctrl-C. It must not bind a public interface.
 
-## 11. Improvement-cycle records (S)
-
-```bash
-loadout improve --json
-loadout improve --write --output "$TEST_PROJECT/improvements" --json
-```
-
-Copy the exact cycle id printed by the second command:
-
-```bash
-loadout improve-feedback --id <cycle-id> --outcome partial \
-  --note "Disposable matrix verification" \
-  --directory "$TEST_PROJECT/improvements"
-```
-
-Expected: the first command is read-only; `--write` persists a local prompt/cycle
-record only; feedback requires a human-selected outcome and stores no project source
-or prompt transcript.
-
-## 12. Cleanup and pass criteria
+## 9. Cleanup and pass criteria
 
 First verify snapshot availability and roll back any remaining disposable mutation:
 

package/docs/GITHUB_AUTHORIZATION.md

@@ -19,6 +19,11 @@ paste a token into a manifest, catalog, log, or command argument.
 
 ## Local flow and failure modes
 
+> **Partly unimplemented.** `loadout connect` and `loadout disconnect` do not
+> exist in the CLI today. `loadout discover --private` is real and reads
+> credentials the user has already configured. Steps 1, 2, 3, and 5 describe the
+> intended flow, not current behaviour.
+
 1. `loadout connect github` opens a browser using PKCE and a loopback callback.
 2. The local process verifies `state`, PKCE verifier, expiration, and callback host.
 3. It stores only an OS-keychain reference to the refresh/session material.

package/docs/PROVENANCE_AND_COMPARISON.md

@@ -26,7 +26,7 @@ their `SKILL.md` files, and writes a local index below `LOADOUT_HOME/provenance`
 
 ## Relationship classification
 
-`loadout compare` uses deterministic relationships:
+`loadout scan` uses deterministic relationships:
 
 - `exact-copy`: identical instruction fingerprint;
 - `divergent-same-name`: same normalized name, different instructions;

package/docs/RELEASE_REVIEW.md

@@ -19,7 +19,6 @@ For current behavior and evidence, use:
   catalog/support facts;
 - the [changelog](../CHANGELOG.md) for released behavior;
 - the [feature test matrix](./FEATURE_TEST_MATRIX.md) for adapter evidence;
-- the [repository stabilization record](./REPOSITORY_STABILIZATION.md) and
   [sanitized July 19 live checks](./evidence/live-checks-2026-07-19.json) only when
   investigating that historical release line.
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "loadout-ai",
-  "version": "0.7.0",
+  "version": "0.8.0",
   "private": false,
   "license": "MIT",
   "description": "Universal upgrade manager for AI coding agents",
@@ -15,7 +15,6 @@
     "docs",
     "README.md",
     "CHANGELOG.md",
-    "MASTER_PLAN.md",
     "SECURITY.md",
     "LICENSE"
   ],
@@ -57,6 +56,7 @@
     "test:e2e:readme": "node scripts/readme-product-flow.mjs",
     "test:package": "node scripts/package-smoke.mjs",
     "pretest:performance": "npm run build",
+    "test:coverage": "vitest run --coverage",
     "test:performance": "node scripts/scan-benchmark.mjs",
     "typecheck": "tsc -p tsconfig.json --noEmit",
     "readme:update": "node scripts/update-readme-facts.mjs",
@@ -65,8 +65,9 @@
     "check:live": "node scripts/check-live-evidence.mjs",
     "check:allowlists": "npm run build && node scripts/check-mode-allowlists.mjs",
     "check:catalog-freshness": "npm run build && node scripts/check-catalog-freshness.mjs",
-    "check:evidence": "node scripts/check-catalog-attribution.mjs && node scripts/check-discovery-artifacts.mjs && npm run check:readme-claims",
-    "verify": "npm run format:check && npm run lint && npm run typecheck && npm run check:evidence && npm test -- --run && npm run test:e2e:cli && npm run test:e2e:readme && npm run test:package && npm run test:performance",
+    "check:audit": "npm audit --audit-level=high",
+    "check:evidence": "node scripts/check-catalog-attribution.mjs && node scripts/check-discovery-artifacts.mjs && node scripts/check-documented-commands.mjs && npm run check:readme-claims",
+    "verify": "npm run format:check && npm run lint && npm run typecheck && npm run check:audit && npm run check:evidence && npm test -- --run && npm run test:e2e:cli && npm run test:e2e:readme && npm run test:package && npm run test:performance",
     "verify:full": "npm run verify"
   },
   "dependencies": {
@@ -76,6 +77,7 @@
   "devDependencies": {
     "@eslint/js": "^9.39.5",
     "@types/node": "^22.10.0",
+    "@vitest/coverage-v8": "^4.1.10",
     "eslint": "^9.39.5",
     "eslint-config-prettier": "^10.1.8",
     "prettier": "^3.9.5",

package/skills/loadout-router/SKILL.md

@@ -1,120 +1,85 @@
 ---
 name: loadout-router
-description: Choose the right model tier and agent for a coding task, and hand work off between Claude Code and Codex. Use when the user asks which model to use, mentions running low on usage or quota, wants to save tokens or cost, asks whether to switch to Opus/Sonnet/Haiku or a GPT tier, or wants to delegate a task to another agent.
+description: Decide which model to use for a coding task and hand work to another agent. Use when the user asks which model to use, mentions running low on usage or cost, asks whether to switch to Opus or Sonnet or a GPT tier, or wants to delegate a task to Codex or Claude Code.
 ---
 
 # Loadout Router
 
-Route each coding task to the cheapest model that still does it well, and hand
-work between agents when a different one is better suited.
+Pick the model that fits the work, using the user's own routing policy rather
+than your guess or mine.
 
-This skill wraps the `loadout` CLI, so the model catalog and pricing stay
-current with the installed version rather than going stale in this file.
-
-## Prerequisite
-
-Check once per session:
-
-```bash
-loadout --version
-```
-
-If that fails, tell the user to install it (`npm install --global loadout-ai`)
-and answer from general knowledge instead of guessing at specifics.
-
-## Choosing a model
-
-Run the router with the task described in plain words:
+## The policy is the user's, not yours
 
 ```bash
-loadout route <task description>
+loadout route
 ```
 
-It classifies the task into one of six phases — plan, implement, review, test,
-debug, document — and prints the recommended tier, the models in that tier with
-current prices, which of the user's installed agents can run them, and a cheaper
-fallback with its tradeoff.
+That prints three buckets and the model the user has chosen for each:
 
-Report the recommendation and the reason. Name the actual model, not just the
-tier. If the output lists a `Conserve:` alternative, mention it only when the
-user cares about cost or quota, otherwise it is noise.
+- **hard** — architecture, security, migrations, tricky debugging, risky review
+- **normal** — most implementation, ordinary debugging, refactors
+- **cheap** — tests, docs, boilerplate, renames, mechanical edits
 
-When the user is explicit about the phase, skip classification:
+Read the policy before advising. If the user disagrees with a recommendation,
+the fix is to change the policy, not to argue:
 
 ```bash
-loadout route --phase review
+loadout route --set cheap=claude-sonnet-5
 ```
 
-## When the user is low on usage
+## Your job is the bucket, not the model
 
-If the user mentions running out, being rate limited, conserving quota, or
-stretching a plan, add `--conserve`:
+The CLI can guess a bucket from wording, and it says so when it does. **You
+should do better**, because you have the conversation, the code, and the stakes.
+Decide the bucket yourself and state it:
 
 ```bash
-loadout route --conserve <task description>
+loadout route --bucket hard
 ```
 
-This drops each phase one tier and prints the tradeoff you are accepting. Say
-what is being given up — "shallower architectural reasoning, so review the plan
-more carefully" — rather than presenting it as a free win.
+Judge by consequence, not vocabulary:
 
-Neither Claude Code nor Codex exposes remaining quota programmatically, so never
-claim to know how much the user has left. `--conserve` is a user-driven choice,
-not a measurement.
-
-## Comparing models and cost
+- Anything touching auth, payments, migrations, or data deletion is **hard**,
+  however small the diff.
+- Unfamiliar code is harder than familiar code doing the same thing.
+- A one-line change in a hot path is not cheap.
+- Genuinely mechanical work — a rename, a docstring, a test for code you just
+  wrote — is **cheap**, and paying frontier prices for it is waste.
 
-```bash
-loadout route --models
-loadout route --models --provider anthropic
-loadout route --models --tier fast
-loadout route --cost
-```
+Report the model and why that bucket. One or two sentences.
 
-Use these when the user asks what is available or what something costs. Prices
-are per-million-token list rates; actual spend depends on prompt size, so give
-ratios ("roughly 4x cheaper") rather than predicting a dollar total.
+## When cost matters
 
-## Handing work to another agent
+If the user mentions running low, being rate limited, or wanting to spend less,
+say what a cheaper bucket would cost them in quality rather than presenting it
+as free. `loadout route` shows real per-million prices for the comparison.
 
-When a different agent suits the task better — or the user asks to delegate —
-use the handoff log. Check it is set up:
+Neither Claude Code nor Codex exposes remaining quota programmatically, so never
+claim to know how much the user has left.
 
-```bash
-loadout handoff status
-```
+## Handing work to the other agent
 
-If uninitialized, run `loadout handoff init` first. Then send the task:
+One command sends a task; it sets up the shared log on first use:
 
 ```bash
-loadout handoff send codex "write unit tests for the auth module" --context "see src/auth.ts"
+loadout handoff codex "write vitest coverage for src/auth.ts" --context "zod schemas already exist"
 ```
 
-Put anything the receiving agent needs into `--context`: file paths, the
-decision you already made, what you deliberately left out. The other agent has
-none of this conversation.
+Put everything the receiver needs into `--context` — file paths, decisions you
+already made, what you deliberately left out. It has none of this conversation.
 
-Sending a task is a real side effect on a shared file. Confirm with the user
-before sending unless they asked for the handoff themselves.
+Sending writes to a shared file in the user's repository, so confirm first
+unless they asked for the handoff themselves.
 
 ## Reading your own inbox
 
-At the start of a session, and after finishing a task, check whether another
-agent left you work:
+At session start, and after finishing a task:
 
 ```bash
-loadout handoff inbox claude-code
+loadout handoff claude-code
 ```
 
-If it lists tasks, work them in order and run the `loadout handoff done <id>`
-command it prints for each one. If it reports none, continue as normal and do
-not mention it.
-
-## Judgment this skill does not replace
+Work anything listed in order, then run the `loadout handoff --done <id>`
+command it prints. If nothing is pending, say nothing and carry on.
 
-- A "simple" task in an unfamiliar or high-risk area still deserves a stronger
-  model. The classifier reads keywords, not stakes.
-- Security-sensitive, migration, and data-loss paths are worth frontier tier
-  regardless of what phase they classify as.
-- If the user has already chosen a model, do not argue unless the choice is
-  clearly wrong for the work.
+`loadout handoff` with no arguments shows every pending task, both directions.