Prospecting runs: saved_query_ids and chat_closed - #36
Merged
Merged
Conversation
Large runs now split their saved contact list into parts instead of truncating at 50 MiB (platform commit 2fbd2d6e5). saved_query_ids carries all of them in order, with saved_query_id staying the first entry for backward compatibility.
A chat closed for off-topic use cancels an unapproved or paused run with stop_reason "misuse" while an approved queued or running one keeps working. Callers need chat_closed on the run model to know when to stop chatting and start a new conversation. The API's 403 on a start past the daily new-chat limit already surfaces cleanly through the existing detail-message extraction, so no client change was needed there.
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
…pes in the contract script A key with a newline passed local validation and only failed later when the transport built the Idempotency-Key header. check_models() also compared schema property names only, so a type or requiredness change on a mirrored model (e.g. saved_query_ids, chat_closed) would pass silently.
The platform's GET /prospecting/runs now accepts before (a run ID) to page past a full page of results, ordered by created_at then run_id descending. Forward it through the SDK and CLI so callers can walk past the 50-row max instead of only ever seeing the newest page.
Prospecting runs can now pause at decision points (a pilot check on large lists, a search drifting off target, running short, reaching the target) and ask the user, or decide on their own in auto mode. The API defaults to auto so unattended callers never stall; the web chat defaults to asking, and the CLI is a person at a terminal, so it follows the chat rather than the API. The brief's checkpoints field is hidden from the platform's OpenAPI schema to keep the MCP tool listing small, so the request generator pins it by hand and the contract check allows it instead of flagging drift. Approval's field is pending until the platform deploys it. At a checkpoint, `prospecting wait` asks on a terminal and keeps waiting. Scripts and agents get a distinct exit code 7 with the question and suggested replies on stderr, so they can answer through `prospecting message` without parsing prose or mistaking a pause for success.
… list id The platform declares prospecting status, kind, role and stage as named type aliases, which its OpenAPI spec emits as $ref schemas. The contract check read every $ref as "object", so all four fields reported drift on models that actually matched. It now follows the ref to the target's type, so a genuine type change behind a ref is still caught. Runs gain companies_saved_query_id. The spec marks it required, but the SDK defaults it to None so it still parses servers from before the field existed; the contract check allows that one difference instead of failing on it.
The owner changed the rule: a poor pilot in auto mode now sharpens the criteria once and keeps the run going with a notice, instead of stopping outright. pilot_failed is now reserved for a re-pilot fit still under 20% (the searches are broken); a failed sharpening attempt falls back to the original criteria and continues.
--summary and --negate-summary help said semantic search, which the API never did. It matches every word of a term in any order; quoting keeps words together as one term but no longer means an exact phrase, since the persona summary index dropped phrase positions.
A released SDK should keep working when the API adds a status, stage or message kind, or raises a server-side cap. Response enums are now open (Literal | str), and a run's brief parses into a response model without the request-side limits. The contract check flags only drift that would break parsing, and CONTRIBUTING spells out the rules for new fields. The polling deadline is renamed timeout -> max_wait (signup takes request_timeout) so it no longer reads like the HTTP timeout, and the floor moves to Python 3.11 ahead of 3.10's end of life.
Customers go in as their own field so the platform can group them into segments and find lookalikes of each; approval can narrow which segments are expanded and turn results grouping on or off. The CLI reads the customer list from a file, the same format as --domains-file. The spec now carries every property that was pending in the generator (geo filters, approve checkpoints), so those overlays are gone and the merge path is tested with a stand-in entry instead.
The agentic prospecting default is moving from 25 companies and 2 contacts per company to 1000 and 1, matching the platform's own fallback when a brief omits both and planning can't infer a count. Unset fields are still dropped from the wire, so this only changes what an unset ProspectingBrief reports locally. _generated/requests.py is hand-edited ahead of a spec regen: the platform OpenAPI spec still carries the old defaults, so the next `gen_requests.py` run reproduces this once that spec change lands. Also fixes stale doc claims from the task-17 review: the CHANGELOG's seed_segments=None description said it keeps "every segment but Other" when it actually keeps the run's current selection (the plan card's default, or whatever a chat edit set) -- the approve() docstrings already had this right.
The API now keeps reply_pending true while a "segment these" request on a finished run is in flight, which can last more than an hour. Callers waiting on the flag need to know the acknowledgement is not the end, and that the flag expires after about 90 minutes if grouping dies.
Prospecting has never shipped, so its within-release iterations, server-only notes and a 'breaking' default change to an unreleased model were noise; one entry describes it. The daily conversation limit is a silent guardrail and stays out of public notes.
The API now uses DELETE on a run to remove the chat (cancelling it first if active), so cancel() moved to POST .../cancel and delete() wraps the new behaviour. No released version has prospecting.
The API can now retitle a prospecting chat in any status, so SDK users get the same control the app has. Takes a generated ProspectingRunUpdate like the other request-taking methods and returns the list() summary. No CLI command, matching delete(). Regenerated request models from the dev spec, which also picks up the reworded customer_domains description.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds saved_query_ids (every saved contact list part, in order) and chat_closed to ProspectingRun, and documents the misuse stop reason. Matches discolike-platform !908; merge when that ships. The contract check stays red for prospecting until the endpoints are deployed.
checkpointssetting (ask/auto) and how to answer a check-in; auto runs sharpen a weak pilot once and keep going.