Best possible solution:

Keep the Telegram-side accumulator, but complete the contract by preserving or recomputing partial delta/replace metadata through agent-runner-execution and add focused coverage for the full callback path before merge.

Do we have a high-confidence way to reproduce the issue?

Yes, source-level reproduction is high confidence: current main keeps the 30-character first-preview debounce and the draft stream suppresses first previews below that threshold. A token-shaped partial forwarded as text-only can therefore remain invisible until final delivery; I did not run a live Telegram bot in this read-only review.

Is this the best way to solve the issue?

No. The Telegram accumulator is a maintainable shape, but the PR must also carry or recompute delta/replace through the normal agent-runner callback wrappers so production Telegram sees the same payload shape as the new tests.

Full review comments:

• [P2] Forward partial deltas through the runner — src/auto-reply/get-reply-options.types.ts:80
  The new Telegram accumulator only works when onPartialReply receives delta/replace, but the normal runner wrappers still call params.opts.onPartialReply({ text }) or { text, mediaUrls } after typing normalization, dropping those fields in agent-runner-execution.ts. Token-sized fragments can therefore still reach Telegram as text-only snapshots and stay below the first-preview debounce in the production path.
  Confidence: 0.87

Overall correctness: patch is incorrect
Overall confidence: 0.86

Acceptance criteria:

• pnpm test src/auto-reply/reply/agent-runner-execution.test.ts extensions/telegram/src/bot-message-dispatch.test.ts extensions/telegram/src/draft-stream.test.ts src/plugin-sdk/channel-streaming.test.ts
• pnpm exec oxfmt --check --threads=1 src/auto-reply/reply/agent-runner-execution.ts src/auto-reply/reply/agent-runner-execution.test.ts extensions/telegram/src/bot-message-dispatch.ts extensions/telegram/src/bot-message-dispatch.test.ts
• git diff --check upstream/main

What I checked:

• Current-main Telegram debounce path: Current main still passes minInitialChars: 30 for Telegram partial/block draft streams, so a first preview is intentionally withheld until the rendered preview reaches the threshold. (extensions/telegram/src/bot-message-dispatch.ts:527, cb60a06c9e58)
• Current-main draft suppression: The draft stream returns without sending while no preview message exists and rendered text is shorter than minInitialChars, matching the reported no-preview symptom for token-sized fragments. (extensions/telegram/src/draft-stream.ts:284, cb60a06c9e58)
• Delta fields are produced upstream: The embedded subscription layer builds assistant stream data with text, delta, and replace, then calls onPartialReply(data), so the missing contract is downstream forwarding rather than upstream derivation. (src/agents/pi-embedded-subscribe.handlers.messages.ts:605, cb60a06c9e58)
• Delta fields are dropped before channel callbacks: The normal agent-runner wrappers call params.opts.onPartialReply with only { text } or { text, mediaUrls }, dropping delta and replace before Telegram can use the PR's accumulator. (src/auto-reply/reply/agent-runner-execution.ts:1474, cb60a06c9e58)
• PR-head Telegram accumulator: PR head 0cab701b2afb adds DraftPartialTextUpdate, resolves updates using delta/replace, and passes the whole partial payload into Telegram ingestion, but this only works when those fields survive to the channel callback. (extensions/telegram/src/bot-message-dispatch.ts:113, 0cab701b2afb)
• PR discussion and proof: The PR body and comments include terminal proof from a local Telegram dispatcher/draft-stream smoke and contributor validation, and the related issue remains the active bug target while earlier debounce-removal attempts were closed. (0cab701b2afb)

Likely related people:

• obviyus: Recent commits on Telegram draft streaming and draft-stream behavior include 10bbed8a6d30, 30e079dd89b4, and 7a2cc4b8d65c; this PR is also assigned to this handle and its latest commit is by this handle. (role: recent Telegram streaming contributor; confidence: high; commits: 10bbed8a6d30, 30e079dd89b4, 7a2cc4b8d65c; files: extensions/telegram/src/bot-message-dispatch.ts, extensions/telegram/src/draft-stream.ts, extensions/telegram/src/bot-message-dispatch.test.ts)
• steipete: Recent agent-runner-execution.ts history includes CLI preview bridge serialization and related bridge work, which is the path currently dropping the new partial metadata before channel callbacks. (role: recent agent preview bridge contributor; confidence: high; commits: 560c7440fb88, 2afffccee3af, 434b3d387608; files: src/auto-reply/reply/agent-runner-execution.ts, src/auto-reply/reply/agent-runner-execution.test.ts)
• medns: Recent agent streaming/refactor history touches the embedded subscription and stream-data area that produces text/delta payloads consumed by this PR's intended contract. (role: recent adjacent agent streaming contributor; confidence: medium; commits: e7277b4e3a4b; files: src/agents/pi-embedded-subscribe.handlers.messages.ts, extensions/telegram/src/bot-message-dispatch.ts)

Remaining risk / open question:

• The supplied terminal smoke proves the Telegram dispatcher/draft-stream behavior when delta metadata reaches it, but it does not cover the full agent-runner-to-Telegram callback path where the metadata is currently lost.

Codex review notes: model gpt-5.5, reasoning high; reviewed against cb60a06c9e58.