Skip to main content
Glama
Still True

Still True

by Marc-Dvci

Still True re-checks every scheduled Buffer post against the world just before it goes out, and holds the posts the world has made wrong.

In the IT audits I worked on, a control counted only if it still held on the day it mattered, not the day it was approved. Still True applies that to a queue. Marc Donovici

Film (1:50) · Live app · Fire drill, no sign-in · For judges · How it works · Evidence · Run it

Live

still-true.chatisthisreal.workers.dev, any Buffer account connects with OAuth

Film

Still True in 1:50: a real queue held, tagged, explained and released inside Buffer

Built for

Buffer Commit 2026

Built with

Buffer GraphQL API and OAuth 2.0 with PKCE · Cloudflare Workers, Durable Objects (SQLite), Workers AI · TypeScript

Headline

410 labelled cases written before the code: every verdict matches the rules, no healthy post held or flagged. Five real days replayed with the alerts, news and weather recorded for them

For judges

Criterion

Where to look

Real usefulness

The problem in one paragraph; the five checks and the fire drill on real days; holds that never delete and are explained inside Buffer (src/durable/org.ts)

Execution and craft on the Buffer API

How it uses Buffer: 13 GraphQL operations, each validated against Buffer's introspected schema in CI; the live check on a real channel; PKCE with rotating refresh tokens owned by one Durable Object; request budget; the Worker tests that hold, note, tag, release and fix through a recorded Buffer stand-in

Creativity and originality

A queue that reads the world before it speaks: weather, remembrance days, disaster alerts and the linked page, checked at the moment of publishing. The fire drill replays your own queue on a real day the world changed

Related MCP server: agent-mandate-mcp

See it work

A product page that changed

A forecast that changed

A day of remembrance

The problem

A scheduled post is a promise written on Monday about Friday. Writing ahead is Buffer's whole value, and it has one failure: the world moves in between. The shirt sold out. The concert was postponed. "Tomorrow" now means another day. It is raining on the "terrace open" post. A wildfire reached the audience's city an hour ago. Today the protection is a manual, all-or-nothing pause button, in Buffer and in every other scheduler, and it only helps someone who is watching. Small businesses are not watching at 08:55, and agents now fill queues days ahead.

Still True is for the café owner, the shop, the venue and the agency running client queues. It reads each post again 60 and 10 minutes before it goes out.

How it works

Rules decide every verdict; a model drafts wording only when a person asks. Each finding names its evidence: what the post says, what the source says now, and when it was read.

Check

Example

Source of truth

Verdict

Calendar

"Join us Friday 14 November" when 14 November 2026 is a Saturday; "demain, dimanche 15" read on Friday 13; "tonight at 6pm" going out at 18:30; a sale "until 1 December" going out on 3 December; "good morning" at 20:30

The post's own words, read on the publish day in the channel's timezone

Review, or hold for a date already passed

Your own page

"Back in stock, €49" while the page says out of stock; a price the page no longer shows; an event cancelled, postponed or moved

schema.org JSON-LD (Shopify, WooCommerce, Wix, ticketing sites), Shopify's .js product endpoint, HTTP status

Hold, or review with the corrected price or date

Weather

"Terrace open this afternoon ☀️" into 90 % rain

Open-Meteo forecast for the due hour

Review, with the next dry, bright slot proposed

Remembrance

A flash sale in Paris on 13 November

31 national days in 17 countries, each with its source (data/remembrance.json)

Review

The world

"Our prices are on fire 🔥" the day a wildfire alert hits the audience's country

GDACS red and orange alerts; Wikipedia's Current Events

Hold for wording that echoes the hazard, review for promotions

The full rule for each of the 18 codes is on the How it works page. Supportive posts ("our thoughts are with", "collecting donations", "we will be closed") are never flagged by the disaster and news checks.

What happens to a flagged post

  1. Held, never deleted. A certain problem (sold out, cancelled, a date already passed, wording that echoes an active disaster) moves the post to drafts with editPost(saveToDraft: true). Its original time is kept.

  2. Explained inside Buffer. A note on the post (createCollaborationNote) and a tag, "Still True: held" or "Still True: review", say what changed and where it was read. The team sees it in Buffer without opening Still True.

  3. One tap to fix or release. The corrected price, the right weekday or the next dry slot is applied with editPost. Release puts the post back at its original time. A setting turns holds off entirely: notes only.

The fire drill

/drill replays a queue on a real day the world changed, with the records published for that day: GDACS alerts, Wikipedia's Current Events and Open-Meteo's archived weather. Signed-in accounts replay their own queue at the same times of day.

Day

Place

Held

Review

Still true

8 Jan 2025, Palisades and Eaton fires

Los Angeles

2

3

3

29 Oct 2024, floods

Valencia

3

3

2

6 Feb 2023, earthquake

Istanbul

2

4

2

15 Jun 2025, storm

Strasbourg

1

3

4

13 Nov 2026, anniversary of the 2015 attacks

Paris

1

3

4

For agents

POST /mcp is an MCP server (Streamable HTTP) with one read-only tool, check_post. An agent that schedules through Buffer's own MCP server can ask Still True first.

{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"check_post","arguments":{"text":"Join us Friday 14 November!","due_at":"2026-11-12T18:00:00+01:00","timezone":"Europe/Paris"}}}

How it uses Buffer

Need

Buffer API

Code

Connect any Buffer account

OAuth 2.0 Authorization Code with PKCE (S256), offline_access; single-use refresh tokens refreshed by exactly one owner, the account's Durable Object

src/buffer/oauth.ts

Read the queue

posts filtered by status (scheduled, timed draft) and a seven-day dueAt window, sorted by dueAt

src/buffer/operations.ts

Local day and hour

channels: timezone, service, isDisconnected; account.timezone for the default country

src/durable/org.ts

Hold, release, fix

editPost with saveToDraft, mode: customScheduled + dueAt, text, tagIds

src/durable/org.ts

Explain inside Buffer

createCollaborationNote, tagsV2, createTag

src/durable/org.ts

Sample drafts

createPost with saveToDraft: true

src/demo/samples.ts

Every operation is validated against schema/buffer.graphql, the schema introspected from api.buffer.com, in CI. Typed mutation errors (... on MutationError), 401s, RATE_LIMIT_EXCEEDED with Retry-After, and the RateLimit headers are all handled. Details and the request budget: docs/BUFFER_API.md.

What it will not do

  • It never deletes a post and never publishes one. A hold is a move to drafts.

  • No model decides anything. Workers AI is called only when a person presses "Suggest wording", and a suggestion is refused if it adds a number, date, link, handle or hashtag the post did not contain (src/ai/assist.ts).

  • It stores only what a check needs, for posts in the next seven days, and deletes everything on disconnect (privacy, system card).

Evidence

Measure

Result

Reproduce

Labelled cases, written and committed before the detectors

267 development cases, 143 held-out cases (191 healthy in total): every verdict and every finding code matches the written rules; 0 healthy posts held or flagged

npm run eval and npm run eval -- heldout · docs/EVALUATION.md

Live Buffer API

Through the live app and its OAuth App Client (11 Nov 2026): connected a real account; the first check, 7 Buffer requests, moved a scheduled post linking an out-of-stock product to drafts with the "Still True: held" tag and a note; "Release as written" on the live board put it back at its exact original time without the tag. With the local Worker on the same channel: the weekday fix and the five sample drafts; Buffer's responses recorded as fixtures

BUFFER_API_KEY=... npx tsx scripts/live_check.ts <channel> · docs/LIVE.md

Real markup

Prices, stock and event status read correctly from Shopify, WooCommerce, Wix and ticketing-site markup, including a malformed JSON-LD block and an event postponed only in its title

test/unit/page.test.ts

Real days

The five drills above, on the recorded GDACS, Wikipedia and Open-Meteo data

test/unit/drill.test.ts

Buffer contract

13 operations valid against the introspected schema

test/unit/operations.test.ts

Tests

171 unit · 48 Worker and Durable Object tests in workerd · 12 browser tests with axe-core (WCAG 2.2 AA, desktop and mobile). Decision core coverage: 99.4 % of lines, 85.5 % of branches

npm test, npm run test:unit, npm run test:e2e

IMPORTANT

What is real: the disaster alerts, news items and weather in the fire drill are the records GDACS, Wikipedia and Open-Meteo published for those days, and the page fixtures reproduce the markup of real Shopify, WooCommerce, Wix and ticketing pages. Atelier Nord and Salle Rivage are the demo shop and venue Still True serves itself, so the sample drafts have pages to link to.

Claims and the tests that hold them

Claim

Test

A certain problem moves the post to drafts, adds a note and the held tag, and leaves healthy posts untouched

test/worker/org.test.ts "holds an out-of-stock post, explains it in Buffer, and leaves a healthy post alone"; live on Buffer in docs/LIVE.md, recorded in test/unit/buffer-recorded.test.ts

Every edit sends the post's text, which Buffer requires

test/worker/org.test.ts "sends the post's text with every edit, as Buffer requires"

The note never claims a hold that Buffer refused

test/worker/org.test.ts "says 'flagged', not 'moved to drafts', when Buffer refuses the hold"

A post is held and noted once, not on every check

test/worker/org.test.ts "does not hold again a post it already explained"

Two checks at once (the first check after connecting, then "Add sample drafts") hold and note a post once

test/worker/org.test.ts "runs one check at a time: two at once hold and note a post once"

A held post deleted in Buffer leaves the board; one still in drafts stays

test/worker/org.test.ts "forgets a held post once it is deleted in Buffer, and keeps one that is still a draft"

A channel connected in Buffer is listed at the next check

test/worker/org.test.ts "lists a channel connected in Buffer at the next check, not an hour later"

Buffer's RateLimit header is read per window, keeping the tightest

test/worker/primitives.test.ts "reads Buffer's RateLimit headers and keeps the window with the fewest requests left"

Notes-only mode never changes a post

test/worker/org.test.ts "only writes notes in 'notes' mode, and never holds"

Release restores the original time and removes the held tag

test/worker/org.test.ts "releases a held post at its original time and drops the held tag"

An expired token is refreshed once and Buffer's replacement kept

test/worker/org.test.ts "refreshes an expired token once and keeps Buffer's replacement"

The action log is hash-chained and detects an edited entry

test/worker/org.test.ts "chains the action log by hash and detects an edited entry"

Disconnect revokes the token and deletes everything

test/worker/org.test.ts "revokes the token and deletes everything on disconnect"

Tokens are bound to their account by AES-GCM associated data

test/worker/primitives.test.ts "round-trips, and fails for another owner or a tampered value"

PKCE matches RFC 7636

test/worker/primitives.test.ts "computes the PKCE challenge of RFC 7636's test vector"

Supportive posts are never flagged during a disaster

test/unit/world.test.ts "never flags supportive posts"

A suggestion that adds a claim is refused

test/unit/assist.test.ts "refuses a suggestion that adds a claim"

Dates and weekdays inside links are never read as words

test/unit/calendar.test.ts "ignores weekdays and dates inside URLs"

Run it

TIP

No install, no key: open thefire drill. With Docker: docker compose up, then http://127.0.0.1:8788.

Open in GitHub Codespaces

Node 22 or later.

npm ci
npm test                      # unit tests and Worker tests in workerd, random order
npm run eval                  # the labelled cases
npm run test:e2e              # browser tests with axe-core
node scripts/dev_secrets.mjs && npx vite build && npx wrangler dev --env e2e --port 8788

To connect a real Buffer account locally, create an App Client in Buffer (Settings, API, App Clients) with the redirect URI https://localhost:8787/auth/callback, copy .dev.vars.example to .dev.vars, fill it in, and run npm run dev.

Layout

src/core/        the decision core: pure functions, no network, no clock
src/buffer/      GraphQL operations, API client, OAuth with PKCE
src/durable/     OrgDO (one per account) and WorldDO (shared)
src/fetchers/    linked pages and Open-Meteo, guarded and cached
src/demo/        the fire drill, the demo shop, the sample drafts
src/mcp.ts       the MCP endpoint
web/             the departure board (Vite, vanilla TypeScript)
data/            remembrance days, drill scenarios, place and timezone tables
eval/            labelled cases and the scorer
schema/          Buffer's introspected GraphQL schema

Licence

MIT. Data sources and their licences are listed on the How it works page: GDACS, Wikipedia (CC BY-SA 4.0), Open-Meteo (CC BY 4.0). Still True is built on the Buffer API and is not an official Buffer product.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      A
      quality
      C
      maintenance
      A pre-action risk gate for AI agents. Your agent calls the forecast tool before any irreversible action — send email, run SQL, make a payment, delete a file — and gets a risk score (0–100) and a GO / CONFIRM / STOP verdict in a few seconds.
      1
      73 npm
      MIT
    • A
      license
      A
      quality
      B
      maintenance
      Verifies agent actions against mandates via a single tool, returning allow, deny, or requires_approval receipts without executing the actions.
      1
      49 npm
      MIT
    • A
      license
      Not graded
      quality
      C
      maintenance
      Provides agents with five read-only advisory tools for bounded semantic decisions, returning redacted receipts while leaving execution and authorization to the host.
      17,996 npm
      Apache 2.0
    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables AI assistants to check Trustpilot and Google review text against the platforms' published rules using five read-only tools.
      10 npm
      MIT