Skip to content

fix(web): match ChatGPT conversations by id across project paths - #227

Merged
ChesterRa merged 1 commit into
ChesterRa:mainfrom
chriscoveries:up/chatgpt-project-conversation-id
Nov 16, 2026
Merged

ChesterRa merged 1 commit into
ChesterRa:mainfrom
chriscoveries:up/chatgpt-project-conversation-id

Conversation

@chriscoveries

Copy link
Copy Markdown
Contributor

Fixes bound ChatGPT conversations returning chatgpt_bound_conversation_unavailable after a chat is moved into a Project.

ChatGPT serves the same chat at both /c/<id> and /g/g-p-<project>/c/<id>. The bound-target check compared full paths, so the move looked like a different conversation. This compares the conversation id parsed from the path when both URLs are https ChatGPT URLs on the same host and port. URLs without an id still use the existing normalized URL comparison.

Added one unit test covering /c/<id> against /g/g-p-<project>/c/<id> (match) and a different conversation id (no match).

@ChesterRa
ChesterRa merged commit 0214699 into ChesterRa:main Nov 16, 2026
7 checks passed
ChesterRa added a commit that referenced this pull request Nov 16, 2026
MCP recipes: `capability_import` accepted `command` as a string or argv
array, and a package recipe's fallback command without an identifier,
but installation used only the two candidate lists, so such records
imported and then failed to install. Import validation, installation and
the recipe fingerprint now share one parser. A string names one
executable and is never split by a shell; arguments need an argv array,
kept as written. Candidates are tried as `command_candidates`,
`command`, `fallback_command_candidates`, `fallback_command`; a package
recipe uses its registry only with a supported registry and a non-empty
identifier. Malformed command fields are rejected before the catalog
changes, and a string that looks like a whole command line gets a hint
to use an argv array. The fingerprint of an existing command recipe
containing only argv-array candidates is unchanged. Previously ignored
`command`, `fallback_command` or string candidates now affect the
fingerprint, so an old installation may report a changed recipe until
explicitly updated; its installed invoker is not replaced automatically.

Automation: a cron rule that already owes a run, for example one enabled
again after missing its time, reported the following occurrence as its next
fire, although the owed run could fire on the next eligible tick. It now
reports the run it owes, as interval and one-time rules already did; the Web
compares that with the daemon's `server_now` and labels the time as pending.
The run still fires once, on the first tick that can carry it out.

Capability page: the Enabled card counted only the page on screen, while the
Skills, MCP and Packs cards counted the whole search. `capability_overview`
now returns `enabled_count` over the same matches as `kind_counts`, after
seeding the Group's default capabilities as a state read does, and the Web
uses it.

Documentation: the home page's getting-started button links to the page's
own address, so opening it directly no longer returns 404. The operations
guide gains a table of what a daemon restart keeps, with a FAQ entry. The
daemon IPC standard documents the command forms, `kind_counts`,
`enabled_count` and the due `next_fire_at`.

ChatGPT: with #227 a delivery receipt also matches on the other route form
of the same conversation. Recovering an uncertain delivery from such a
receipt recorded the Project address as the conversation that received the
session bootstrap, so the next delivery sent the bootstrap into the same
chat again; it now records the bound conversation. The changes merged from
#226 and #227 gain regressions: both provisional URL forms stay pending
through pairing, and a conversation moved into a Project keeps its
exact-batch receipt without navigating or sending again, while another
conversation's receipt is rejected.

Validation: both maintainers reviewed the whole batch independently, and a
fresh-context review traced the four user paths across entry points before
the gate. The bootstrap-record and due-run wording findings are fixed here.
On the final tree: `npm run check`, `npm run build` and `npm test` (668
files, 5019 tests); the repository pytest (126); `git diff --check`;
`cargo fmt --all -- --check`;
`cargo clippy --workspace --all-targets --locked -- -D warnings`;
cargo-nextest's `ci` profile with real Chrome forbidden (4031 passed), its
`chrome` profile (47 passed) and the `daemon_self_launch` tests run serially
(3 passed), so each of the 4081 non-ignored tests among the 4091 listed ran
exactly once across the three gate groups. Four selected regressions also
ran before the full gate.
@ChesterRa

Copy link
Copy Markdown
Owner

Thanks for keeping the comparison scoped to the conversation ID and the same HTTPS origin. It is merged into main in 0214699. During validation we found a related recovery case: a receipt on the alternate route form could cause CCCC to send the session bootstrap again on the next delivery. Follow-up commit f4280c8 records the bound conversation as the bootstrap target, with passing regressions for both route forms.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants