feat(config): configurable agent-loop retry limits with pause-and-ask (#897) - #904
Open
shoryabansalgithub wants to merge 19 commits into
Open
Conversation
… loops Adds a nanocoder.retries section to agents.config.json exposing the previously hardcoded conversation-loop caps: maxRepeatedToolCalls (default 3), maxEmptyTurns (default 2), maxMalformedRetries (default 2). Defaults preserve existing behaviour exactly. When the repeated-tool-call limit is hit in an interactive session the loop now pauses and asks the user whether to continue (granting another window of attempts) or stop, instead of always hard-stopping. Headless and non-interactive runs keep the hard stop since there is nobody to ask. Named 'retries' rather than the issue's proposed 'maxRetries' because maxRetries is already a public per-provider setting that caps network request retries. Closes Nano-Collective#897
Review follow-up: keep currentRepeatedCount const (it names the detected streak used in user-facing messages) and carry the post-prompt streak in repeatedCountForNextTurn so the recursion site shows where a reset can come from.
…, duplicate flush
The plain runtime (auto-selected for 'nanocoder run' in CI and non-TTY environments) previously had no repeated-tool-call, empty-turn, or malformed-retry protection at all - it was bounded only by headless.maxTurns (default 200), which is exactly the unattended token-drain scenario issue Nano-Collective#897 targets. The loop now honors the same nanocoder.retries limits as the interactive runtime: consecutive identical tool calls, consecutive empty turns (nudged up to the cap), and malformed tool-call self-correction retries each hard-stop with a clear error naming the setting once their cap is hit - there is no user to ask in a plain run. Empty assistant messages are no longer appended to history (providers reject them), matching the interactive loop. Docs and the changeset now describe both runtimes.
feat(chat-handler): configurable agent-loop retry limits with pause-and-ask on tool loops
feat(config): configurable agent-loop retry limits with pause-and-ask
…nd cap subagent loops Apply the four captain-decided fixes from review round 2 of Nano-Collective#897: - Unknown-tool turns now count toward the maxRepeatedToolCalls streak in both runtimes, so a model stuck calling a nonexistent tool trips the cap (interactive: pause-and-ask; plain: hard stop) instead of looping until the turn ceiling. The streak/prompt logic is shared between the valid and unknown-tool branches. - The subagent executor loop, previously uncapped, now applies maxRepeatedToolCalls and stops with an error naming the setting. - Docs: loud CI warning that legitimate polling patterns in --plain runs hard-stop at the cap and need maxRepeatedToolCalls raised; fixed the off-by-one in the retry-limits table (a limit of 3 executes the repeated call twice and pauses on the third emission) and reworded the continue option to 'check again after N more' to match the implementation. - Tests for the unknown-tool path in both loops and for all subagent cap paths (trip, one-under, unknown-tool, custom limit). Also fixes a misplaced biome-ignore in chat-panel-harness.ts that left two standing lint warnings.
shoryabansalgithub
requested review from
Avtrkrb,
akramcodez and
will-lamerton
as code owners
August 19, 2026 12:47
Contributor
There was a problem hiding this comment.
Pull request overview
Adds configurable agent-loop retry limits (nanocoder.retries) so the interactive and --plain conversation loops can’t silently spin forever (and drain tokens) when a model repeats tool calls, emits empty turns, or produces malformed tool-call text.
Changes:
- Introduces
RetryLimitsConfig+ config loader support (getRetryLimits) with defaults matching the prior hardcoded caps. - Updates both the interactive loop and the
--plainruntime to enforce the configured caps (including pause-and-ask behavior for repeated tool calls in interactive sessions). - Adds/updates AVA tests and docs describing the new configuration and behavior.
Reviewed changes
Copilot reviewed 14 out of 14 changed files in this pull request and generated 1 comment.
Show a summary per file
| File | Description |
|---|---|
| source/types/config.ts | Adds RetryLimitsConfig and AppConfig.retries typing. |
| source/config/index.ts | Loads nanocoder.retries with clamping/defaults; exports getRetryLimits(). |
| source/constants.ts | Documents constants as defaults for the new config surface. |
| source/hooks/chat-handler/conversation/conversation-loop.tsx | Uses configurable limits; adds interactive pause-and-ask flow for repeated tool calls. |
| source/plain/conversation.ts | Enforces the same limits in the --plain loop (repeats/empty/malformed). |
| source/config/index.spec.ts | Adds unit coverage for retry-limits config loading, defaults, clamping, invalid types. |
| source/hooks/chat-handler/conversation/conversation-loop.spec.ts | Adds/extends tests for the interactive pause-and-ask repeated-tool-call behavior and limit overrides. |
| source/plain/conversation.spec.ts | Updates tests to cover empty-turn nudging and retry-limit behavior in --plain. |
| docs/configuration/index.md | Documents the new nanocoder.retries section and semantics. |
| docs/configuration/providers/index.md | Clarifies provider maxRetries vs agent-loop retry limits. |
| docs/features/development-modes.md | Notes repeated-tool-call safeguard remains even in yolo mode. |
| docs/features/commands.md | Documents hard-stop behavior for retry limits in non-interactive run/--plain. |
| CLAUDE.md | Updates tool-call parsing description to mention malformed retry limit. |
| .changeset/configurable-retry-limits.md | Adds changeset entry for the feature. |
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
Comment on lines
+346
to
+363
| // Skip appending a fully-empty assistant message (no content, no tool | ||
| // calls): providers reject them, and the empty-turn nudge below re-asks | ||
| // without one — same rule the interactive loop applies. | ||
| const hasAssistantPayload = | ||
| cleanedContent.trim() || | ||
| validToolCalls.length > 0 || | ||
| errorResults.length > 0; | ||
| if (hasAssistantPayload) { | ||
| messages = [ | ||
| ...messages, | ||
| { | ||
| role: 'assistant', | ||
| content: cleanedContent, | ||
| tool_calls: validToolCalls.length > 0 ? validToolCalls : undefined, | ||
| reasoning: streamedReasoning || undefined, | ||
| }, | ||
| ]; | ||
| } |
chat-panel.js reads globalThis.NanocoderMentionUtils at boot since the @-mention autocomplete landed, and the real webview loads mention-utils.js first (see chat-panel.html). The VM harness only evaluated chat-panel.js, so every chat-panel spec failed with an undefined-global TypeError.
…the non-interactive approval exit The non-interactive approval exit saved history where the assistant message announced confirmTools' tool_calls but none of them received a tool result. A later session resume replays that history, and strict OpenAI-compatible providers reject announced tool_calls without results. Pair each unconfirmed tool with a cancellation result, the same pattern the interactive decline path already uses.
…subagent failure text
Author
|
Pushed a hardening round on top of the initial implementation: repeated invalid (unknown) tool calls now count toward the retry cap in all runtimes, assistant/tool message pairing is preserved on every error and abort path (prevents provider API errors on strict OpenAI-compatible backends), subagent loops get the same caps with partial output preserved, and docs/changeset wording was tightened. Full suite: 6,534 tests passing. |
akramcodez
approved these changes
Aug 21, 2026
akramcodez
left a comment
akramcodez
left a comment
Collaborator
There was a problem hiding this comment.
Hey @shoryabansalgithub, I went through the latest changes and did a deep review of the PR. Everything looks really solid now.
The implementation is well scoped, the retry-limit behavior is properly covered, the tool-call/result pairing issues are handled, and the test coverage is comprehensive. Typecheck, lint, and the relevant test suite are all passing as well.
I don't see any blocking issues from my side. The remaining points around ACP scope and the --plain behavior are already documented and look like reasonable design decisions.
Great work on this one.
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
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.
Description
Implements #897: configurable retry limits so a stuck agent cannot silently drain API tokens by retrying a failing operation forever.
Adds a new
nanocoder.retriessection toagents.config.jsonexposing three previously hardcoded agent-loop caps:maxRepeatedToolCalls(default 3, minimum 2) - consecutive identical tool calls before interventionmaxEmptyTurns(default 2, minimum 0) - consecutive empty assistant turns before erroringmaxMalformedRetries(default 2, minimum 0) - malformed tool-call self-correction attempts before erroringWhen the repeated-tool-call limit is hit in an interactive session, the agent pauses and asks the user whether to continue or stop instead of hard-stopping. Choosing continue grants another window and re-prompts when it is spent, reporting the cumulative repeat total. Any non-exact answer (including Esc) fails safe to stop. Non-interactive and headless runs keep a hard stop, and the same caps are enforced in the
--plainruntime loop so unattended CI runs are protected too.The issue proposed the name
maxRetries, but that is already a public per-provider setting meaning network request retries (default 2). Reusing it would silently change meaning for existing configs, so the agent-loop caps live under a distinctnanocoder.retriessection instead. Defaults preserve current behavior.MAX_TOOL_STEPSandMAX_COMPACT_RETRIESare deliberately left hardcoded (per-request step cap and a sub-branch of the empty-turn path, not user-facing retry drains).Supersedes #902, reopened from a dedicated feature branch per maintainer request.
Type of Change
Changeset
pnpm changeset) describing this change for the changelogTesting
Automated Tests
.spec.ts/tsxfilespnpm test:allcompletes successfully)New ava tests cover: config loader defaults, custom values, partial config, clamping, and invalid types; stop at the default limit; continue grants a new window then re-prompts; one-under-limit boundary; custom limits; non-interactive never prompts; custom malformed and empty-turn limits; plain-runtime cap enforcement.
Manual Testing
End-to-end proof was captured against a local mock OpenAI-compatible server deliberately repeating the same failing tool call: interactive pause-and-ask flow (pause, continue, resume, cumulative re-prompt, stop), Esc fail-safe, and the non-interactive hard stop, with measured API request counts confirming the limits (default stops at exactly 3 calls, limit 5 at exactly 5, limit 1 clamps to the minimum 2). The built CLI was smoke-tested including real
agents.config.jsonresolution. No live paid provider was used.Checklist
docs/configuration/index.mdand related pages)