AGENTS.md

1# Telegram Plugin Guide
2 
3Read this before any change under `extensions/telegram/`. These are intentional
4maintainer decisions and review-binding invariants, not incidental
5implementation details. Also read `extensions/AGENTS.md` for the plugin
6boundary rules.
7 
8Verified against Telegram Bot API 10.2, July 14 2026.
9 
10## Reliability Invariants
11 
12### Core drain contracts (do not re-implement in Telegram)
13 
14Owned by `src/channels/message/ingress-drain.ts` (+ claim-owner, retry-policy).
15Proof: `src/channels/message/ingress-drain.test.ts`,
16`ingress-claim-owner.test.ts`, `ingress-retry-policy.test.ts`.
17 
18- Completed rows tombstone via `complete()`, never `delete`.
19- Complete at turn adoption, not settle. Deferred holds the claim; watchdog
20  stays armed through deferral; dead-letter reason `handler-timeout`.
21- One retry policy: attempt floor **and** age gate (defaults 8 / 24h).
22- Claim refresh heartbeat while dispatching/deferred (`claimLeaseMs / 3`).
23- Never silently complete on transient failure — release/fail via disposition.
24- Pre-adoption supersede tombstones; post-adoption interruption is core-owned
25  (reply-run registry / queue interrupt).
26 
27### Telegram-owned (transport + channel policy)
28 
29- Durable-before-ack on both transports. Polling: ingress worker advances its
30  offset only after the parent's committed spool enqueue
31  (`writeTelegramSpooledUpdate`). Webhook: respond 200 only after the spool
32  write; non-200 on write failure is the redelivery contract.
33- `update_id`↔event-id encoding and per-chat/topic lane derivation stay in
34  `telegram-ingress-spool.ts` / sequential-key.
35- Polling and webhook both: enqueue then pump
36  `createTelegramTransportIngressDrain(...).drainOnce()` — no private claim
37  loops.
38- Stall timeout: `OPENCLAW_TELEGRAM_SPOOLED_HANDLER_TIMEOUT_MS` →
39  `adoptionStallTimeoutMs` (default 5 min) via
40  `resolveTelegramAdoptionStallTimeoutMs`.
41- Non-retryable classifier: `telegram-ingress-non-retryable.ts`
42  (missing harness, dispatch-dedupe rollback).
43- Supersede predicate: `telegram-ingress-supersede.ts` — only abort text /
44  authorized-looking explicit commands (and ambient room_event pending) may
45  supersede pre-adoption work. Normal messages never supersede.
46- room_event ambient work shares the sequential lane so a later user turn can
47  supersede it pre-adoption; adopted user turns are never touched (core drain
48  supersede is pre-adoption only).
49- No per-message full-store writes. Hot-path SQLite writes are per-entry.
50  Rewriting a cache on every send or read stalls the event loop, and that
51  stall masquerades as a polling stall (the sent-message-cache regression).
52- Transport error classification. The getUpdates worker retries Bot API 5xx
53  and 429 locally, honoring `parameters.retry_after`; 401/404 stay fatal; 409
54  must propagate to the parent session, which owns webhook-conflict recovery.
55  Bot API errors carry `error_code`, not `.code`; parse non-2xx bodies
56  defensively (a 502 HTML page is not JSON).
57- Send funnel parity. The durable funnel (`send.ts`) and the streaming funnel
58  (`bot/delivery.*`) must degrade identically: rich-entity 400 falls back to
59  plain text, caption parse 400 falls back to a plain caption, quote-not-found
60  400 falls back to a legacy reply. New recoveries go into the shared
61  predicates (`send-error-predicates.ts`, `reply-parameters.ts`), never into
62  one funnel only.
63- Outbound flood waits honor `retry_after` up to
64  `TELEGRAM_OUTBOUND_RETRY_AFTER_CAP_MS`; do not re-clamp Telegram sends to the
65  generic channel retry ceiling.
66- Webhook security ordering. The secret header is validated first
67  (constant-time compare, single-header enforcement, connection close on 401);
68  the request rate limit budgets only failed-auth attempts so Telegram's own
69  delivery is never throttled.
70- Every owned undici transport gets closed on all exit paths: polling session,
71  webhook shutdown and startup failure, probe-cache eviction.
72 
73## Streaming
74 
75- Do not reintroduce `sendMessageDraft` for answer streaming. Telegram drafts
76  are ephemeral 30-second previews in private chats; final delivery still
77  requires a separate `sendMessage`. OpenClaw uses `sendMessage` plus
78  `editMessageText`, then finalizes in place so the user sees one persistent
79  answer.
80- Streaming owns one visible preview message. Edit it forward. Do not send an
81  extra final bubble unless the final edit genuinely failed.
82- Keep the first-preview debounce. If a provider sends token-sized deltas,
83  coalesce them into cumulative preview text instead of removing the debounce.
84- Respect Telegram limits in the Telegram layer. Text over 4096 chars chains
85  into continuation messages. Polls keep the current Bot API 12-option cap.
86 
87## Telegram API Ownership
88 
89- Prefer grammY primitives and Telegram-native helpers when they model the
90  behavior directly. Avoid custom Bot API wrappers for behavior grammY already
91  owns.
92- Throttling is bot-token scoped. All Telegram API clients for the same token
93  share one grammY `apiThrottler()` instance.
94- Do not silently retry failed topic sends without topic metadata. A
95  wrong-surface success is worse than a loud Telegram error.
96- DM topics and forum topics are distinct. `direct_messages_topic_id` and
97  `message_thread_id` are not interchangeable.
98 
99## Context And Authorization
100 
101- Reply context comes from OpenClaw-observed messages. Bot API updates expose
102  `reply_to_message`, but there is no arbitrary `getMessage(chat, id)`
103  hydration path later.
104- Current local chat context must outrank stale reply ancestry in the prompt.
105  Old replied-to messages should not look like the active conversation.
106- The group history window is always on for groups and bounded by
107  `historyLimit`. Do not reintroduce prompt-history gating modes; that
108  regression blinded ambient rooms.
109- The group history window is rolling. Use self-entry watermark selection for
110  "since your last reply" views; do not reintroduce destructive clears because
111  room events are not persisted to the session and cleared context is
112  unrecoverable.
113- Pairing is DM-only. Group and topic authorization need explicit config
114  allowlists.
115- Telegram allowlists use numeric sender IDs. Usernames are optional, mutable,
116  and not a reliable arbitrary-user lookup key in the Bot API.
117- Group and channel visible replies are policy-controlled. Normal room replies
118  stay private unless `messages.groupChat.visibleReplies: "automatic"` is set
119  or the agent explicitly calls `message.send`.
120 
121## Interactive Surfaces
122 
123- Native callbacks stay structured. Approval, native command, plugin, select,
124  and multiselect callbacks must not fall through as raw callback text.
125- Preserve callback values exactly, including delimiters such as `env|prod`.
126- Native slash commands should remain fast-pathable before full workspace and
127  agent-turn setup.
128 
129## Review Standard
130 
131- Telegram behavior PRs need real Telegram proof when they touch transport,
132  streaming, topics, callbacks, authorization, or reply context. Prefer the
133  bot-to-bot QA lane or an equivalent live Telegram probe over synthetic-only
134  validation.
135- Reliability PRs (spool, drain, retry, ack, offset paths) need crash-window
136  or restart-replay test proof, not just happy-path tests.
137