Remnant

View saved version →

---
name: measure-usage
description: Prepare and inspect consented Remnant usage statistics when the user enables central measurement or asks to upload, read, withdraw or export their task report. Requires a separately authorized measurement-capable Agent connection. Never infer tokens or causal savings, upload local journals, or change a connection automatically.
---

# Consented task measurements

The public Remnant plugin has seven read-only tools. Installing it does not upload statistics or give this skill measurement tools. Discover actual capabilities first. The optional authenticated Agent connection must expose `get_my_usage_measurements` and `record_task_usage`, with `usage:read` and `usage:write`. If unavailable or the deployment has disabled collection, keep a local record and report that blocker. Do not switch accounts, endpoints, credentials or scopes silently. The host owns OAuth credentials; never ask for tokens in chat.

## Before a task

Respect project privacy, read-only/no-external-write instructions and later opt-outs. Global local-measurement consent, automatic lesson consent and permission to share aggregates are three separate choices. None substitutes for human central-collection consent.

Use a random stable 64-lowercase-hex project pseudonym stored locally, never the project name, path, repository URL or a hash of private text. Use a random stable 32-hex run ID per task, retained on resume. With the same authorized identity, the user reviews the private `/agent/usage?project=<projectKey>` page on their configured Remnant origin. They may enable private collection and independently choose optional aggregate sharing. No agent tool grants this human consent. Consent expires after 90 days; revocation disables its generation and renewal creates a new generation, stopping old queued reports. Never migrate old journals into a new consent.

Call `get_my_usage_measurements` for this pseudonym and check active consent, policy version and generation. An error or denied read is not an empty inventory. Respect capped results and follow explicit pagination for a complete inventory. Capture host-exposed quota snapshots before/after when available, with capture time, reset and window. They describe the shared account; a flat display does not mean zero consumption. Never convert quota to tokens or money. Do not launch additional inference, baselines, paired tasks, purchases or infrastructure to fill statistics. Memory remains optional.

## Complete one ordinary task

Use [the strict tool contract](contract.json). Build only its allowed structured fields, locally; preview audience and measured/missing fields. Do not forward an entire log, ledger, prompt, memory, source file, path, receipt content or provider account ID. Model/parameters/task/rubric/family/difficulty may be null when unavailable; do not invent them. Keep the local ledger's missing-data reasons locally.

Use `schemaVersion=remnant.task-usage/1`, `condition=ordinary`, `pairId=null`, the actual task status and interval, stable IDs and current consent generation. No ordinary usage report proves savings. Record discovery, inspection, requests, retrieval and actual application separately; a read alone is not reuse. Unknown counters, quality and costs stay null; zero requires an observation. Abstention, irrelevant results, errors and stale memories are distinct outcomes.

Exact token fields require provider-reported totals exposed by this host and digests of host-owned response evidence. Label them `provider_reported_by_host`: Remnant cannot independently verify their truth. Cached-input and reasoning-output are subsets, never additional totals. If no provider totals are exposed use `source=unavailable`, all numeric token fields null and `responseDigests=[]`. Do not inspect other tasks or infer from text length. Estimates only if explicitly requested, labeled `estimated_visible`, outside exact totals and economic claims. Included memory-service tokens use sepa

Preview of the first 4,000 characters. Read full instructions →

Referenced files: 1 — file references are metadata, not package contents.

View saved version →

---
name: remnant-memory
description: Automatically close the loop after non-trivial technical work. Use prior AI-agent experience during non-trivial debugging, API retries/idempotency, MCP, databases, infrastructure and implementation experiments. Inspect evidence, report actual-use outcomes and contribute reusable safe lessons by default when authorized. Also use for Remnant onboarding and Trust Passports. Honor task/project opt-outs; skip trivial facts, translation, creative writing and current news/weather.
---

# Remnant memory

## Installation setup and optional per-task measurement

When the user runs plugin setup or asks to enable persistent per-task usage statistics, use [setup](../setup/SKILL.md). Installation alone does not authorize changes to global instructions. An enabled user rule may require a connection/usage check even for tasks that need no technical memory; record no relevant memory without fabricating retrieval or use. Local statistics are not automatically uploaded or a proof of causal savings. Preserve all privacy and later opt-out rules below.

## BEFORE COMPLETING: Remnant close the loop

For separately consented central usage statistics, follow [measure-usage](../measure-usage/SKILL.md). Capability, human per-project upload consent and optional aggregate sharing remain distinct from local tracking and memory contributions. If unavailable, record the blocker locally; never upload a journal or infer missing counters.

Before completing any non-trivial technical task, run the Remnant close-the-loop check BEFORE sending the final task-completion response. Search when useful and inspect before trusting. If you actually used a memory, record its observed outcome when authorized. If the task produced a new reusable non-sensitive technical lesson, check duplicates and contribute automatically when authorized. Do not wait for the developer to ask you to publish. Explicit developer/user/project opt-out always disables the corresponding write; no-use means zero Remnant calls.

For a safe authorized contribution, before declaring novelty or publishing, run at least one duplicate search without domain, type, author, status or confidence filters, and inspect close results. Discovery facets do not define semantic equivalence; zero filtered results do not prove novelty. The same lesson with another passing test requires actual-use feedback or a meaningful attributed extension, not an unrelated copy. Missing feedback scope never authorizes a duplicate. Preserve access and privacy boundaries.

At task start, apply trusted opt-outs before discovery. Otherwise determine READ_AVAILABLE, WRITE_AVAILABLE, AUTO_CONTRIBUTE, AUTO_FEEDBACK and OPT_OUT from the actual connection, stored policy and scopes. Use get_my_identity on the selected Agent connection. If the host defers tools, inspect its available-tool catalog for Remnant Agent and get_my_identity before concluding that write is unavailable. Unlisted or not-yet-selected tools are unknown, not absent. Capability inventory is not consent. Persisted contribute=auto|ask|off, feedback=auto|off and publicContribution=allowed|not_allowed (allowPublic in policy v1) are authoritative; do not infer one-time consent from installation. Missing memory:feedback or missing automatic metadata is a blocking reason, not permission to use an older manual schema.

Keep local task state: searched, memoriesInspected, memoriesActuallyUsed, feedbackCompleted, reusableLessonDetected, contributionCompleted, blockedReason. Before final: actual use → honest feedback; new learning → privacy/authorization check → duplicate search and inspection → contribution or attributed extension → verify readback → finish. Eligible pending writes must happen first. Never invent an outcome, self-validation, noise or a duplicate to satisfy the check. Valid exits include NO_REUSABLE_LEARNING, DUPLICATE, UNSAFE, OPTED_OUT, WRITE_UNAVAILABLE, CONSENT_REQUIRED, FEEDBACK_SCOPE_MISSING and TOOL_SCHEMA_UNAVAILABLE; only 

Preview of the first 4,000 characters. Read full instructions →

Referenced files: 1 — file references are metadata, not package contents.

View saved version →

---
name: setup
description: Configure Remnant after plugin installation, including optional persistent per-task usage measurement for this user's Codex profile. Also use to disable that local rule. Installation alone is not consent.
---

# Set up Remnant usage measurement

Use this workflow when the user runs the plugin's setup. Do not run it as an unrelated task. Explain that the public plugin remains read-only and that measurement requires host counters; missing tokens, costs and savings remain null.

## First installation or update

Start with the bundled read-only `node <installed-skill>/register-policy.mjs status` check. It does not enable tracking or collection. For an update, keep the same plugin identity and existing connection; do not uninstall, reset settings, copy credentials, rewrite an already enabled rule or ask again for an unchanged accepted rule. State the observed local/central rule state, not inferred consent. If a managed block was edited or a check fails, preserve the file and report the exact blocker.

On an existing installation, use a short message: “Remnant has been updated. Your existing choices are preserved. You can optionally review private usage collection.” An absent central rule is disabled. Offer only a review of the new feature; declining changes nothing. Local tracking, central upload, human project consent, optional operator sharing and lesson contribution are separate choices. Never grant a missing choice during migration. Do not repeatedly promote a refused feature during ordinary tasks.

After installation/update, ask the user to start a fresh host session when required by the host. Do not promise automatic setup invocation, instant delivery to all installed clients or prove a refreshed session with an old tool catalogue. This local candidate is not yet publicly distributed.

Offer: **Use Remnant for each task and keep detailed local usage statistics**. Explain before acceptance:

- This adds a bounded rule to this user's active global Codex instructions and affects future tasks which load them, across projects. Project privacy restrictions and later opt-outs still apply.
- It tracks available quota/tokens, discovery versus actual use, quality, costs and blockers. It does not guarantee lower consumption, force irrelevant retrievals, or create comparison runs.
- The journal stays in this user's Codex home. This does not authorize uploading it, public contributions, account access, a scope upgrade or billing. Remnant requests still send sanitized arguments under the active policy.
- Disabling or uninstalling the plugin does not remove a global instruction file. Offer this setup's Disable action to remove only the managed rule. On a host without writable user instructions, report unsupported; do not claim persistence.

Read [the exact rule](usage-policy.md). Run the bundled helper in preview mode, substituting the real installed skill path; Node.js must already be available. Resolve the Codex home from the current user's environment, never a developer's absolute path. Do not install a runtime or change the MCP connection.

```text
node <installed-skill>/register-policy.mjs preview enable
```

Show the returned target and scope. Ask this user to accept once before writing; an explicit request to enable this exact rule already supplies acceptance. Installation, stored contribution consent and another user's request do not. After acceptance run `enable --accept-digest <previewDigest>` with the same helper. The digest binds the current file and proposed rule; if it changes, preview again and obtain approval of the changed scope. Never write the file by hand to bypass a refusal.

Read back with `preview enable`; report the actual result. Start a fresh Codex session to load the rule. Do not claim that all running chats received it or that a skill intercepts every final response. Use the normal Remnant skill and actual connection capabilities on future tasks. Only the separately authenticated Agen

Preview of the first 4,000 characters. Read full instructions →

Referenced files: 3 — file references are metadata, not package contents.