recordAssistantUsage accumulated cacheRead across the entire multi-turn
run, and totalTokens was clamped to contextTokens. This caused
session_status to report 100% context usage regardless of actual load.
Changes:
- run.ts: capture lastTurnTotal from the most recent model call and
inject it into the normalized usage before it reaches agentMeta.
- usage-reporting.test.ts: verify usage.total reflects current turn,
not accumulated total.
Fixes#17016
Wire media URL tracking through the embedded agent pipeline so that
media already sent via messaging tools is not delivered again by the
reply dispatcher.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Enable plugins to override the model and provider for agent runs by
returning modelOverride/providerOverride from the before_agent_start
hook. The hook is now invoked early in run.ts (before resolveModel)
so overrides take effect. The result is passed to attempt.ts via
earlyHookResult to prevent double-firing.
This enables security-critical use cases like routing PII-containing
prompts to local models instead of cloud providers.
* Agents: add subagent orchestration controls
* Agents: add subagent orchestration controls (WIP uncommitted changes)
* feat(subagents): add depth-based spawn gating for sub-sub-agents
* feat(subagents): tool policy, registry, and announce chain for nested agents
* feat(subagents): system prompt, docs, changelog for nested sub-agents
* fix(subagents): prevent model fallback override, show model during active runs, and block context overflow fallback
Bug 1: When a session has an explicit model override (e.g., gpt/openai-codex),
the fallback candidate logic in resolveFallbackCandidates silently appended the
global primary model (opus) as a backstop. On reinjection/steer with a transient
error, the session could fall back to opus which has a smaller context window
and crash. Fix: when storedModelOverride is set, pass fallbacksOverride ?? []
instead of undefined, preventing the implicit primary backstop.
Bug 2: Active subagents showed 'model n/a' in /subagents list because
resolveModelDisplay only read entry.model/modelProvider (populated after run
completes). Fix: fall back to modelOverride/providerOverride fields which are
populated at spawn time via sessions.patch.
Bug 3: Context overflow errors (prompt too long, context_length_exceeded) could
theoretically escape runEmbeddedPiAgent and be treated as failover candidates
in runWithModelFallback, causing a switch to a model with a smaller context
window. Fix: in runWithModelFallback, detect context overflow errors via
isLikelyContextOverflowError and rethrow them immediately instead of trying the
next model candidate.
* fix(subagents): track spawn depth in session store and fix announce routing for nested agents
* Fix compaction status tracking and dedupe overflow compaction triggers
* fix(subagents): enforce depth block via session store and implement cascade kill
* fix: inject group chat context into system prompt
* fix(subagents): always write model to session store at spawn time
* Preserve spawnDepth when agent handler rewrites session entry
* fix(subagents): suppress announce on steer-restart
* fix(subagents): fallback spawned session model to runtime default
* fix(subagents): enforce spawn depth when caller key resolves by sessionId
* feat(subagents): implement active-first ordering for numeric targets and enhance task display
- Added a test to verify that subagents with numeric targets follow an active-first list ordering.
- Updated `resolveSubagentTarget` to sort subagent runs based on active status and recent activity.
- Enhanced task display in command responses to prevent truncation of long task descriptions.
- Introduced new utility functions for compacting task text and managing subagent run states.
* fix(subagents): show model for active runs via run record fallback
When the spawned model matches the agent's default model, the session
store's override fields are intentionally cleared (isDefault: true).
The model/modelProvider fields are only populated after the run
completes. This left active subagents showing 'model n/a'.
Fix: store the resolved model on SubagentRunRecord at registration
time, and use it as a fallback in both display paths (subagents tool
and /subagents command) when the session store entry has no model info.
Changes:
- SubagentRunRecord: add optional model field
- registerSubagentRun: accept and persist model param
- sessions-spawn-tool: pass resolvedModel to registerSubagentRun
- subagents-tool: pass run record model as fallback to resolveModelDisplay
- commands-subagents: pass run record model as fallback to resolveModelDisplay
* feat(chat): implement session key resolution and reset on sidebar navigation
- Added functions to resolve the main session key and reset chat state when switching sessions from the sidebar.
- Updated the `renderTab` function to handle session key changes when navigating to the chat tab.
- Introduced a test to verify that the session resets to "main" when opening chat from the sidebar navigation.
* fix: subagent timeout=0 passthrough and fallback prompt duplication
Bug 1: runTimeoutSeconds=0 now means 'no timeout' instead of applying 600s default
- sessions-spawn-tool: default to undefined (not 0) when neither timeout param
is provided; use != null check so explicit 0 passes through to gateway
- agent.ts: accept 0 as valid timeout (resolveAgentTimeoutMs already handles
0 → MAX_SAFE_TIMEOUT_MS)
Bug 2: model fallback no longer re-injects the original prompt as a duplicate
- agent.ts: track fallback attempt index; on retries use a short continuation
message instead of the full original prompt since the session file already
contains it from the first attempt
- Also skip re-sending images on fallback retries (already in session)
* feat(subagents): truncate long task descriptions in subagents command output
- Introduced a new utility function to format task previews, limiting their length to improve readability.
- Updated the command handler to use the new formatting function, ensuring task descriptions are truncated appropriately.
- Adjusted related tests to verify that long task descriptions are now truncated in the output.
* refactor(subagents): update subagent registry path resolution and improve command output formatting
- Replaced direct import of STATE_DIR with a utility function to resolve the state directory dynamically.
- Enhanced the formatting of command output for active and recent subagents, adding separators for better readability.
- Updated related tests to reflect changes in command output structure.
* fix(subagent): default sessions_spawn to no timeout when runTimeoutSeconds omitted
The previous fix (75a791106) correctly handled the case where
runTimeoutSeconds was explicitly set to 0 ("no timeout"). However,
when models omit the parameter entirely (which is common since the
schema marks it as optional), runTimeoutSeconds resolved to undefined.
undefined flowed through the chain as:
sessions_spawn → timeout: undefined (since undefined != null is false)
→ gateway agent handler → agentCommand opts.timeout: undefined
→ resolveAgentTimeoutMs({ overrideSeconds: undefined })
→ DEFAULT_AGENT_TIMEOUT_SECONDS (600s = 10 minutes)
This caused subagents to be killed at exactly 10 minutes even though
the user's intent (via TOOLS.md) was for subagents to run without a
timeout.
Fix: default runTimeoutSeconds to 0 (no timeout) when neither
runTimeoutSeconds nor timeoutSeconds is provided by the caller.
Subagent spawns are long-running by design and should not inherit the
600s agent-command default timeout.
* fix(subagent): accept timeout=0 in agent-via-gateway path (second 600s default)
* fix: thread timeout override through getReplyFromConfig dispatch path
getReplyFromConfig called resolveAgentTimeoutMs({ cfg }) with no override,
always falling back to the config default (600s). Add timeoutOverrideSeconds
to GetReplyOptions and pass it through as overrideSeconds so callers of the
dispatch chain can specify a custom timeout (0 = no timeout).
This complements the existing timeout threading in agentCommand and the
cron isolated-agent runner, which already pass overrideSeconds correctly.
* feat(model-fallback): normalize OpenAI Codex model references and enhance fallback handling
- Added normalization for OpenAI Codex model references, specifically converting "gpt-5.3-codex" to "openai-codex" before execution.
- Updated the `resolveFallbackCandidates` function to utilize the new normalization logic.
- Enhanced tests to verify the correct behavior of model normalization and fallback mechanisms.
- Introduced a new test case to ensure that the normalization process works as expected for various input formats.
* feat(tests): add unit tests for steer failure behavior in openclaw-tools
- Introduced a new test file to validate the behavior of subagents when steer replacement dispatch fails.
- Implemented tests to ensure that the announce behavior is restored correctly and that the suppression reason is cleared as expected.
- Enhanced the subagent registry with a new function to clear steer restart suppression.
- Updated related components to support the new test scenarios.
* fix(subagents): replace stop command with kill in slash commands and documentation
- Updated the `/subagents` command to replace `stop` with `kill` for consistency in controlling sub-agent runs.
- Modified related documentation to reflect the change in command usage.
- Removed legacy timeoutSeconds references from the sessions-spawn-tool schema and tests to streamline timeout handling.
- Enhanced tests to ensure correct behavior of the updated commands and their interactions.
* feat(tests): add unit tests for readLatestAssistantReply function
- Introduced a new test file for the `readLatestAssistantReply` function to validate its behavior with various message scenarios.
- Implemented tests to ensure the function correctly retrieves the latest assistant message and handles cases where the latest message has no text.
- Mocked the gateway call to simulate different message histories for comprehensive testing.
* feat(tests): enhance subagent kill-all cascade tests and announce formatting
- Added a new test to verify that the `kill-all` command cascades through ended parents to active descendants in subagents.
- Updated the subagent announce formatting tests to reflect changes in message structure, including the replacement of "Findings:" with "Result:" and the addition of new expectations for message content.
- Improved the handling of long findings and stats in the announce formatting logic to ensure concise output.
- Refactored related functions to enhance clarity and maintainability in the subagent registry and tools.
* refactor(subagent): update announce formatting and remove unused constants
- Modified the subagent announce formatting to replace "Findings:" with "Result:" and adjusted related expectations in tests.
- Removed constants for maximum announce findings characters and summary words, simplifying the announcement logic.
- Updated the handling of findings to retain full content instead of truncating, ensuring more informative outputs.
- Cleaned up unused imports in the commands-subagents file to enhance code clarity.
* feat(tests): enhance billing error handling in user-facing text
- Added tests to ensure that normal text mentioning billing plans is not rewritten, preserving user context.
- Updated the `isBillingErrorMessage` and `sanitizeUserFacingText` functions to improve handling of billing-related messages.
- Introduced new test cases for various scenarios involving billing messages to ensure accurate processing and output.
- Enhanced the subagent announce flow to correctly manage active descendant runs, preventing premature announcements.
* feat(subagent): enhance workflow guidance and auto-announcement clarity
- Added a new guideline in the subagent system prompt to emphasize trust in push-based completion, discouraging busy polling for status updates.
- Updated documentation to clarify that sub-agents will automatically announce their results, improving user understanding of the workflow.
- Enhanced tests to verify the new guidance on avoiding polling loops and to ensure the accuracy of the updated prompts.
* fix(cron): avoid announcing interim subagent spawn acks
* chore: clean post-rebase imports
* fix(cron): fall back to child replies when parent stays interim
* fix(subagents): make active-run guidance advisory
* fix(subagents): update announce flow to handle active descendants and enhance test coverage
- Modified the announce flow to defer announcements when active descendant runs are present, ensuring accurate status reporting.
- Updated tests to verify the new behavior, including scenarios where no fallback requester is available and ensuring proper handling of finished subagents.
- Enhanced the announce formatting to include an `expectFinal` flag for better clarity in the announcement process.
* fix(subagents): enhance announce flow and formatting for user updates
- Updated the announce flow to provide clearer instructions for user updates based on active subagent runs and requester context.
- Refactored the announcement logic to improve clarity and ensure internal context remains private.
- Enhanced tests to verify the new message expectations and formatting, including updated prompts for user-facing updates.
- Introduced a new function to build reply instructions based on session context, improving the overall announcement process.
* fix: resolve prep blockers and changelog placement (#14447) (thanks @tyler6204)
* fix: restore cron delivery-plan import after rebase (#14447) (thanks @tyler6204)
* fix: resolve test failures from rebase conflicts (#14447) (thanks @tyler6204)
* fix: apply formatting after rebase (#14447) (thanks @tyler6204)
The UsageAccumulator sums cacheRead/cacheWrite across all API calls
within a single turn. With Anthropic prompt caching, each call reports
cacheRead ≈ current_context_size, so after N tool-call round-trips the
accumulated total becomes N × actual_context, which gets clamped to
contextWindow (200k) by deriveSessionTotalTokens().
Fix: track the most recent API call's cache fields separately and use
them in toNormalizedUsage() for context-size reporting. This makes
/status Context display accurate while preserving accumulated output
token counts.
Fixes#13698Fixes#13782
Co-authored-by: akari-musubi <259925157+akari-musubi@users.noreply.github.com>
* fix: suggest /reset in context overflow error message
When the context window overflows, the error message now suggests
using /reset to clear session history, giving users an actionable
recovery path instead of a dead-end error.
Closes#12940
Co-Authored-By: Claude <noreply@anthropic.com>
* fix: suggest /reset in context overflow error message (#12973) (thanks @RamiNoodle733)
---------
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Rami Abdelrazzaq <RamiNoodle733@users.noreply.github.com>
* initial commit
* feat: implement deriveSessionTotalTokens function and update usage tests
* Added deriveSessionTotalTokens function to calculate total tokens based on usage and context tokens.
* Updated usage tests to include cases for derived session total tokens.
* Refactored session usage calculations in multiple files to utilize the new function for improved accuracy.
* fix: restore overflow truncation fallback + changelog/test hardening (#11551) (thanks @tyler6204)
* fix: gracefully handle oversized tool results causing context overflow
When a subagent reads a very large file or gets a huge tool result (e.g.,
gh pr diff on a massive PR), it can exceed the model's context window in
a single prompt. Auto-compaction can't help because there's no older
history to compact — just one giant tool result.
This adds two layers of defense:
1. Pre-emptive: Hard cap on tool result size (400K chars ≈ 100K tokens)
applied in the session tool result guard before persistence. This
prevents extremely large tool results from being stored in full,
regardless of model context window size.
2. Recovery: When context overflow is detected and compaction fails,
scan session messages for oversized tool results relative to the
model's actual context window (30% max share). If found, truncate
them in the session via branching (creating a new branch with
truncated content) and retry the prompt.
The truncation preserves the beginning of the content (most useful for
understanding what was read) and appends a notice explaining the
truncation and suggesting offset/limit parameters for targeted reads.
Includes comprehensive tests for:
- Text truncation with newline-boundary awareness
- Context-window-proportional size calculation
- In-memory message truncation
- Oversized detection heuristics
- Guard-level size capping during persistence
* fix: prep fixes for tool result truncation PR (#11579) (thanks @tyler6204)
* fix(errors): return clear billing error message instead of cryptic raw error (#8136)
When an LLM API provider returns a credit/billing-related error (HTTP 402,
insufficient credits, low balance, etc.), OpenClaw now shows a clear,
actionable message instead of passing through the raw/cryptic error text:
⚠️ API provider returned a billing error — your API key has run out of
credits or has an insufficient balance. Check your provider's billing
dashboard and top up or switch to a different API key.
Changes:
- formatAssistantErrorText: detect billing errors via isBillingErrorMessage()
and return a user-friendly message (placed before the generic HTTP/JSON
error fallthrough)
- sanitizeUserFacingText: same billing detection for the sanitization path
- pi-embedded-runner/run.ts: add billingFailure detection in the profile
exhaustion fallback, so the FailoverError message is billing-specific
- Added 3 new tests for credit balance, HTTP 402, and insufficient credits
* fix: extract billing error message to shared constant
Previously, overflowCompactionAttempted was a boolean flag set once, preventing
recovery when a single compaction wasn't enough. Change to a counter allowing up
to 3 attempts before giving up. Also add diagnostic logging on overflow events to
help debug early-overflow issues.
Fixes sessions that hit context overflow during long agentic turns with many tool
calls, where one compaction round isn't sufficient to bring context below limits.
- Change MAX_IMAGE_BYTES from 6MB to 5MB to match Anthropic API limit
- Add isImageSizeError() to detect image size errors from API
- Handle image size errors with user-friendly message instead of retry
- Prevent failover for image size errors (not retriable)
Fixes#2271
* fix: detect Anthropic 'Request size exceeds model context window' as context overflow
Anthropic now returns 'Request size exceeds model context window' instead of
the previously detected 'prompt is too long' format. This new error message
was not recognized by isContextOverflowError(), causing auto-compaction to
NOT trigger. Users would see the raw error twice without any recovery attempt.
Changes:
- Add 'exceeds model context window' and 'request size exceeds' to
isContextOverflowError() detection patterns
- Add tests that fail without the fix, verifying both the raw error
string and the JSON-wrapped format from Anthropic's API
- Add test for formatAssistantErrorText to ensure the friendly
'Context overflow' message is shown instead of the raw error
Note: The upstream pi-ai package (@mariozechner/pi-ai) also needs a fix
in its OVERFLOW_PATTERNS regex: /exceeds the context window/i should be
changed to /exceeds.*context window/i to match both 'the' and 'model'
variants for triggering auto-compaction retry.
* fix(tests): remove unused imports and helper from test files
Remove WorkspaceBootstrapFile references and _makeFile helper that were
incorrectly copied from another test file. These caused type errors and
were unrelated to the context overflow detection tests.
* fix: trigger auto-compaction on context overflow promptError
When the LLM rejects a request with a context overflow error that surfaces
as a promptError (thrown exception rather than streamed error), the existing
auto-compaction in pi-coding-agent never triggers. This happens because the
error bypasses the agent's message_end → agent_end → _checkCompaction path.
This fix adds a fallback compaction attempt directly in the run loop:
- Detects context overflow in promptError (excluding compaction_failure)
- Calls compactEmbeddedPiSessionDirect (bypassing lane queues since already in-lane)
- Retries the prompt after successful compaction
- Limits to one compaction attempt per run to prevent infinite loops
Fixes: context overflow errors shown to user without auto-compaction attempt
* style: format compact.ts and run.ts with oxfmt
* fix: tighten context overflow match (#1627) (thanks @rodrigouroz)
---------
Co-authored-by: Claude <claude@anthropic.com>
Co-authored-by: Peter Steinberger <steipete@gmail.com>
Auth profiles in cooldown (due to rate limiting) were being attempted,
causing unnecessary retries and delays. This fix ensures:
1. Initial profile selection skips profiles in cooldown
2. Profile rotation (after failures) skips cooldown profiles
3. Clear error message when all profiles are unavailable
Tests added:
- Skips profiles in cooldown during initial selection
- Skips profiles in cooldown when rotating after failure
Fixes#1316
