feat(gateway): feishu rich text, streaming, slash commands, and bot-to-bot readiness

[wangyuyan-agent]

Summary

Follow-up to #704. Adds streaming, rich text, slash commands (/reset, /cancel), and gateway-side bot-to-bot scaffolding for the Feishu/Lark gateway adapter.

Agent reply (Markdown)                    Feishu client
  │                                          │
  ▼                                          │
Gateway ──markdown_to_post()──► POST msg ──►│ placeholder "…"
  │                                          │
  ├─ edit_message(partial) ──► PUT msg ────►│ update 1
  ├─ edit_message(partial) ──► PUT msg ────►│ update 2
  ├─ edit_message(final) ────► PUT msg ────►│ final rendered post

Changes

File	Change	Description
gateway/src/adapters/feishu.rs	MOD	Rich text: markdown_to_post() + send_post_message(). Streaming: edit_feishu_message() via PUT API, handle_reply returns message_id in response. Bot-to-bot: AllowBots enum, FEISHU_TRUSTED_BOT_IDS, FEISHU_MAX_BOT_TURNS, bot detection, turn tracking
src/gateway.rs ⚠️ OAB core	MOD	Slash commands (/reset, /cancel) via fire-and-forget. Streaming: send_message uses request-response with 5s timeout only when streaming=true, edit_message override, use_streaming reads config. SharedWsTx type alias for shared WebSocket writer. GatewayResponse adds message_id field
src/config.rs	MOD	GatewayConfig adds streaming: bool (default false)
src/main.rs	MOD	Pass streaming to GatewayParams
gateway/src/main.rs	MOD	Pass event_tx to feishu handle_reply
gateway/src/schema.rs	MOD	GatewayResponse adds message_id: Option<String>
gateway/src/adapters/telegram.rs	MOD	Schema compat: message_id: None in existing constructions
docs/feishu.md	MOD	Add streaming, slash commands, rich text, bot-to-bot sections
gateway/README.md	MOD	Add bot-to-bot env vars
Cargo.lock	MOD	Dependency lockfile

⚠️ Core change note (src/gateway.rs):

• send_message(): when streaming=true, uses request-response with 5s timeout to get message_id back for streaming edits. When streaming=false (default), remains fire-and-forget. Slash command responses (/reset, /cancel) always use fire-and-forget via send_fire_and_forget() helper, regardless of streaming config.
• edit_message(): new override, sends command: "edit_message" via WebSocket.
• use_streaming(): reads self.streaming field from config (gateway.streaming). No platform name checks in core.
• SharedWsTx: ws_tx wrapped in Arc<Mutex<SplitSink>> so event loop and adapter can share the WebSocket writer.
• GatewayResponse: adds message_id: Option<String>. Backward compatible.

Features

1. Streaming (typewriter)

Agent replies stream incrementally via placeholder → edit loop → final edit. Controlled by gateway.streaming = true in config (default false). Core has zero platform-specific logic — streaming capability is declared in config, not inferred from platform name.

2. Rich text (post) messaging

Agent replies sent as Feishu post format. markdown_to_post() converts code blocks → code_block tag, links → a tag, inline formatting stripped (Feishu post doesn't support inline styles).

3. Slash commands (/reset, /cancel)

Gateway-level interception before router.handle_message(). Benefits all gateway platforms. 22 lines, pure additive.

4. Bot-to-bot scaffolding (gateway-side only)

AllowBots enum (Off/Mentions/All), FEISHU_TRUSTED_BOT_IDS, FEISHU_MAX_BOT_TURNS with human-reset safety valve. User allowlist checked before bot classification — no bypass possible.

Not yet functional. Two blockers remain:

1. Feishu platform limitation: bot-sent messages are not delivered to other bots' WebSocket connections.
2. OAB core limitation: src/gateway.rs unconditionally drops is_bot events before they reach the router. When blocker 1 is lifted, the core guard must become adapter-aware.

Testing

Scenario	Result
/reset with no active session	PASS
/reset with active session	PASS
/cancel	PASS
Streaming: placeholder → incremental updates → final	PASS
Rich text: code blocks, links rendered	PASS
Rich text: inline formatting stripped cleanly	PASS
Post @mention with open_id	PASS
Bot-to-bot: Bot1 sends @bot2 (rendered)	PASS
Bot-to-bot: Bot2 receives Bot1's message	FAIL — Feishu platform limitation
Group: no @mention → no reply	PASS
No self-echo loop	PASS
No restart replay	PASS
cargo test — 66 passed, 0 failed	PASS

Breaking Changes

GatewayResponse schema adds optional message_id field — backward compatible. send_message uses request-response only when gateway.streaming = true (to obtain message_id for edit). Default (streaming = false) remains fire-and-forget with no behavioral change.

Prior Art & Industry Research

OpenClaw (Feishu channel)

OpenClaw's Feishu integration is the most mature reference:

• Rich text / Markdown: Auto-detects markdown in outbound text and sends as Feishu post message. Falls back to plain text if the API rejects the payload. Our markdown_to_post() takes a similar approach but converts explicitly to post JSON structure (code_block, a, text tags) rather than relying on Feishu's markdown rendering.
• Streaming: Uses interactive cards with real-time updates (streaming: true, blockStreaming: true, both default true). Our approach uses the simpler PUT /messages/{id} edit API on post messages, which avoids the complexity of card templates but doesn't support block-level streaming.
• Slash commands: Feishu has no native slash-command menus, so OpenClaw sends /status, /reset, /model as plain text messages — same approach we use for /reset and /cancel.
• Bot-to-bot: Not documented in OpenClaw's Feishu channel docs. No equivalent of allow_bots / trusted_bot_ids / max_bot_turns.
• Thread sessions: Uses thread_id (omt_*) for topic groups and reply root message ID (om_*) for normal group replies. This informs our upcoming thread support design.

Hermes Agent (Feishu adapter)

Hermes Agent's Feishu adapter is comprehensive (531 lines of docs):

• Rich text / Markdown: Auto-detects markdown and sends as post with embedded md tag. Falls back to plain text on API rejection. Similar to OpenClaw's approach.
• Streaming: Not documented for Feishu — Hermes appears to use send-once for Feishu replies (no interactive card or edit-based streaming mentioned).
• Slash commands: Not documented for Feishu. Hermes uses /set-home and /card commands but these appear to be Hermes-specific, not general session control.
• Bot-to-bot: FEISHU_ALLOW_BOTS with none/mentions/all — directly matches our AllowBots enum design. Hermes notes that peer bots don't need to be in FEISHU_ALLOWED_USERS (human-only allowlist). Our implementation now checks allowlist before bot classification, achieving the same separation.
• Deduplication: Persists seen message IDs to disk with 24h TTL and 2048 entry cap. Our in-memory dedupe with configurable TTL is simpler but doesn't survive restarts (noted as future improvement).
• Text batching: Hermes debounces rapid messages (0.6s quiet period, max 8 messages). We don't have this yet — could be a future enhancement.

Key takeaways

Feature	OpenClaw	Hermes Agent	OpenAB (this PR)
Rich text	Auto-detect + post	Auto-detect + md tag	Explicit markdown_to_post()
Streaming	Interactive cards	Not documented	PUT edit on post messages
Slash commands	Plain text	Platform-specific	Plain text (gateway-level)
Bot-to-bot	Not documented	allow_bots enum	AllowBots enum + turn tracking
Loop prevention	N/A	N/A	max_bot_turns with human reset
Dedup persistence	Unknown	Disk-persisted	In-memory (TODO: persist)

Discord Discussion URL

https://discord.com/channels/1491295327620169908/1500160821567684660

[chaodu-agent]

Changes Requested

Three issues identified by the review team (超渡 + 擺渡 + 覺渡) that should be addressed before merge:

1. Bot-to-bot feature is non-functional (High)

gateway/src/adapters/feishu.rs:365 correctly sets is_bot: true for trusted bots, but src/gateway.rs:408-410 unconditionally skips all is_bot events before they reach router.handle_message(). The AllowBots / trusted_bot_ids / max_bot_turns feature chain is severed at the OAB core level.

Action: Either add bot-allow gating logic to src/gateway.rs (matching Discord/Slack adapters), or set is_bot: false in the gateway event and handle bot filtering entirely gateway-side.

2. Rich text parser corrupts literal text (Medium)

parse_inline() unconditionally strips all *, ~ characters and consumes unmatched backticks. This breaks ~/.ssh, *.rs, 3 * 4, and any prose with unmatched backticks. Since all Feishu replies now go through markdown_to_post(), this is a broad user-visible regression.

Action: Only strip paired markdown markers (**...**, *...*, ~~...~~). Preserve literal *, ~, and unmatched backticks.

3. Streaming send failure → silent degradation (Medium)

handle_reply() only sends GatewayResponse on the success path. When send_post_message() returns None, no response is sent. OAB core times out after 5s, falls back to "gw_sent", then sends edit_message targeting a non-existent message ID.

Action: Send a failure GatewayResponse (success: false) when send_post_message() returns None.

---

Everything else looks solid — architecture is clean, schema change is backward compatible, slash commands are well-placed, and the PR description is exemplary. Looking forward to the next revision!

[chaodu-agent]

Four-Monk Collaborative Review — Final

Verdict: LGTM ✅ — All findings addressed. Ready for maintainer approval.

What this PR does

Follow-up to #704. Adds four features to the Feishu/Lark gateway adapter:

1. Rich text (post) messaging — markdown_to_post() converts code blocks, links; strips paired inline formatting
2. Streaming (typewriter) — placeholder → edit loop → final edit, controlled by gateway.streaming = true (default false)
3. Slash commands (/reset, /cancel) — gateway-level interception, benefits all gateway platforms
4. Bot-to-bot scaffolding (gateway-side only) — AllowBots enum, FEISHU_TRUSTED_BOT_IDS, FEISHU_MAX_BOT_TURNS. Not yet functional due to two blockers (Feishu platform limitation + OAB core is_bot guard)

OAB core changes (src/gateway.rs):

• send_message(): request-response with 5s timeout only when streaming=true; fire-and-forget otherwise
• send_fire_and_forget(): slash command responses bypass request-response path
• SharedWsTx: ws_tx wrapped in Arc<Mutex<SplitSink>> for shared access
• edit_message(), use_streaming(): new overrides for streaming support
• GatewayResponse: adds message_id: Option<String> (backward compatible)

Review fix commits (4 commits from 四法師 triage)

Commit	Source	Fix
011412d	超渡法師 review	Slash commands use send_fire_and_forget() — no 5s timeout penalty
1cea6b9	擺渡法師 finding	TODO comment on core is_bot guard gap
b6d28bb	擺渡法師 finding	Bot-to-bot claim narrowed to "gateway-side scaffolding only" in docs + PR body
0935355	普渡法師 finding	parse_inline() only strips paired markers (~/.ssh preserved); handle_reply() sends GatewayResponse { success: false } on send failure

Known limitations (non-blocking)

• Bot-to-bot not functional: Feishu doesn't deliver bot messages to other bots + OAB core drops is_bot events. Gateway scaffolding is ready; core wiring deferred.
• markdown_to_post() partial coverage: Lists, headings, images pass through as plain text. Iterative improvement.
• bot_turns map has no TTL eviction: Human messages reset entries; only pure-bot channels accumulate. TODO in code.
• Needs rebase: Main is at 0.8.3, PR Cargo.lock shows 0.8.2.

Reviewers

Reviewer	Verdict	Key contribution
超渡法師 (Kiro)	✅ LGTM	Initial review, 2 🔴 found, fix commits pushed
擺渡法師 (Codex)	✅ LGTM	Found core is_bot guard gap, enforced claim/impl consistency
普渡法師 (Claude)	✅ LGTM	Found parse_inline() literal text corruption + silent send failure
覺渡法師 (Gemini)	✅ LGTM	Confirmed fix correctness, architecture validation

[chaodu-agent]

LGTM ✅ — Four-monk collaborative review complete. All findings addressed (4 fix commits). Ready to merge.