Conversation
When a bot's provider quota runs out mid-conversation, raise a harness-native card in its thread offering the other configured engines (server/quota-switch.ts, following the peer-approval.ts card pattern). Choosing one switches modelSelection to a valid model for that engine (reusing default-model-selection.ts) and re-dispatches the last user message; "Not now" just settles the card. At most one card is pending per thread at a time. Also enrich the inline replay startTurn builds for a rewound/fresh engine with a compact line per tool call (server/turn-context.ts's new formatToolActivityLine), so an engine joining mid-thread learns which files were read or commands ran, not just what was said — the text-only transcript used for engineIsFresh, recovery text, and native-replay drivers is untouched. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…xact message Round 1 review fixes: - maybeRaiseQuotaCard now checks each candidate's snapshot() before offering it, so an unavailable engine never reaches the card (the click-time check in pickSelection stays as a second line of defence, but now posts a visible chip instead of failing silently). - OptionCardData gains fixedOptions (server/store.ts and its frontend duplicate in src/state/store.tsx); OptionCard.tsx stops rendering the free-text box for a fixedOptions card. An answer naming none of the offered engines still settles the card as "Not now" server-side. - switchAndRedispatch re-dispatches the exact user message captured at card-raise time (threaded through from the runtime.error handler's active-path lookup) instead of re-scanning at answer time, fixing stale/wrong/dropped resends including the image-only case; a message that can no longer be found posts a "could not be resent" chip instead of a false "continuing" one. - switchAndRedispatch also bails without writing a chip when patchBot returns null (bot deleted meanwhile). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
… an engine fallback ladder (phase 2 part 2) Every harness call a tool makes is bounded (60 s), a read that fails on the network is retried once, errors say what to do next, and a result over 24k characters is cut with the whole text kept by the harness and read back through tool_result_read. On a terminal provider failure a direct turn continues on the bot's named alternate (fallback.alternate), the switch is recorded on the task and said in the thread, and the exact failed message is re-sent once the failed turn settles; with side effects or no alternate, #1184's quota card asks the person instead (decision 12). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
|
Important Review skippedAuto reviews are disabled on base/target branches other than the default branch. Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Advanced Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
This was referenced Sep 19, 2026
milind-soni
added a commit
that referenced
this pull request
Sep 20, 2026
…1561) * fix(tools): bound agent replies with scoped temporary retrieval Adapt the bounded-result capability from #1293 without its retry or provider fallback behavior. Co-authored-by: aivsomkar <aivsomkar@gmail.com> * test: always close the tool-result verification fixture --------- Co-authored-by: aivsomkar <aivsomkar@gmail.com>
Owner
|
Closing at the maintainer’s request: we are deferring the Phase 2/3 harness stack to keep the current release scope smaller and reduce regression risk. This is a scope decision, not a finding that every change here is defective. The branch is being preserved so focused changes can be revisited separately. |
This branch was successfully deployed
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
2 participants
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.
Phase 2, part 2 of the harness programme (
docs/plans/2026-09-15-phase-2.md). Stacked on #1288 (part 1). Spec:docs/superpowers/specs/2026-09-15-phase-2-reliability-design.md.In plain terms
Tools that fail honestly. Every call a harness tool makes is bounded at 60 seconds; a read that fails on the network is tried once more; an error tells the model which tool, how long, whether it was retried, and what to do next, instead of a bare "fetch failed". A result over 24,000 characters is cut to its first 16,000 and the whole text is kept by the harness (redacted, 0600, per conversation) and read back in slices with
tool_result_read, so a long listing no longer floods the context or gets lost.Engines that fall back (decision 12). A bot may name an alternate engine (
fallback.alternateonPATCH /api/bots/:id). When a direct turn fails on a terminal provider error (quota, auth, overloaded, rate limits past the driver's own retries, server errors) the harness switches the task to the alternate, recordstask.fallback = { from, to, reason, at }, says so in the thread ("continuing on Grok after quota on claude"), and re-sends the exact message that failed once the failed turn has settled; the fresh-session replay carries the history. Two things stop the silent rung and leave it to the person: the failed turn had a side effect, or there is no alternate. Then, for a quota error, the card from #1184 (imported here) offers the other engines; otherwise the thread shows the error as today. A task falls back silently at most once.Measured
server/fallback.e2e.test.ts(real harness, fake Claude failing the way an out-of-quota CLI does, fake ACP as the alternate): the bot with an alternate gets its reply from the alternate, the task carries the record, the chip is there, the ledger's row names the alternate's driver; the bot without one gets the card and no reply.Every engine
Full on every engine: every engine mounts the same proxy, and the alternate can be any engine the workspace has; the failed message is re-dispatched with the same fresh-session replay an engine switch uses.
PRs checked (AGENTS.md rule 2)
Platforms
branching.test.tsChecks
pnpm typecheck,pnpm lint,pnpm i18n:checkclean; the suites the wiring touches green (the two that failed in one batch run pass alone and in the full suite: a shared temp folder, not the code). Full suite on the Phase 2 tip: see #1294 (parts 3–4), which carries the run for the whole branch.🤖 Generated with Claude Code