| name | riela-chat-example-verification |
| description | Verify Riela chat examples and live chat memory regressions through RielaApp, especially Telegram, Discord, Matrix, persona trio chats, image attachment prompts, raw-log/daily-summary memory examples, event-source receipts, and chat reply visibility checks. |
Riela Chat Example Verification
Use this skill when asked to prove that Riela chat examples actually work, not just that messages were posted. Treat "sent in chat" as preparation only; verification requires a running event listener and observable processing evidence.
Required Evidence
Collect at least one of these before calling a chat example verified:
- A visible bot reply in the target chat UI.
- A new event receipt from
riela events list for the matching source.
- A completed workflow session created after the test message, with expected memory read/write or reply output.
For RielaApp verification, also confirm the app process launched and whether its daemon riela events serve child stayed running or exited. Preserve command logs under tmp/<task-name>/; do not write scratch files at the repository root.
Preflight
Run these checks from the repo root:
swift build --product RielaApp
riela events validate --workflow-definition-dir ./examples --event-root ./examples/event-sources/.riela-events --output json
ps aux | rg 'RielaApp|riela events serve' | rg -v 'rg|Brave Helper|Discord Helper' || true
Check environment readiness by listing variable names only, never values:
kinko status || true
direnv exec . sh -lc 'env | cut -d= -f1 | rg "^(RIELA_TELEGRAM|RIELA_DISCORD|DISCORD_BOT|RIELA_MATRIX|GEMINI|OPENAI)"' || true
find "$HOME/.riela" -maxdepth 2 -type f \( -name 'rielaapp.env' -o -name 'env' \) -print 2>/dev/null
RielaApp-launched daemon children read ~/.riela/rielaapp.env, ~/.riela/env, and the app process environment. Shell-only direnv state is not enough for GUI-launched RielaApp.
RielaApp Launch Check
Start RielaApp and inspect the daemon result:
swift run RielaApp
Expected log shape:
profile=... discovered ... user daemon workflow candidate(s)launch ... riela events serve ...- either
pid=... running after startup grace or a precise exit message such as a missing token/env var.
If using Computer Use, inspect the RielaApp status-bar app or "Riela Workflows" window. If accessibility cannot inspect the accessory app, record the timeout and use process/log evidence instead.
Chat Regression Flow
Use this sequence for Telegram, Discord, and Matrix:
- Start RielaApp or
riela events serve for the exact event root and workflow set under test.
- Confirm a matching listener process is running.
- Send a text-only sanity message to the primary bot. The message should invite a normal conversation, not only a ping.
- Confirm the primary bot's visible response includes mentions or handoff text that targets the other two bots/personas.
- Confirm bot1 and bot2 each produce meaningful visible replies, or collect event receipts/completed sessions that prove their turns ran.
- After that conversation, verify chat memory by asking the primary bot, bot1, and bot2 to summarize what was just discussed. Each summary should reference the actual conversation content, not a generic "I do not know" answer.
- Send the image test to Mika with the Yui image asset and ask what is visible in the image.
- Confirm the image answer describes image content. Do not accept replies that only say a photo, square image, or attachment exists.
- Ask about the same image again without reattaching it to verify image memory reuse.
- Confirm the follow-up reply describes the previously stored image content, or record the exact blocker.
Do not mark image behavior verified when imagePaths is empty unless the workflow intentionally reasons from attachment metadata only.
Required Live Chat Scenario
When the user asks for end-to-end chat memory regression, run this scenario on
each requested surface rather than only sending isolated prompts:
- Talk to the primary bot with an arbitrary but concrete conversation topic.
The bot's reply must mention or hand off to bot1 and bot2.
- Verify bot1 and bot2 answer meaningfully in the same chat flow.
- Ask the primary bot, bot1, and bot2 to summarize the conversation. Treat this
as the text-memory check.
- Show an image and ask what is in it.
- Ask again later what was in the image without reattaching it. Treat this as
the file/image-memory check.
Acceptable proof is visible browser/chat UI evidence, fresh event receipts, or
completed workflow sessions with memory read/write details. A sent message alone
is never enough.
Shared Persona NodeRef And Handoff Guard Regression
Use this when verifying the shared Yui/Mika/Rina persona nodes, cross-workflow
nodeRef, or the deterministic handoff loop guard. This is the regression that
prevents prompt-only handoff flags from creating Yui/Mika/Rina loops.
First validate the shared workflow and each vendor workflow:
direnv exec . .build/arm64-apple-macosx/debug/riela workflow validate \
shared-agent-trio-personas --workflow-definition-dir ./examples
direnv exec . .build/arm64-apple-macosx/debug/riela workflow validate \
telegram-agent-trio-chat --workflow-definition-dir ./examples
direnv exec . .build/arm64-apple-macosx/debug/riela workflow validate \
discord-agent-trio-chat --workflow-definition-dir ./examples
direnv exec . .build/arm64-apple-macosx/debug/riela workflow validate \
matrix-agent-trio-chat --workflow-definition-dir ./examples
Run deterministic mock checks for all three chat vendors:
mkdir -p tmp/skill-chat-verification/logs tmp/skill-chat-verification/sessions
for workflow in telegram-agent-trio-chat discord-agent-trio-chat matrix-agent-trio-chat; do
direnv exec . .build/arm64-apple-macosx/debug/riela workflow run "$workflow" \
--workflow-definition-dir ./examples \
--mock-scenario "./examples/$workflow/mock-scenario.json" \
--session-store tmp/skill-chat-verification/sessions \
--output json |
tee "tmp/skill-chat-verification/logs/run-$workflow-node-ref-handoff.json"
done
Expected mock evidence:
status is completed.exitCode is 0.- Final
rootOutput.replyAs is rina.
- Final
rootOutput.handoffTrail is ["yui","mika","rina"].
- Final
rootOutput.handoff_yui, handoff_mika, and handoff_rina are all false.
Run an adversarial loop check when the file is available in tmp/ or recreate
it under tmp/skill-chat-verification/adversarial/. The adversarial Rina output
must try to set handoff_yui: true; the memory-write add-on must block it:
direnv exec . .build/arm64-apple-macosx/debug/riela workflow run \
telegram-agent-trio-chat \
--workflow-definition-dir ./examples \
--mock-scenario tmp/skill-chat-verification/adversarial/telegram-agent-trio-loop-mock.json \
--session-store tmp/skill-chat-verification/sessions \
--output json |
tee tmp/skill-chat-verification/logs/run-agent-trio-adversarial-loop.json
Expected adversarial evidence:
rootOutput.handoffGuard.blocked is true.rootOutput.handoffGuard.reason is target-persona-already-replied.rootOutput.handoffGuard.selectedTarget is yui.rootOutput.handoff_yui is false.rootOutput.replyText no longer contains a visible request to send the chat
back to Yui or another already visited persona.
For a live Telegram browser check, start a listener for the trio event root under
test, then post in Telegram Web. Keep all transient event roots and logs under
tmp/:
rm -rf tmp/live-agent-trio-node-ref
mkdir -p \
tmp/live-agent-trio-node-ref/events/sources \
tmp/live-agent-trio-node-ref/events/bindings \
tmp/live-agent-trio-node-ref/events/destinations \
tmp/live-agent-trio-node-ref/sessions
cp examples/event-sources/.riela-events/sources/telegram-gateway-personas.json \
tmp/live-agent-trio-node-ref/events/sources/
cp examples/event-sources/.riela-events/bindings/telegram-gateway-personas-to-workflow.json \
tmp/live-agent-trio-node-ref/events/bindings/
cp examples/event-sources/.riela-events/destinations/telegram-gateway-persona-replies.json \
tmp/live-agent-trio-node-ref/events/destinations/
direnv exec . .build/arm64-apple-macosx/debug/riela events validate \
--workflow-definition-dir ./examples \
--event-root tmp/live-agent-trio-node-ref/events
direnv exec . .build/arm64-apple-macosx/debug/riela events serve \
--workflow-definition-dir ./examples \
--event-root tmp/live-agent-trio-node-ref/events \
--session-store tmp/live-agent-trio-node-ref/sessions \
2>&1 | tee tmp/live-agent-trio-node-ref/events-serve.log
Open the exact Telegram chat URL in Brave Browser with Computer Use and send a
fresh marker message such as:
@YuiCodexF0529Bot nodeRef-live-HHMM final sanitizer retest. Ask @mikatrend0529bot then @rinacursor0529bot for short opinions. Mika must not answer for Rina.
Live Telegram is verified only when all of these are true:
- Browser shows the user marker message.
- Browser shows visible Yui, Mika, and Rina replies for that same marker.
serve-record.json reports lastReplyAs as yui,mika,rina.- The matching Telegram history file contains the marker plus three assistant
entries with
replyAs yui, mika, and rina.
- The newest SQLite workflow snapshot is
completed, has final
rootOutput.replyAs = rina, and has
rootOutput.handoffTrail = ["yui","mika","rina"].
- Rina's final visible reply does not include a handoff request or continuation
cue back to Yui or Mika, such as
@Yui, @Mika, 次はミカ, or another
already visited persona request.
Useful evidence commands:
cat tmp/live-agent-trio-node-ref/events/serve-record.json
tail -n 60 tmp/live-agent-trio-node-ref/events/telegram-history/*/*.json
sqlite3 tmp/live-agent-trio-node-ref/sessions/runtime-records/runtime-message-log.sqlite \
"select workflow_execution_id, json_extract(root_output_json,'$.replyAs'), json_extract(root_output_json,'$.replyText'), json_extract(root_output_json,'$.handoffTrail'), json_extract(session_json,'$.status'), updated_at from workflow_runtime_snapshots order by updated_at desc limit 5;"
For a live Discord browser check, use the checked-in persona event source as a
template and inject the target guild/channel from the local environment. The
example source may intentionally contain placeholder IDs, so do not run it
directly without this rewrite:
rm -rf tmp/live-discord-node-ref
mkdir -p \
tmp/live-discord-node-ref/events/sources \
tmp/live-discord-node-ref/events/bindings \
tmp/live-discord-node-ref/events/destinations \
tmp/live-discord-node-ref/sessions \
tmp/live-discord-node-ref/logs
direnv exec . sh -lc '
set -eu
jq --arg guild "$RIELA_DISCORD_SERVER_ID" --arg channel "$RIELA_DISCORD_CHANNEL_ID" \
".guildIds = [\$guild] | .channels = [(.channels[0] | .id = \$channel)]" \
examples/event-sources/.riela-events/sources/discord-gateway-personas.json \
> tmp/live-discord-node-ref/events/sources/discord-gateway-personas.json
'
cp examples/event-sources/.riela-events/bindings/discord-gateway-personas-to-workflow.json \
tmp/live-discord-node-ref/events/bindings/
cp examples/event-sources/.riela-events/destinations/discord-gateway-persona-replies.json \
tmp/live-discord-node-ref/events/destinations/
direnv exec . .build/arm64-apple-macosx/debug/riela events validate \
--workflow-definition-dir ./examples \
--event-root tmp/live-discord-node-ref/events \
--output json | tee tmp/live-discord-node-ref/logs/events-validate.json
direnv exec . .build/arm64-apple-macosx/debug/riela events serve \
--workflow-definition-dir ./examples \
--event-root tmp/live-discord-node-ref/events \
--session-store tmp/live-discord-node-ref/sessions \
2>&1 | tee tmp/live-discord-node-ref/logs/events-serve.log
Open the Discord channel URL for $RIELA_DISCORD_SERVER_ID and
$RIELA_DISCORD_CHANNEL_ID in Brave Browser with Computer Use, then send a fresh
marker such as:
Yui, nodeRef-discord-live-HHMM final sanitizer retest. Ask Mika then Rina for short opinions. Mika must not answer for Rina.
Live Discord is verified only when all of these are true:
- Browser shows the marker in the target channel.
- Browser shows visible Yui, Mika, and Rina replies for that same marker.
serve-record.json reports lastReplyAs as yui,mika,rina and
lastReplyDispatchCount as 3.- The matching Discord history file contains the marker plus three assistant
entries with
replyAs yui, mika, and rina.
- The newest SQLite workflow snapshot is
completed, has final
rootOutput.replyAs = rina, and has
rootOutput.handoffTrail = ["yui","mika","rina"].
- Rina's final visible reply does not include a handoff request or continuation
cue back to Yui or Mika.
Useful Discord evidence commands:
cat tmp/live-discord-node-ref/events/serve-record.json
tail -n 80 tmp/live-discord-node-ref/events/discord-history/*/*.json
sqlite3 tmp/live-discord-node-ref/sessions/runtime-records/runtime-message-log.sqlite \
"select workflow_execution_id, json_extract(root_output_json,'$.replyAs'), json_extract(root_output_json,'$.replyText'), json_extract(root_output_json,'$.handoffTrail'), json_extract(session_json,'$.status'), updated_at from workflow_runtime_snapshots order by updated_at desc limit 5;"
If a marker appears in Telegram or Discord but no workflow session is created,
check the event source offset and seen-message files before retrying. Do not call
the run verified until either the browser shows fresh bot replies or
SQLite/session evidence proves the listener processed the marker. Stop the manual
listener with Ctrl-C after the check unless the user asks to keep it running.
Known Example Targets
- Telegram trio:
examples/telegram-agent-trio-chat
- Discord trio:
examples/discord-agent-trio-chat
- Matrix trio:
examples/matrix-agent-trio-chat
- Shared Yui/Mika/Rina persona nodes:
examples/shared-agent-trio-personas
- Raw log and daily summary memory split:
examples/chat-memory-raw-and-daily-summary
- Yui image asset:
examples/telegram-agent-trio-chat/assets/icons/yui-codex.png
For Matrix, prefer the deterministic local sample when live homeserver credentials are unavailable. Read matrix-local.md when Matrix needs local verification.
Recent History Evidence
When asked to reuse previous Codex evidence, search only session logs and redact credentials before displaying anything:
find "$HOME/.codex/sessions" -type f -mtime -4 -name '*.jsonl' -print0 |
xargs -0 rg -n -i 'telegram|discord|matrix|RielaApp|events serve|events list|completed|reply|records|receipt'
Do not trust broad search output directly because Codex session files include developer prompts, tool schemas, and possibly secret-looking strings. Read history-evidence.md for the known successful artifacts and the safer interpretation checklist.
For the 2026-06-22 four-day history search, treat tmp/trio-memory-recreate/{telegram,discord,matrix} as the known good three-platform memory regression. It proves deterministic Telegram/Discord/Matrix workflow memory behavior when each run is completed, exitCode=0, nodeExecutions=13, transitions=12, the SQLite counts are 1|2|3, and final rootOutput.replyAs is rina. It is not live browser-chat proof unless paired with a fresh visible reply, receipt, or post-message session from the same run.
Attachment Blocker Check
Before testing the image flow, inspect gateway support:
rg -n 'caption|photo|imagePaths|attachments|message\.text' Sources/RielaCLI/EventLiveServe*.swift
The image flow is blocked if:
- Telegram accepts only
message.text and ignores caption or photo.
- Telegram/Discord populate attachment descriptors but leave
imagePaths: [] while the workflow expects real image analysis.
- The event source config says
resolveFilePaths: true but the gateway implementation does not download or materialize image files.
Report these as implementation blockers, not test success.
Telegram File Memory Live Check
Use this when asked to prove that Telegram image memory is stored as file-backed memory and can be reused by a later persona turn.
Start the listener for the installed Telegram SDK trio workflow:
direnv exec . .build/arm64-apple-macosx/debug/riela events serve \
--workflow-definition-dir "$HOME/.riela/workflows" \
--event-root "$HOME/.riela/workflows/telegram-sdk-trio-chat/.riela-events" \
--session-store "$HOME/.riela/sessions"
Confirm the listener before sending messages:
ps aux | rg -i 'riela events serve' | rg -v 'rg -i'
cat "$HOME/.riela/workflows/telegram-sdk-trio-chat/.riela-events/serve-record.json"
In Telegram Web, open the exact target chat URL, attach the Yui image asset as Photo or Video, and send a caption that mentions Mika, for example:
@mikatrend0529bot Live file-memory retest HHMM: What is visible in this image? Reply briefly.
After Mika replies, verify all three layers:
find "$HOME/.riela/workflows/telegram-sdk-trio-chat/.riela-events/attachments" \
-type f -mmin -10 -print -exec file {} \;
sqlite3 "$PWD/.riela/memory/chat-memory.sqlite" \
'select file_id,record_id,path,media_type,kind,name,size_bytes from memory_files order by file_id desc limit 10;'
jq -r '.session | [.sessionId,.status,.entryStepId,.createdAt,.updatedAt] | @tsv' \
"$HOME/.riela/sessions"/telegram-sdk-trio-chat-session-*.json | tail
The expected evidence is:
- Telegram event root has a fresh downloaded image under
.riela-events/attachments/....
memory_files has a row for the same logical image copied under .riela/memory/files/chat-memory/<record-id>/....- The session starts at
save-chat-event-memory, not a persona node.
load-mika-chat-memory returns memoryAttachmentCountRead > 0 and non-empty imagePaths.- The visible Mika reply describes the image content, not only that a square photo exists.
Then ask Rina about the same image without reattaching it:
@rinacursor0529bot Live file-memory retest HHMM: What image did I just show Mika? Reply briefly.
Verify the follow-up session:
jq -r '.session.executions[] |
[.stepId,.status,(.acceptedOutput.payload.operation // .acceptedOutput.payload.replyAs // ""),
(.acceptedOutput.payload.memoryAttachmentCountRead // ""),
(.acceptedOutput.payload.imagePaths // "" | tostring),
(.acceptedOutput.payload.replyText // "")] | @tsv' \
"$HOME/.riela/sessions/telegram-sdk-trio-chat-session-<id>.json"
The Rina turn is verified only when load-rina-chat-memory returns non-empty imagePaths and the visible Rina reply describes the previously stored image. Stop the manual listener after the check with Ctrl-C unless the user asks to keep it running.
Final Report
Include:
- RielaApp launch result and daemon child status.
- Which chat surfaces were tested.
- Message timestamp or latest event/session id.
- Whether memory write/update/read was observed.
- Whether image attachment content was actually available to the model.
- Any missing env vars by name only.
