feat(external-history): find Claude Code, Pi and OpenCode conversations without typed paths - #116

Problem

External history found only Codex conversations automatically. Every other source needed a file path typed into the dialog, and the dialog opened on that path box. Conversations from Claude Code, Pi and OpenCode on the same machine were invisible.

Change

• External history now finds Claude Code, Pi and OpenCode conversations in each provider's own folder, next to Codex. One table on the server decides how each provider is read, and a test fails when a built-in provider has no entry.
• One search request walks many conversations, newest first, and both clients follow empty pages by themselves. Sessions that T3 already owns are hidden by their stored ids.
• The typed ChatGPT export path feature is removed: the setting, the reader, and the path box on web and mobile.

Scope

• Covers web, desktop and mobile, the t3_external_history tool text, and the user docs.
• Grok, Antigravity, Cursor, Kiro, Muse, FX and ACP agents are listed in the dialog as not readable yet. Older OpenCode database formats show as unsupported.
• Search is still per source. One search box across all sources is a follow-up.
• This is a feature on the fork's own External history, requested by the maintainer in the T3 thread that produced this branch. Upstream has no External history.

Tradeoffs

• No persistent index. A quick text check on the raw files skips sessions that cannot match, so a no-hit search over 1,645 Claude sessions takes about 2 to 3 seconds.
• Codex keeps its app-server reader. Moving it to disk reads would lose archived threads.

Blast Radius

Saved export-path keys are ignored during settings decoding. The legacy export reader, path setting, and path-entry UI are fully removed. New source kinds are sent only to clients that ask for them, so older web and mobile builds still decode the source list. Reads never write to a provider folder, and folders under the T3 home or ~/Downloads are skipped.

Before and after

Before	After

Web was rechecked on the current head after the export path removal: the dialog has no path input, Codex, Claude Code and OpenCode show as Ready with nothing typed, and a Claude Code search for one word returned matches from 2,108 items with no extra clicks.

Mobile (iPhone 17e simulator, iOS 27)

Sources, nothing typed	Keyboard before	Keyboard after

Mobile was checked on a simulator paired to the same isolated dev server: the same three sources show as Ready, a Codex search and a Claude Code search both returned results (Claude Code: 2,108 items searched), and switching source clears the query. The run found that the keyboard covered the search field and the Search button; the screen now scrolls above the keyboard like the other settings screens.

Two JSONL history store tests failed on macOS only, because the temp folder resolves through a symlink there. The tests now use the resolved path. vp run check:changed passes on the pushed head.

Still not checked: the tool text against a live model (no ANTHROPIC_API_KEY and no Codex proxy on this machine; both cases build in a dry run), and Pi against real data (no Pi sessions on this machine; fixtures cover it).

Additional audit fixes

• Version search continuations with a digest of the ordered conversation listing. Additions, removals, reordering, or changes to the resumed transcript now return the existing changed-history error instead of silently skipping or repeating results. Search cursors are now V3; read/evidence cursors retain V2.
• Skip individual missing or inaccessible filesystem entries without aborting healthy history sources. Incomplete listings are not cached, so recovered access is retried even when the parent directory timestamp stays unchanged. Root and unexpected filesystem errors still propagate.
• Ignore malformed Claude profile settings while discovering sources, so one invalid disabled profile cannot prevent valid Claude profiles or other providers from loading.
• Reproduced four regressions before fixing them. The complete focused history selection passes: 23 tests across 5 suites, including file-access recovery and cursor consistency.
• Required changed-file validation passed the native addon, formatting, lint, both knip stages, server GraphQL validation, and 22 selected tests across 5 files. The cached typecheck could not start because this workspace lacks the required task-cache sandbox; the fresh CI run supplies full typecheck/build/browser coverage.
• The legacy ChatGPT export reader, export-path setting, and path-entry UI remain fully deleted. No model-facing prompt/schema changes were needed for these fixes.
• Published commit: f2c52b2. Its Git tree exactly matches the tested local files.
• Fresh checks for this commit: Fork CI and Agent Context Budget. Both were queued when this section was written. The prior commit's complete Fork CI run passed.

Initial audit follow-up

• Merged current main and retained full deletion of the legacy ChatGPT export implementation. No export reader, export-path setting, or path-entry UI remains. The only old-key reference is a regression test confirming stored unknown keys are ignored.
• Fixed JSONL and OpenCode prefilters dropping decoded matches containing JSON escapes or Unicode case folds.
• Refresh cached transcript metadata so appending an existing conversation updates newest-first ordering.
• Match Pi ownership by the session-file path T3 actually stores, so T3-owned Pi sessions are excluded.
• Keep OpenCode evidence tied to its own session while retaining database consistency checks and item-digest validation; unrelated conversations no longer invalidate references.
• Skip individual unavailable transcripts and keep conversation imports from allocating discarded tool payloads.
• Shortened the external-history tool description. The measured tool budget is 67,387 bytes, 32 bytes below current main (67,419); no budget gate was weakened.
• Seven backend regression failures were reproduced before fixing them. Afterward, 232 tests pass across all 18 selected non-browser files: 183 server, 8 mobile component, 23 web unit, 7 client-runtime, 3 contracts, and 8 OpenCode usage tests. Changed-file formatting, lint, generated GraphQL, dependency-patch checks, and the React compiler checks (5 changed files / 10 assertions) pass.
• The context-budget GitHub workflow passes on the published commit.
• The required changed-file checker was invoked. This workspace cannot start the task-cache sandbox; its broad root test and unused-export processes exited without diagnostics. A direct no-cache typecheck also terminated with exit 137 at the server package, even with one-CPU affinity, without TypeScript error diagnostics. Full typecheck verification remains outstanding for CI. Focused tests pass as listed above.
• Both external-history model-evaluation cases pass dry-run prompt/schema construction. Live evaluation is blocked by missing ANTHROPIC_API_KEY; browser verification is blocked before assertions by the environment's Unix-socket restriction. Native device verification remains incomplete.
• Published commit: 8e267c3. Its Git tree exactly matches the audited local files.
• Initial workflows: Fork CI and Agent Context Budget.

Original implementation verification

• Web, real browser, dev server on isolated state against real provider folders: with the switch on and nothing typed, Codex, Claude Code and OpenCode showed as Ready. A Claude search returned 9 matches from 2,108 messages in 5.3 s with no clicks, and a match opened its original text. The same flow was repeated on the current head; see Before and after.
• A read-only count over 1,645 real Claude sessions: 5 failed to load before the size fix, 0 after.
• vp run check:changed passes after the rebase (format, lint, knip, typecheck of 42 packages, 19 test files).

Built by Claude Fable 5.2 in Claude Code inside T3 Code, with Codex GPT-6.1 Sol and Claude Sonnet 5.5 as delegates.

🤖 Generated with Claude Code