| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
fix(tui): input auto-wrap, Alt+Enter newline, and intelligent arrow key navigation By the way, apply cargo fmt for code style consistency. Co-Authored-By: GLM-5 | 16 天前 | |
fix(bash): reclaim spilled output via TTL sweep + per-session byte budget Spilled bash output (>64KiB) was written to ~/.xiaoo/bash-output/<session>/ and never deleted: no Drop, no TTL, no session-close or startup cleanup, so distinct large outputs accumulated on disk indefinitely. Add two best-effort reclaimers in spill_full_output: - sweep_stale_spills: once-per-process sweep across all sessions, deleting files older than SPILL_TTL (7d) and pruning emptied session dirs. TTL-based so a killed session's orphans are still reclaimed on the next run. - enforce_session_budget: on each write, evict least-recently-modified files until the session dir is under SPILL_SESSION_BUDGET_BYTES (64MiB), never evicting the file just written (needed for later grep/read retrieval). Both swallow fs errors so cleanup never blocks the bash result. Add unit tests covering oldest-first eviction, kept-file protection, and the under-budget no-op. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> | 14 天前 | |
feat: initial commit with core features | 2 个月前 | |
feat(tool): trim built-in tool output to cut re-billed context Tool results are re-sent with the whole history every later turn, so oversized output is re-billed on every subsequent turn. Across the built-in tools: cap bash per-stream output 1MiB -> 64KiB with a head+tail window, and add retrieval-affordance notes to the file_read "unchanged" and glob "truncated" results so the model reuses what is already in context instead of re-reading or re-globbing. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> | 15 天前 | |
feat(tool): trim built-in tool output to cut re-billed context Tool results are re-sent with the whole history every later turn, so oversized output is re-billed on every subsequent turn. Across the built-in tools: cap bash per-stream output 1MiB -> 64KiB with a head+tail window, and add retrieval-affordance notes to the file_read "unchanged" and glob "truncated" results so the model reuses what is already in context instead of re-reading or re-globbing. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> | 15 天前 | |
feat(tool): trim built-in tool output to cut re-billed context Tool results are re-sent with the whole history every later turn, so oversized output is re-billed on every subsequent turn. Across the built-in tools: cap bash per-stream output 1MiB -> 64KiB with a head+tail window, and add retrieval-affordance notes to the file_read "unchanged" and glob "truncated" results so the model reuses what is already in context instead of re-reading or re-globbing. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> | 15 天前 | |
feat(tool): trim built-in tool output to cut re-billed context Tool results are re-sent with the whole history every later turn, so oversized output is re-billed on every subsequent turn. Across the built-in tools: cap bash per-stream output 1MiB -> 64KiB with a head+tail window, and add retrieval-affordance notes to the file_read "unchanged" and glob "truncated" results so the model reuses what is already in context instead of re-reading or re-globbing. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> | 15 天前 | |
fix(tui): input auto-wrap, Alt+Enter newline, and intelligent arrow key navigation By the way, apply cargo fmt for code style consistency. Co-Authored-By: GLM-5 | 16 天前 | |
refactor: improve spawn_subagent and join_subagent schemas with examples, fix compilation errors | 2 个月前 | |
Move conch backend ownership to gateway | 1 个月前 | |
feat(provider): fix deepseek hello-world && add deepseek v4 support | 2 个月前 | |
refactor: remove legacy Tool - Deleted legacy implementations. - Removed associated validation modules for both tools to eliminate unused code. - Updated module imports to reflect the removal of legacy components. | 1 个月前 | |
feat(backend): wire backend into the current design Co-authored-by: yaozhenhe<yaozhenhe@huawei.com> | 2 个月前 | |
feat(subagent): slim read-only subagent profile, turn budget, static repo map - Subagents get a dedicated slim system prompt and are restricted to read-only search tools (file_read/glob/grep), enforced even under an explicit --tools allowlist; skills are withheld from the child. - Caller-settable turn budget with a larger exploration default (~20); a fan-out prompt issues independent grep/glob/file_read in one turn. - A bounded static repository map is injected once into the MAIN agent's cache-stable prefix; subagents are excluded. - The subagent prompt gains a "disposable exploration worker" distillation instruction; the coordinator prompt-builder tests move to substring checks. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> | 15 天前 | |
feat(core): concurrent tool batch, verified finish, plan on control path Loop-level execution quality, independent of the inner model: - Independent tool calls in a turn run concurrently while preserving call-order history; a suspending join_subagent is ordered last and same-path side-effecting calls fall back to sequential, so a batch can neither strand a committed mutation across a suspend nor race two writers on one file. Several suspends in one turn each resolve on resume (no stranded tool_use blocks). - Verified finish: the first stop is not accepted; a one-shot completion checklist re-anchors on the original task and only a second stop completes. Gated to runs that expose tools (a conversational turn is not nudged). - Plan on the control path: todo_write exposes a reader; open items are re-injected each turn and consulted before a stop is accepted. - Error classification: transient (stream/io/timeout/throttle) vs permanent; a repeated identical failing OR successful call injects a strategy-change nudge; malformed tool names are normalized before rejection. - On the final turn tools are withheld so the model commits a final answer rather than a cut-off tool call. Stale tool-result bodies beyond a recent window are pruned to a marker (token-keyed). - Declares futures-util in crates/core for the concurrent join. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> | 15 天前 | |
feat(webfetch/websearch): auto-load system CA certs via rustls-native-certs Add build_http_client() to reqwest_util that loads native certificates before building the reqwest client. Both webfetch and websearch now trust the system certificate store (e.g. /etc/ssl/certs/ca-bundle.crt) in addition to the built-in webpki-roots, resolving SSL errors caused by enterprise MITM proxies or internal CAs. | 1 个月前 | |
feat(webfetch/websearch): auto-load system CA certs via rustls-native-certs Add build_http_client() to reqwest_util that loads native certificates before building the reqwest client. Both webfetch and websearch now trust the system certificate store (e.g. /etc/ssl/certs/ca-bundle.crt) in addition to the built-in webpki-roots, resolving SSL errors caused by enterprise MITM proxies or internal CAs. | 1 个月前 | |
feat(core): concurrent tool batch, verified finish, plan on control path Loop-level execution quality, independent of the inner model: - Independent tool calls in a turn run concurrently while preserving call-order history; a suspending join_subagent is ordered last and same-path side-effecting calls fall back to sequential, so a batch can neither strand a committed mutation across a suspend nor race two writers on one file. Several suspends in one turn each resolve on resume (no stranded tool_use blocks). - Verified finish: the first stop is not accepted; a one-shot completion checklist re-anchors on the original task and only a second stop completes. Gated to runs that expose tools (a conversational turn is not nudged). - Plan on the control path: todo_write exposes a reader; open items are re-injected each turn and consulted before a stop is accepted. - Error classification: transient (stream/io/timeout/throttle) vs permanent; a repeated identical failing OR successful call injects a strategy-change nudge; malformed tool names are normalized before rejection. - On the final turn tools are withheld so the model commits a final answer rather than a cut-off tool call. Stale tool-result bodies beyond a recent window are pruned to a marker (token-keyed). - Declares futures-util in crates/core for the concurrent join. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> | 15 天前 | |
feat: implement todo management tool and UI enhancements - Introduced a new todo_write tool for managing tasks within the application, allowing users to update their todo lists dynamically. - Enhanced the TUI to display the current task list and its status, integrating the todo management into the user interface. - Updated the application state to include a plan_state for tracking todo items, ensuring persistence across sessions. - Refactored rendering logic to accommodate the new sidebar for displaying todo items, improving user experience. - Added tests to verify the functionality of the todo management tool and its integration with the application state. | 1 个月前 |
| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
| 16 天前 | ||
| 14 天前 | ||
| 2 个月前 | ||
| 15 天前 | ||
| 15 天前 | ||
| 15 天前 | ||
| 15 天前 | ||
| 16 天前 | ||
| 2 个月前 | ||
| 1 个月前 | ||
| 2 个月前 | ||
| 1 个月前 | ||
| 2 个月前 | ||
| 15 天前 | ||
| 15 天前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 15 天前 | ||
| 1 个月前 |