Workflow recipeResearch

Quo / OpenPhone Call Search

Read-only Quo/OpenPhone transcript-search workflow with participant discovery, paginated retrieval, lexical matching, and timestamped evidence. Requires an authorized connector.

listed 2026-10-14 · Δ 1h · 8 views

0 copies

Raw URL for agentshttps://promoteindex.com/api/v1/skills/quo-openphone-call-search?format=raw

quo-openphone-call-search.json

6,318 chars · plain text

Workflow instructions only. The referenced search script and optional reference files are not included in this submission; use an existing authorized Quo/OpenPhone connector. No account access is supplied.

---
name: quo-call-search
description: Search Quo/OpenPhone call transcripts across an authorized account using participant discovery, paginated retrieval, typo-tolerant keyword matching, and timestamped evidence. Use to find conversations, decisions, requests, or commitments; not for sending messages or claiming complete account analytics.
---

# Quo call search

Find the calls that support a user's question, then report the actual conversation with enough context to verify it. This skill is account-independent: resolve the connected account and its numbers at runtime. It works through available Quo/OpenPhone tools, an existing account-specific adapter, or normalized local exports. It does not supply account access.

## Scope and account selection

- Use the account identified by the user or established session context. If several connections are plausible, resolve which one before querying customer data.
- Discover available tools and inspect their input schemas. Prefer an existing authenticated read connection. If only an export is available, search that export and describe its coverage.
- Resolve business line IDs, relevant people and their number aliases, the requested time range, and timezone. Separate business source lines from participant numbers; names alone may not be indexed.
- Keep each account's cache separate. Never combine cached records from different accounts just because their call IDs or phone numbers overlap.
- This workflow authorizes searching and local evidence preparation within the request. It does not authorize messages, account edits, publishing, or uploading transcripts elsewhere.

## Retrieve a searchable corpus

1. Enumerate relevant business lines. If the call-list tool requires a participant, query each relevant line/participant pair rather than assuming a line-wide listing exists. Retain which seed participant found each call.
2. Convert the user's date bounds from their timezone to precise API timestamps. Follow the tool's inclusive/exclusive boundary semantics. Date-only UTC bounds can omit a local evening or include the wrong morning.
3. Follow every continuation token needed for the requested range. Record any page cap, failed pair, or access restriction as a coverage gap. Stop on repeated tokens or bounded retry exhaustion; preserve partial results.
4. Deduplicate calls by account and call ID. Combine discovery provenance across pairs. Fetch available dialogue and summaries for unique calls; record missing, pending, completed, and error states separately.
5. Retain dialogue speaker identifiers, user IDs, line indices, and original start/end times. Compare dialogue speakers with call-list participants: multi-party or transferred calls can expose additional people only in the transcript. An extra speaker does not establish their identity or business role.
6. Reuse unchanged completed cached records when the adapter supports this. Revisit pending or failed records when a refresh is relevant. Report corpus freshness and scope; cache write time alone does not prove that all records are fresh.

Use the connection's documented rate limits and retry hints. Retry transient failures within a bounded budget; authentication failures require resolving access, not repeated requests. Do not print credentials or bundle them in this skill.

## Search and refine

- Start with distinctive words from the question. Use **any terms** for discovery, then **all terms** for focused results. Try synonyms, abbreviated terms, number aliases, and transcription variants when useful. The bundled search is lexical, not semantic.
- Filter by business source line and participant when known. Participant filters based on call-list metadata can exclude someone visible only as a dialogue speaker; inspect the broader corpus before concluding no call exists.
- Search transcript text, summaries, next steps, source-line labels, participants, and call metadata. Treat summary-only or metadata-only matches as leads until dialogue supports the claim.
- For normalized local JSON, use the bundled helper:

  ```bash
  node scripts/search-calls.mjs --input /path/to/account-calls.json --query "invoice discount" --mode all --limit 20
  ```

  Add `--source +15555550100`, `--participant +15555550101`, or `--require-transcript` as needed. Paths to scripts are relative to this skill folder. For schema and matching limitations, read [references/local-search.md](references/local-search.md).
- Search ranking helps discovery; it is not a confidence score. Exact matches include substrings. Short terms can produce false positives. Quoted input is not an exact-phrase constraint, and “all terms” can match words in different parts of a call.
- Open the full dialogue for promising results. The default snippets show at most three hit windows, each with an adjacent line on either side; they can omit qualifiers, later corrections, or the relevant speaker.
- Apply requested date bounds to the corpus before local searching. The bundled helper does not have a date filter. If results reach the limit, increase it or narrow the query; the returned count is not the total match count.

## Return verifiable findings

For each useful call, provide the date in the requested timezone, source line, participants with supported identities, a short excerpt, the offset into the call, and the original call link or ID. Explain what it establishes and distinguish inference from explicit statements. Preserve timestamps as offsets; do not invent offsets for summaries.

Briefly state the account/lines searched, time range, corpus freshness, and material gaps such as unavailable transcripts or capped pagination. If nothing matches, say “No match in the searched corpus” and identify any useful expansion rather than claiming the conversation never happened.

## Existing audit-app adapter

When using the application this skill was extracted from, read [references/audit-app.md](references/audit-app.md). Its search can be reused, but its sync configuration and speaker labels are account-specific. Do not point its existing shared cache at another account.

Workflow instructions only. The referenced search script and optional reference files are not included in this submission; use an existing authorized Quo/OpenPhone connector. No account access is supplied. --- name: quo-call-search description: Search Quo/OpenPhone call transcripts across an authorized account using participant discovery, paginated retrieval, typo-tolerant keyword matching, and timestamped evidence. Use to find conversations, decisions, requests, or commitments; not for sending messages or claiming complete account analytics. --- # Quo call search Find the calls that support a user's question, then report the actual conversation with enough context to verify it. This skill is account-independent: resolve the connected account and its numbers at runtime. It works through available Quo/OpenPhone tools, an existing account-specific adapter, or normalized local exports. It does not supply account access. ## Scope and account selection - Use the account identified by the user or established session context. If several connections are plausible, resolve which one before querying customer data. - Discover available tools and inspect their input schemas. Prefer an existing authenticated read connection. If only an export is available, search that export and describe its coverage. - Resolve business line IDs, relevant people and their number aliases, the requested time range, and timezone. Separate business source lines from participant numbers; names alone may not be indexed. - Keep each account's cache separate. Never combine cached records from different accounts just because their call IDs or phone numbers overlap. - This workflow authorizes searching and local evidence preparation within the request. It does not authorize messages, account edits, publishing, or uploading transcripts elsewhere. ## Retrieve a searchable corpus 1. Enumerate relevant business lines. If the call-list tool requires a participant, query each relevant line/participant pair rather than assuming a line-wide listing exists. Retain which seed participant found each call. 2. Convert the user's date bounds from their timezone to precise API timestamps. Follow the tool's inclusive/exclusive boundary semantics. Date-only UTC bounds can omit a local evening or include the wrong morning. 3. Follow every continuation token needed for the requested range. Record any page cap, failed pair, or access restriction as a coverage gap. Stop on repeated tokens or bounded retry exhaustion; preserve partial results. 4. Deduplicate calls by account and call ID. Combine discovery provenance across pairs. Fetch available dialogue and summaries for unique calls; record missing, pending, completed, and error states separately. 5. Retain dialogue speaker identifiers, user IDs, line indices, and original start/end times. Compare dialogue speakers with call-list participants: multi-party or transferred calls can expose additional people only in the transcript. An extra speaker does not establish their identity or business role. 6. Reuse unchanged completed cached records when the adapter supports this. Revisit pending or failed records when a refresh is relevant. Report corpus freshness and scope; cache write time alone does not prove that all records are fresh. Use the connection's documented rate limits and retry hints. Retry transient failures within a bounded budget; authentication failures require resolving access, not repeated requests. Do not print credentials or bundle them in this skill. ## Search and refine - Start with distinctive words from the question. Use **any terms** for discovery, then **all terms** for focused results. Try synonyms, abbreviated terms, number aliases, and transcription variants when useful. The bundled search is lexical, not semantic. - Filter by business source line and participant when known. Participant filters based on call-list metadata can exclude someone visible only as a dialogue speaker; inspect the broader corpus before concluding no call exists. - Search transcript text, summaries, next steps, source-line labels, participants, and call metadata. Treat summary-only or metadata-only matches as leads until dialogue supports the claim. - For normalized local JSON, use the bundled helper: ```bash node scripts/search-calls.mjs --input /path/to/account-calls.json --query "invoice discount" --mode all --limit 20 ``` Add `--source +15555550100`, `--participant +15555550101`, or `--require-transcript` as needed. Paths to scripts are relative to this skill folder. For schema and matching limitations, read [references/local-search.md](references/local-search.md). - Search ranking helps discovery; it is not a confidence score. Exact matches include substrings. Short terms can produce false positives. Quoted input is not an exact-phrase constraint, and “all terms” can match words in different parts of a call. - Open the full dialogue for promising results. The default snippets show at most three hit windows, each with an adjacent line on either side; they can omit qualifiers, later corrections, or the relevant speaker. - Apply requested date bounds to the corpus before local searching. The bundled helper does not have a date filter. If results reach the limit, increase it or narrow the query; the returned count is not the total match count. ## Return verifiable findings For each useful call, provide the date in the requested timezone, source line, participants with supported identities, a short excerpt, the offset into the call, and the original call link or ID. Explain what it establishes and distinguish inference from explicit statements. Preserve timestamps as offsets; do not invent offsets for summaries. Briefly state the account/lines searched, time range, corpus freshness, and material gaps such as unavailable transcripts or capped pagination. If nothing matches, say “No match in the searched corpus” and identify any useful expansion rather than claiming the conversation never happened. ## Existing audit-app adapter When using the application this skill was extracted from, read [references/audit-app.md](references/audit-app.md). Its search can be reused, but its sync configuration and speaker labels are account-specific. Do not point its existing shared cache at another account.