fix(opencode): deterministic queued message wrapping for prompt cache#21535
Open
Sathvik-1007 wants to merge 1 commit intoanomalyco:devfrom
Open
fix(opencode): deterministic queued message wrapping for prompt cache#21535Sathvik-1007 wants to merge 1 commit intoanomalyco:devfrom
Sathvik-1007 wants to merge 1 commit intoanomalyco:devfrom
Conversation
Sathvik-1007
force-pushed
the
fix/queued-message-cache-invalidation
branch
2 times, most recently
from
c998041 to
8f82457
Compare
Author
|
CI e2e failures unrelated to this PR. our change touches only |
fairyhunter13
added a commit
to fairyhunter13/opencode
that referenced
this pull request
Apr 8, 2026
Core cache optimizations: - Move mindContext from dynamicSystem to stableSystem (500-2000+ tokens/turn cached at BP1 for sessions with SessionMind context) - Split failureContext into stableFailures (prior turns, BP1 cached) and dynamicFailures (current turn only) using signature-based dedup - Add markLargeToolResults() pre-pass: cache_control on tool-result content parts >7000 chars (~2000 tokens), Anthropic direct + OpenRouter Claude - Fix stale parts reference bug in markLargeToolResults for multi-tool messages - Add compressImages() async pre-pass via sharp (PR anomalyco#21371): 3-phase quality->dimension->fallback compression prevents 5MB API limit errors - Session snapshot resets (resetFailureSnapshot/resetEnvDynamicSent) in cleanup - prompt_async idle race condition fix: check new messages before loop break Upstream PR cherry-picks: - PR anomalyco#21535: deterministic queued message wrapping eliminates per-turn cache miss - PR anomalyco#21492: tool evidence digest (evidence.ts) preserves context through compaction - PR anomalyco#21507: session processor single-flight summary dedup improvements - PR anomalyco#21528: prompt_async idle wakeup race condition fix - PR anomalyco#21500: Levenshtein O(min(N,M)) space with Int32Array two-row algorithm New tools (PR anomalyco#21399): - ContextUsageTool (check_context_usage): real-time token/cache usage reporting - NewSessionTool (new_session): TUI-only, abort + create new session - TuiEvent.SessionNew bus event and app.tsx handler - SDK types.gen.ts/sdk.gen.ts EventTuiSessionNew type Test infrastructure: - E2E cache tests (OPENCODE_E2E=1) verified 100% cache hit rate on T2+ - Unit tests for large-tool cache breakpoints (4 scenarios) - Fix pre-existing lsp-deps.test.ts assertion bug (LspTool in make() not all()) - Add await to all ProviderTransform.message() call sites (now async)
Sathvik-1007
force-pushed
the
fix/queued-message-cache-invalidation
branch
from
bb60b2b to
c7f6aff
Compare
fairyhunter13
added a commit
to fairyhunter13/opencode
that referenced
this pull request
Apr 8, 2026
…rchestrator, multi-credential, codebase indexer Core Features: - Session Mind with persistent memory across sessions - Orchestrator + Worker subagent architecture - Multi-credential OAuth with auto-refresh - Codebase indexer and watcher connectors - Footer status bar with live metrics Cache & Prompt Optimizations: - Move mindContext/failureContext to stable system prefix (BP1 cached) - Large tool result cache_control breakpoints (>7000 chars) - Deterministic message wrapping (PR anomalyco#21535) - Tool evidence digest through compaction (PR anomalyco#21492) - O(1) queue dequeue + single-flight summary (PR anomalyco#21507) - Levenshtein O(min(N,M)) space optimization (PR anomalyco#21500) - Three-phase image auto-compression (PR anomalyco#21371) - ContextUsage and NewSession tools (PR anomalyco#21399) - E2E cache integration tests with real Anthropic OAuth Session snapshot resets prevent memory leaks on session delete.
fairyhunter13
added a commit
to fairyhunter13/opencode
that referenced
this pull request
Apr 8, 2026
…rchestrator, multi-credential, codebase indexer Core Features: - Session Mind with persistent memory across sessions - Orchestrator + Worker subagent architecture - Multi-credential OAuth with auto-refresh - Codebase indexer and watcher connectors - Footer status bar with live metrics Cache & Prompt Optimizations: - Move mindContext/failureContext to stable system prefix (BP1 cached) - Large tool result cache_control breakpoints (>7000 chars) - Deterministic message wrapping (PR anomalyco#21535) - Tool evidence digest through compaction (PR anomalyco#21492) - O(1) queue dequeue + single-flight summary (PR anomalyco#21507) - Levenshtein O(min(N,M)) space optimization (PR anomalyco#21500) - Three-phase image auto-compression (PR anomalyco#21371) - ContextUsage and NewSession tools (PR anomalyco#21399) - E2E cache integration tests with real Anthropic OAuth Session snapshot resets prevent memory leaks on session delete.
fairyhunter13
pushed a commit
to fairyhunter13/opencode
that referenced
this pull request
Apr 8, 2026
…rchestrator, multi-credential, codebase indexer Core Features: - Session Mind with persistent memory across sessions - Orchestrator + Worker subagent architecture - Multi-credential OAuth with auto-refresh - Codebase indexer and watcher connectors - Footer status bar with live metrics Cache & Prompt Optimizations: - Move mindContext/failureContext to stable system prefix (BP1 cached) - Large tool result cache_control breakpoints (>7000 chars) - Deterministic message wrapping (PR anomalyco#21535) - Tool evidence digest through compaction (PR anomalyco#21492) - O(1) queue dequeue + single-flight summary (PR anomalyco#21507) - Levenshtein O(min(N,M)) space optimization (PR anomalyco#21500) - Three-phase image auto-compression (PR anomalyco#21371) - ContextUsage and NewSession tools (PR anomalyco#21399) - E2E cache integration tests with real Anthropic OAuth Session snapshot resets prevent memory leaks on session delete.
Sathvik-1007
force-pushed
the
fix/queued-message-cache-invalidation
branch
from
c7f6aff to
ec6261f
Compare
…che consistency Queued user message text was wrapped in <system-reminder> via in-memory mutation (prompt.ts:1483-1499) that never persisted to DB. Next turn re-fetched from DB, got unwrapped text, broke provider prompt cache. Move wrapping into toModelMessages with a 'remind' option so the same messages always serialize identically regardless of step counter. Fixes anomalyco#21518
Sathvik-1007
force-pushed
the
fix/queued-message-cache-invalidation
branch
from
ec6261f to
20a973f
Compare
fairyhunter13
added a commit
to fairyhunter13/opencode
that referenced
this pull request
Apr 9, 2026
Core cache optimizations: - Move mindContext from dynamicSystem to stableSystem (500-2000+ tokens/turn cached at BP1 for sessions with SessionMind context) - Split failureContext into stableFailures (prior turns, BP1 cached) and dynamicFailures (current turn only) using signature-based dedup - Add markLargeToolResults() pre-pass: cache_control on tool-result content parts >7000 chars (~2000 tokens), Anthropic direct + OpenRouter Claude - Fix stale parts reference bug in markLargeToolResults for multi-tool messages - Add compressImages() async pre-pass via sharp (PR anomalyco#21371): 3-phase quality->dimension->fallback compression prevents 5MB API limit errors - Session snapshot resets (resetFailureSnapshot/resetEnvDynamicSent) in cleanup - prompt_async idle race condition fix: check new messages before loop break Upstream PR cherry-picks: - PR anomalyco#21535: deterministic queued message wrapping eliminates per-turn cache miss - PR anomalyco#21492: tool evidence digest (evidence.ts) preserves context through compaction - PR anomalyco#21507: session processor single-flight summary dedup improvements - PR anomalyco#21528: prompt_async idle wakeup race condition fix - PR anomalyco#21500: Levenshtein O(min(N,M)) space with Int32Array two-row algorithm New tools (PR anomalyco#21399): - ContextUsageTool (check_context_usage): real-time token/cache usage reporting - NewSessionTool (new_session): TUI-only, abort + create new session - TuiEvent.SessionNew bus event and app.tsx handler - SDK types.gen.ts/sdk.gen.ts EventTuiSessionNew type Test infrastructure: - E2E cache tests (OPENCODE_E2E=1) verified 100% cache hit rate on T2+ - Unit tests for large-tool cache breakpoints (4 scenarios) - Fix pre-existing lsp-deps.test.ts assertion bug (LspTool in make() not all()) - Add await to all ProviderTransform.message() call sites (now async)
fairyhunter13
added a commit
to fairyhunter13/opencode
that referenced
this pull request
Apr 10, 2026
… cache hit-rate maximization
## Subagent Policy Refactor (packages/opencode/src/provider/provider.ts)
Introduce SUBAGENT_PROVIDER_POLICY as a declarative per-provider allow/deny
map for subagent model selection, replacing hardcoded Anthropic branching in
resolveSubagentModel. Future providers plug in via single map entries without
touching algorithm code. Universalizes the normal algorithm path so non-policy
providers (openai, google, xai, etc.) use fair dynamic tier-based selection
without Anthropic leakage.
Provider changes:
- Add SUBAGENT_PROVIDER_POLICY with anthropic entry (allow claude-sonnet-4-6)
- Add lookupSubagentPolicy() with proxy detection for claude-* via bedrock/vertex
- Add passesPolicy() helper (deny-before-allow precedence, undefined-safe)
- Refactor resolveSubagentModel Branches 1/2/4 to use policy lookup
- Universalize normal algorithm: remove Anthropic-specific leakage from
Branch 3/4, use cost comparison instead of includes('opus') for tiered log
- Raise FLOOR_CEILINGS.critical to 10 (unlocks premium non-Anthropic models;
Anthropic path returns early via policy, unaffected)
- Fix misleading comment in resolveSubagentVariant about Perplexity suppression
- Mark SUBAGENT_MODEL_ALLOWLIST @deprecated (superseded by policy map)
## Task Tool Changes (packages/opencode/src/tool/task.ts)
- Raise maxRetries from 2 to 3 for full low->critical escalation ladder
- Surface rate-limit silent downgrade: log.warn with reason 'itpm-headroom-low'
and output note when ITPM headroom forces complexity to low
- Document clampComplexity agent-ceiling bypass (JSDoc + inline escalation comment)
- Improve Agent.get error for unknown subagent_type: list available agents
## Tests
- Add packages/opencode/test/provider/variant-picker.test.ts (32 tests):
table-driven matrix: 7 variant configs x 4 complexity tiers
- Extend packages/opencode/test/tool/task-escalation.test.ts to 76 tests with
SUBAGENT_PROVIDER_POLICY coverage: passesPolicy unit tests, policy shape
assertions, resolveSubagentModel routing (opus/sonnet/haiku), proxy
detection for bedrock/vertex, openai/google/xai non-policy passthrough,
FLOOR_CEILINGS.critical=10 assertions
## Cache Hit-Rate Maximization (feat(cache))
Port and verification of upstream cache optimizations plus new E2E validation
through real Anthropic OAuth subscription path.
- Move git/PR guidance from bash.txt (uncached BP2 tool schema) to
anthropic.txt (cached BP1 system) -- saves ~161 bytes per request
repeated forever (ref: anomalyco#21345)
- Add E2E integration test cache-max-efficiency-e2e.test.ts exercising
the real OAuth subscription path with hafizludyanto credential.
Uses full OAuth plumbing (Authorization Bearer, ?beta=true query,
claude-cli/2.1.2 user-agent, oauth-2025-04-20 beta, Claude Code
system prefix). 11 tests validate cache_creation, cache_read,
hit rate >=85%, extended-cache-ttl-2025-04-11 header, no thinking
tokens, PR anomalyco#21535 queued message regression, model lock stability.
Measured hit rate: 93.7% on claude-sonnet-4-5-20250929 (sonnet-4-6
returns 404 on current OAuth subscription tier).
- Add skill opencode-prompt-cache-maximization/CLAUDE.md documenting
the 4-breakpoint strategy, 13 snapshot Maps, 8-dimension CacheGuard,
13 CacheMonitor assertions, cache-aware compaction, and debug toggles.
- Add skill opencode-anthropic-oauth-subscription/CLAUDE.md documenting
the OAuth subscription request envelope, token refresh flow
(client_id 9d1c250a-e61b-44d9-88ed-5944d1962f5e), auth.json layout,
model availability on subscription tier, and troubleshooting matrix.
Verified:
- bun typecheck clean
- variant-picker 32/32, task-escalation 76/76 (108 total, 0 fail)
- subagent-allowlist 99/106 (7 pre-existing gpt-5.2-codex fails, unchanged)
- Cache E2E 11/11, 93.7% hit rate
- 82 supporting unit tests pass (optimizations, cache-persist-snapshots,
prompt, message-v2)
## Adaptive Thinking Guard (packages/opencode/src/session/llm.ts)
Add a defensive guard in the LLM session layer that strips
`thinking: { type: "adaptive" }` from model parameters when the selected
model does not support adaptive thinking. Non-4.6 models (opus-4-5,
sonnet-4-5, haiku-4-5, and any other non-4.6 variant) would previously
receive the adaptive thinking option through user config or agent options
and respond with "adaptive thinking is not supported on this model" errors.
The guard detects these models by checking that the model ID does not include
"4-6" and removes the adaptive thinking config before the API call is made.
- Add packages/opencode/test/provider/adaptive-thinking-guard.test.ts
(17 tests) proving the guard strips adaptive thinking for non-4.6 models
and preserves it for claude-*-4-6 models
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
fairyhunter13
added a commit
to fairyhunter13/opencode
that referenced
this pull request
Apr 10, 2026
… cache hit-rate maximization
## Subagent Policy Refactor (packages/opencode/src/provider/provider.ts)
Introduce SUBAGENT_PROVIDER_POLICY as a declarative per-provider allow/deny
map for subagent model selection, replacing hardcoded Anthropic branching in
resolveSubagentModel. Future providers plug in via single map entries without
touching algorithm code. Universalizes the normal algorithm path so non-policy
providers (openai, google, xai, etc.) use fair dynamic tier-based selection
without Anthropic leakage.
Provider changes:
- Add SUBAGENT_PROVIDER_POLICY with anthropic entry (allow claude-sonnet-4-6)
- Add lookupSubagentPolicy() with proxy detection for claude-* via bedrock/vertex
- Add passesPolicy() helper (deny-before-allow precedence, undefined-safe)
- Refactor resolveSubagentModel Branches 1/2/4 to use policy lookup
- Universalize normal algorithm: remove Anthropic-specific leakage from
Branch 3/4, use cost comparison instead of includes('opus') for tiered log
- Raise FLOOR_CEILINGS.critical to 10 (unlocks premium non-Anthropic models;
Anthropic path returns early via policy, unaffected)
- Fix misleading comment in resolveSubagentVariant about Perplexity suppression
- Mark SUBAGENT_MODEL_ALLOWLIST @deprecated (superseded by policy map)
## Task Tool Changes (packages/opencode/src/tool/task.ts)
- Raise maxRetries from 2 to 3 for full low->critical escalation ladder
- Surface rate-limit silent downgrade: log.warn with reason 'itpm-headroom-low'
and output note when ITPM headroom forces complexity to low
- Document clampComplexity agent-ceiling bypass (JSDoc + inline escalation comment)
- Improve Agent.get error for unknown subagent_type: list available agents
## Tests
- Add packages/opencode/test/provider/variant-picker.test.ts (32 tests):
table-driven matrix: 7 variant configs x 4 complexity tiers
- Extend packages/opencode/test/tool/task-escalation.test.ts to 76 tests with
SUBAGENT_PROVIDER_POLICY coverage: passesPolicy unit tests, policy shape
assertions, resolveSubagentModel routing (opus/sonnet/haiku), proxy
detection for bedrock/vertex, openai/google/xai non-policy passthrough,
FLOOR_CEILINGS.critical=10 assertions
- Migrate 7 E2E test files to canonical personal-auth fixture for OAuth auth:
api-optimizations-e2e, cache-behavior-e2e, cache-drain-e2e,
cache-monitor-e2e, cache-monitor-lifecycle-e2e, cache-optimizations-e2e,
cache-optimizations-r1-r6-e2e
- Fix await lint warnings in api-optimizations-e2e.test.ts
## Cache Hit-Rate Maximization (feat(cache))
Port and verification of upstream cache optimizations plus new E2E validation
through real Anthropic OAuth subscription path.
- Move git/PR guidance from bash.txt (uncached BP2 tool schema) to
anthropic.txt (cached BP1 system) -- saves ~161 bytes per request
repeated forever (ref: anomalyco#21345)
- Add E2E integration test cache-max-efficiency-e2e.test.ts exercising
the real OAuth subscription path with hafizludyanto credential.
Uses full OAuth plumbing (Authorization Bearer, ?beta=true query,
claude-cli/2.1.2 user-agent, oauth-2025-04-20 beta, Claude Code
system prefix). 11 tests validate cache_creation, cache_read,
hit rate >=85%, extended-cache-ttl-2025-04-11 header, no thinking
tokens, PR anomalyco#21535 queued message regression, model lock stability.
Measured hit rate: 93.7% on claude-sonnet-4-5-20250929 (sonnet-4-6
returns 404 on current OAuth subscription tier).
- Add skill opencode-prompt-cache-maximization/CLAUDE.md documenting
the 4-breakpoint strategy, 13 snapshot Maps, 8-dimension CacheGuard,
13 CacheMonitor assertions, cache-aware compaction, and debug toggles.
- Add skill opencode-anthropic-oauth-subscription/CLAUDE.md documenting
the OAuth subscription request envelope, token refresh flow
(client_id 9d1c250a-e61b-44d9-88ed-5944d1962f5e), auth.json layout,
model availability on subscription tier, and troubleshooting matrix.
Verified:
- bun typecheck clean
- variant-picker 32/32, task-escalation 76/76 (108 total, 0 fail)
- subagent-allowlist 99/106 (7 pre-existing gpt-5.2-codex fails, unchanged)
- Cache E2E 11/11, 93.7% hit rate
- 82 supporting unit tests pass (optimizations, cache-persist-snapshots,
prompt, message-v2)
## Adaptive Thinking Guard (packages/opencode/src/session/llm.ts)
Add a defensive guard in the LLM session layer that strips
`thinking: { type: "adaptive" }` from model parameters when the selected
model does not support adaptive thinking. Non-4.6 models (opus-4-5,
sonnet-4-5, haiku-4-5, and any other non-4.6 variant) would previously
receive the adaptive thinking option through user config or agent options
and respond with "adaptive thinking is not supported on this model" errors.
The guard detects these models by checking that the model ID does not include
"4-6" and removes the adaptive thinking config before the API call is made.
- Add packages/opencode/test/provider/adaptive-thinking-guard.test.ts
(17 tests) proving the guard strips adaptive thinking for non-4.6 models
and preserves it for claude-*-4-6 models
- Add packages/opencode/test/fixture/personal-auth.ts as canonical OAuth
fixture used across all E2E provider tests
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
fairyhunter13
added a commit
to fairyhunter13/opencode
that referenced
this pull request
Apr 10, 2026
… cache hit-rate maximization
## Subagent Policy Refactor (packages/opencode/src/provider/provider.ts)
Introduce SUBAGENT_PROVIDER_POLICY as a declarative per-provider allow/deny
map for subagent model selection, replacing hardcoded Anthropic branching in
resolveSubagentModel. Future providers plug in via single map entries without
touching algorithm code. Universalizes the normal algorithm path so non-policy
providers (openai, google, xai, etc.) use fair dynamic tier-based selection
without Anthropic leakage.
Provider changes:
- Add SUBAGENT_PROVIDER_POLICY with anthropic entry (allow claude-sonnet-4-6)
- Add lookupSubagentPolicy() with proxy detection for claude-* via bedrock/vertex
- Add passesPolicy() helper (deny-before-allow precedence, undefined-safe)
- Refactor resolveSubagentModel Branches 1/2/4 to use policy lookup
- Universalize normal algorithm: remove Anthropic-specific leakage from
Branch 3/4, use cost comparison instead of includes('opus') for tiered log
- Raise FLOOR_CEILINGS.critical to 10 (unlocks premium non-Anthropic models;
Anthropic path returns early via policy, unaffected)
- Fix misleading comment in resolveSubagentVariant about Perplexity suppression
- Mark SUBAGENT_MODEL_ALLOWLIST @deprecated (superseded by policy map)
## Task Tool Changes (packages/opencode/src/tool/task.ts)
- Raise maxRetries from 2 to 3 for full low->critical escalation ladder
- Surface rate-limit silent downgrade: log.warn with reason 'itpm-headroom-low'
and output note when ITPM headroom forces complexity to low
- Document clampComplexity agent-ceiling bypass (JSDoc + inline escalation comment)
- Improve Agent.get error for unknown subagent_type: list available agents
## Tests
- Add packages/opencode/test/provider/variant-picker.test.ts (32 tests):
table-driven matrix: 7 variant configs x 4 complexity tiers
- Extend packages/opencode/test/tool/task-escalation.test.ts to 76 tests with
SUBAGENT_PROVIDER_POLICY coverage: passesPolicy unit tests, policy shape
assertions, resolveSubagentModel routing (opus/sonnet/haiku), proxy
detection for bedrock/vertex, openai/google/xai non-policy passthrough,
FLOOR_CEILINGS.critical=10 assertions
- Migrate 7 E2E test files to canonical personal-auth fixture for OAuth auth:
api-optimizations-e2e, cache-behavior-e2e, cache-drain-e2e,
cache-monitor-e2e, cache-monitor-lifecycle-e2e, cache-optimizations-e2e,
cache-optimizations-r1-r6-e2e
- Fix await lint warnings in api-optimizations-e2e.test.ts
## Cache Hit-Rate Maximization (feat(cache))
Port and verification of upstream cache optimizations plus new E2E validation
through real Anthropic OAuth subscription path.
- Move git/PR guidance from bash.txt (uncached BP2 tool schema) to
anthropic.txt (cached BP1 system) -- saves ~161 bytes per request
repeated forever (ref: anomalyco#21345)
- Add E2E integration test cache-max-efficiency-e2e.test.ts exercising
the real OAuth subscription path with hafizludyanto credential.
Uses full OAuth plumbing (Authorization Bearer, ?beta=true query,
claude-cli/2.1.2 user-agent, oauth-2025-04-20 beta, Claude Code
system prefix). 11 tests validate cache_creation, cache_read,
hit rate >=85%, extended-cache-ttl-2025-04-11 header, no thinking
tokens, PR anomalyco#21535 queued message regression, model lock stability.
Measured hit rate: 93.7% on claude-sonnet-4-5-20250929 (sonnet-4-6
returns 404 on current OAuth subscription tier).
- Add skill opencode-prompt-cache-maximization/CLAUDE.md documenting
the 4-breakpoint strategy, 13 snapshot Maps, 8-dimension CacheGuard,
13 CacheMonitor assertions, cache-aware compaction, and debug toggles.
- Add skill opencode-anthropic-oauth-subscription/CLAUDE.md documenting
the OAuth subscription request envelope, token refresh flow
(client_id 9d1c250a-e61b-44d9-88ed-5944d1962f5e), auth.json layout,
model availability on subscription tier, and troubleshooting matrix.
Verified:
- bun typecheck clean
- variant-picker 32/32, task-escalation 76/76 (108 total, 0 fail)
- subagent-allowlist 99/106 (7 pre-existing gpt-5.2-codex fails, unchanged)
- Cache E2E 11/11, 93.7% hit rate
- 82 supporting unit tests pass (optimizations, cache-persist-snapshots,
prompt, message-v2)
## Adaptive Thinking Guard (packages/opencode/src/session/llm.ts)
Add a defensive guard in the LLM session layer that strips
`thinking: { type: "adaptive" }` from model parameters when the selected
model does not support adaptive thinking. Non-4.6 models (opus-4-5,
sonnet-4-5, haiku-4-5, and any other non-4.6 variant) would previously
receive the adaptive thinking option through user config or agent options
and respond with "adaptive thinking is not supported on this model" errors.
The guard detects these models by checking that the model ID does not include
"4-6" and removes the adaptive thinking config before the API call is made.
- Add packages/opencode/test/provider/adaptive-thinking-guard.test.ts
(17 tests) proving the guard strips adaptive thinking for non-4.6 models
and preserves it for claude-*-4-6 models
- Add packages/opencode/test/fixture/personal-auth.ts as canonical OAuth
fixture used across all E2E provider tests
## Auth Profile Active-Switch Guard (packages/opencode/src/auth/index.ts)
Prevent Auth.set() from silently reassigning the active profile when a
non-default entry is written. The active profile now only changes when
no active profile exists yet OR when the caller explicitly passes
`options.active === true`. Previously any Auth.set() without
`active: false` would clobber the active profile, causing subagent OAuth
auth to silently swap credentials mid-session.
## Usage-Limit Error Recognition (packages/opencode/src/provider/error.ts)
Recognize additional Claude.ai Max subscription usage-limit error signatures
so the TUI surfaces the correct "reached your Claude.ai usage limit" message
(and the opencode go upsell modal) instead of a generic API error.
- Add message regex patterns: /out.of.extra.usage/i, /extra.usage.exceeded/i,
/usage_limit_exceeded/i
- isUsageLimit() now matches body.error.type === "usage_limit_exceeded" and
body.error.code === "extra_usage_exceeded" in addition to the existing
insufficient_quota / usage_not_included codes
- Upstream error mapper branches on body.error.type === "usage_limit_exceeded"
and "extra_usage_exceeded" plus body.error.code === "extra_usage_exceeded",
returning type: "usage_limit" with a friendly message pointing at
claude.ai/settings/usage
- invalid_prompt handler reuses the captured errorMsg local
## Cache Optimization E2E Migration (test/e2e/cache-optimization.test.ts)
Port cache-optimization.test.ts to the canonical personal-auth fixture used
across all E2E provider tests. Removes manual auth.json reading and raw
fetch plumbing.
- Replace getToken() + raw fetch with checkPersonalAuth / setupAuth /
teardownAuth / createOAuthFetch from test/fixture/personal-auth
- Drop OPENCODE_E2E env gate; suite uses describe.skipIf(!HAS_AUTH) so it
runs automatically when ~/.local/share/opencode-personal/opencode/auth.json
has valid OAuth credentials
- Drop oauth-2025-04-20 beta from test headers (fixture handles OAuth
envelope, Bearer token refresh, claude-cli user-agent)
- Wire Auth.get('anthropic') access/expires into createOAuthFetch
## Search-Engine Submodule Bump (cmd/opencode-search-engine)
Bump cmd/opencode-search-engine: 18a207fe → 9a5b99506
- fix(indexer): hybrid search FTS + compaction blocking runtime
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
fairyhunter13
added a commit
to fairyhunter13/opencode
that referenced
this pull request
Apr 10, 2026
… cache hit-rate maximization
Introduce SUBAGENT_PROVIDER_POLICY as a declarative per-provider allow/deny
map for subagent model selection, replacing hardcoded Anthropic branching in
resolveSubagentModel. Future providers plug in via single map entries without
touching algorithm code. Universalizes the normal algorithm path so non-policy
providers (openai, google, xai, etc.) use fair dynamic tier-based selection
without Anthropic leakage.
Provider changes:
- Add SUBAGENT_PROVIDER_POLICY with anthropic entry (allow claude-sonnet-4-6)
- Add lookupSubagentPolicy() with proxy detection for claude-* via bedrock/vertex
- Add passesPolicy() helper (deny-before-allow precedence, undefined-safe)
- Refactor resolveSubagentModel Branches 1/2/4 to use policy lookup
- Universalize normal algorithm: remove Anthropic-specific leakage from
Branch 3/4, use cost comparison instead of includes('opus') for tiered log
- Raise FLOOR_CEILINGS.critical to 10 (unlocks premium non-Anthropic models;
Anthropic path returns early via policy, unaffected)
- Fix misleading comment in resolveSubagentVariant about Perplexity suppression
- Mark SUBAGENT_MODEL_ALLOWLIST @deprecated (superseded by policy map)
- Raise maxRetries from 2 to 3 for full low->critical escalation ladder
- Surface rate-limit silent downgrade: log.warn with reason 'itpm-headroom-low'
and output note when ITPM headroom forces complexity to low
- Document clampComplexity agent-ceiling bypass (JSDoc + inline escalation comment)
- Improve Agent.get error for unknown subagent_type: list available agents
- Add packages/opencode/test/provider/variant-picker.test.ts (32 tests):
table-driven matrix: 7 variant configs x 4 complexity tiers
- Extend packages/opencode/test/tool/task-escalation.test.ts to 76 tests with
SUBAGENT_PROVIDER_POLICY coverage: passesPolicy unit tests, policy shape
assertions, resolveSubagentModel routing (opus/sonnet/haiku), proxy
detection for bedrock/vertex, openai/google/xai non-policy passthrough,
FLOOR_CEILINGS.critical=10 assertions
- Migrate 7 E2E test files to canonical personal-auth fixture for OAuth auth:
api-optimizations-e2e, cache-behavior-e2e, cache-drain-e2e,
cache-monitor-e2e, cache-monitor-lifecycle-e2e, cache-optimizations-e2e,
cache-optimizations-r1-r6-e2e
- Fix await lint warnings in api-optimizations-e2e.test.ts
Port and verification of upstream cache optimizations plus new E2E validation
through real Anthropic OAuth subscription path.
- Move git/PR guidance from bash.txt (uncached BP2 tool schema) to
anthropic.txt (cached BP1 system) -- saves ~161 bytes per request
repeated forever (ref: anomalyco#21345)
- Add E2E integration test cache-max-efficiency-e2e.test.ts exercising
the real OAuth subscription path with hafizludyanto credential.
Uses full OAuth plumbing (Authorization Bearer, ?beta=true query,
claude-cli/2.1.2 user-agent, oauth-2025-04-20 beta, Claude Code
system prefix). 11 tests validate cache_creation, cache_read,
hit rate >=85%, extended-cache-ttl-2025-04-11 header, no thinking
tokens, PR anomalyco#21535 queued message regression, model lock stability.
Measured hit rate: 93.7% on claude-sonnet-4-5-20250929 (sonnet-4-6
returns 404 on current OAuth subscription tier).
- Add skill opencode-prompt-cache-maximization/CLAUDE.md documenting
the 4-breakpoint strategy, 13 snapshot Maps, 8-dimension CacheGuard,
13 CacheMonitor assertions, cache-aware compaction, and debug toggles.
- Add skill opencode-anthropic-oauth-subscription/CLAUDE.md documenting
the OAuth subscription request envelope, token refresh flow
(client_id 9d1c250a-e61b-44d9-88ed-5944d1962f5e), auth.json layout,
model availability on subscription tier, and troubleshooting matrix.
Verified:
- bun typecheck clean
- variant-picker 32/32, task-escalation 76/76 (108 total, 0 fail)
- subagent-allowlist 99/106 (7 pre-existing gpt-5.2-codex fails, unchanged)
- Cache E2E 11/11, 93.7% hit rate
- 82 supporting unit tests pass (optimizations, cache-persist-snapshots,
prompt, message-v2)
Add a defensive guard in the LLM session layer that strips
`thinking: { type: "adaptive" }` from model parameters when the selected
model does not support adaptive thinking. Non-4.6 models (opus-4-5,
sonnet-4-5, haiku-4-5, and any other non-4.6 variant) would previously
receive the adaptive thinking option through user config or agent options
and respond with "adaptive thinking is not supported on this model" errors.
The guard detects these models by checking that the model ID does not include
"4-6" and removes the adaptive thinking config before the API call is made.
- Add packages/opencode/test/provider/adaptive-thinking-guard.test.ts
(17 tests) proving the guard strips adaptive thinking for non-4.6 models
and preserves it for claude-*-4-6 models
- Add packages/opencode/test/fixture/personal-auth.ts as canonical OAuth
fixture used across all E2E provider tests
Prevent Auth.set() from silently reassigning the active profile when a
non-default entry is written. The active profile now only changes when
no active profile exists yet OR when the caller explicitly passes
`options.active === true`. Previously any Auth.set() without
`active: false` would clobber the active profile, causing subagent OAuth
auth to silently swap credentials mid-session.
Recognize additional Claude.ai Max subscription usage-limit error signatures
so the TUI surfaces the correct "reached your Claude.ai usage limit" message
(and the opencode go upsell modal) instead of a generic API error.
- Add message regex patterns: /out.of.extra.usage/i, /extra.usage.exceeded/i,
/usage_limit_exceeded/i
- isUsageLimit() now matches body.error.type === "usage_limit_exceeded" and
body.error.code === "extra_usage_exceeded" in addition to the existing
insufficient_quota / usage_not_included codes
- Upstream error mapper branches on body.error.type === "usage_limit_exceeded"
and "extra_usage_exceeded" plus body.error.code === "extra_usage_exceeded",
returning type: "usage_limit" with a friendly message pointing at
claude.ai/settings/usage
- invalid_prompt handler reuses the captured errorMsg local
Port cache-optimization.test.ts to the canonical personal-auth fixture used
across all E2E provider tests. Removes manual auth.json reading and raw
fetch plumbing.
- Replace getToken() + raw fetch with checkPersonalAuth / setupAuth /
teardownAuth / createOAuthFetch from test/fixture/personal-auth
- Drop OPENCODE_E2E env gate; suite uses describe.skipIf(!HAS_AUTH) so it
runs automatically when ~/.local/share/opencode-personal/opencode/auth.json
has valid OAuth credentials
- Drop oauth-2025-04-20 beta from test headers (fixture handles OAuth
envelope, Bearer token refresh, claude-cli user-agent)
- Wire Auth.get('anthropic') access/expires into createOAuthFetch
Bump cmd/opencode-search-engine: 18a207fe → 9a5b99506
- fix(indexer): hybrid search FTS + compaction blocking runtime
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
fairyhunter13
added a commit
to fairyhunter13/opencode
that referenced
this pull request
Apr 10, 2026
… cache hit-rate maximization
Introduce SUBAGENT_PROVIDER_POLICY as a declarative per-provider allow/deny
map for subagent model selection, replacing hardcoded Anthropic branching in
resolveSubagentModel. Future providers plug in via single map entries without
touching algorithm code. Universalizes the normal algorithm path so non-policy
providers (openai, google, xai, etc.) use fair dynamic tier-based selection
without Anthropic leakage.
Provider changes:
- Add SUBAGENT_PROVIDER_POLICY with anthropic entry (allow claude-sonnet-4-6)
- Add lookupSubagentPolicy() with proxy detection for claude-* via bedrock/vertex
- Add passesPolicy() helper (deny-before-allow precedence, undefined-safe)
- Refactor resolveSubagentModel Branches 1/2/4 to use policy lookup
- Universalize normal algorithm: remove Anthropic-specific leakage from
Branch 3/4, use cost comparison instead of includes('opus') for tiered log
- Raise FLOOR_CEILINGS.critical to 10 (unlocks premium non-Anthropic models;
Anthropic path returns early via policy, unaffected)
- Fix misleading comment in resolveSubagentVariant about Perplexity suppression
- Mark SUBAGENT_MODEL_ALLOWLIST @deprecated (superseded by policy map)
- Raise maxRetries from 2 to 3 for full low->critical escalation ladder
- Surface rate-limit silent downgrade: log.warn with reason 'itpm-headroom-low'
and output note when ITPM headroom forces complexity to low
- Document clampComplexity agent-ceiling bypass (JSDoc + inline escalation comment)
- Improve Agent.get error for unknown subagent_type: list available agents
- Add packages/opencode/test/provider/variant-picker.test.ts (32 tests):
table-driven matrix: 7 variant configs x 4 complexity tiers
- Extend packages/opencode/test/tool/task-escalation.test.ts to 76 tests with
SUBAGENT_PROVIDER_POLICY coverage: passesPolicy unit tests, policy shape
assertions, resolveSubagentModel routing (opus/sonnet/haiku), proxy
detection for bedrock/vertex, openai/google/xai non-policy passthrough,
FLOOR_CEILINGS.critical=10 assertions
- Migrate 7 E2E test files to canonical personal-auth fixture for OAuth auth:
api-optimizations-e2e, cache-behavior-e2e, cache-drain-e2e,
cache-monitor-e2e, cache-monitor-lifecycle-e2e, cache-optimizations-e2e,
cache-optimizations-r1-r6-e2e
- Fix await lint warnings in api-optimizations-e2e.test.ts
Port and verification of upstream cache optimizations plus new E2E validation
through real Anthropic OAuth subscription path.
- Move git/PR guidance from bash.txt (uncached BP2 tool schema) to
anthropic.txt (cached BP1 system) -- saves ~161 bytes per request
repeated forever (ref: anomalyco#21345)
- Add E2E integration test cache-max-efficiency-e2e.test.ts exercising
the real OAuth subscription path with hafizludyanto credential.
Uses full OAuth plumbing (Authorization Bearer, ?beta=true query,
claude-cli/2.1.2 user-agent, oauth-2025-04-20 beta, Claude Code
system prefix). 11 tests validate cache_creation, cache_read,
hit rate >=85%, extended-cache-ttl-2025-04-11 header, no thinking
tokens, PR anomalyco#21535 queued message regression, model lock stability.
Measured hit rate: 93.7% on claude-sonnet-4-5-20250929 (sonnet-4-6
returns 404 on current OAuth subscription tier).
- Add skill opencode-prompt-cache-maximization/CLAUDE.md documenting
the 4-breakpoint strategy, 13 snapshot Maps, 8-dimension CacheGuard,
13 CacheMonitor assertions, cache-aware compaction, and debug toggles.
- Add skill opencode-anthropic-oauth-subscription/CLAUDE.md documenting
the OAuth subscription request envelope, token refresh flow
(client_id 9d1c250a-e61b-44d9-88ed-5944d1962f5e), auth.json layout,
model availability on subscription tier, and troubleshooting matrix.
Verified:
- bun typecheck clean
- variant-picker 32/32, task-escalation 76/76 (108 total, 0 fail)
- subagent-allowlist 99/106 (7 pre-existing gpt-5.2-codex fails, unchanged)
- Cache E2E 11/11, 93.7% hit rate
- 82 supporting unit tests pass (optimizations, cache-persist-snapshots,
prompt, message-v2)
Add a defensive guard in the LLM session layer that strips
`thinking: { type: "adaptive" }` from model parameters when the selected
model does not support adaptive thinking. Non-4.6 models (opus-4-5,
sonnet-4-5, haiku-4-5, and any other non-4.6 variant) would previously
receive the adaptive thinking option through user config or agent options
and respond with "adaptive thinking is not supported on this model" errors.
The guard detects these models by checking that the model ID does not include
"4-6" and removes the adaptive thinking config before the API call is made.
- Add packages/opencode/test/provider/adaptive-thinking-guard.test.ts
(17 tests) proving the guard strips adaptive thinking for non-4.6 models
and preserves it for claude-*-4-6 models
- Add packages/opencode/test/fixture/personal-auth.ts as canonical OAuth
fixture used across all E2E provider tests
Prevent Auth.set() from silently reassigning the active profile when a
non-default entry is written. The active profile now only changes when
no active profile exists yet OR when the caller explicitly passes
`options.active === true`. Previously any Auth.set() without
`active: false` would clobber the active profile, causing subagent OAuth
auth to silently swap credentials mid-session.
Recognize additional Claude.ai Max subscription usage-limit error signatures
so the TUI surfaces the correct "reached your Claude.ai usage limit" message
(and the opencode go upsell modal) instead of a generic API error.
- Add message regex patterns: /out.of.extra.usage/i, /extra.usage.exceeded/i,
/usage_limit_exceeded/i
- isUsageLimit() now matches body.error.type === "usage_limit_exceeded" and
body.error.code === "extra_usage_exceeded" in addition to the existing
insufficient_quota / usage_not_included codes
- Upstream error mapper branches on body.error.type === "usage_limit_exceeded"
and "extra_usage_exceeded" plus body.error.code === "extra_usage_exceeded",
returning type: "usage_limit" with a friendly message pointing at
claude.ai/settings/usage
- invalid_prompt handler reuses the captured errorMsg local
Port cache-optimization.test.ts to the canonical personal-auth fixture used
across all E2E provider tests. Removes manual auth.json reading and raw
fetch plumbing.
- Replace getToken() + raw fetch with checkPersonalAuth / setupAuth /
teardownAuth / createOAuthFetch from test/fixture/personal-auth
- Drop OPENCODE_E2E env gate; suite uses describe.skipIf(!HAS_AUTH) so it
runs automatically when ~/.local/share/opencode-personal/opencode/auth.json
has valid OAuth credentials
- Drop oauth-2025-04-20 beta from test headers (fixture handles OAuth
envelope, Bearer token refresh, claude-cli user-agent)
- Wire Auth.get('anthropic') access/expires into createOAuthFetch
Bump cmd/opencode-search-engine: 18a207fe → 9a5b99506
- fix(indexer): hybrid search FTS + compaction blocking runtime
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
fairyhunter13
added a commit
to fairyhunter13/opencode
that referenced
this pull request
Apr 10, 2026
… cache hit-rate maximization
Introduce SUBAGENT_PROVIDER_POLICY as a declarative per-provider allow/deny
map for subagent model selection, replacing hardcoded Anthropic branching in
resolveSubagentModel. Future providers plug in via single map entries without
touching algorithm code. Universalizes the normal algorithm path so non-policy
providers (openai, google, xai, etc.) use fair dynamic tier-based selection
without Anthropic leakage.
Provider changes:
- Add SUBAGENT_PROVIDER_POLICY with anthropic entry (allow claude-sonnet-4-6)
- Add lookupSubagentPolicy() with proxy detection for claude-* via bedrock/vertex
- Add passesPolicy() helper (deny-before-allow precedence, undefined-safe)
- Refactor resolveSubagentModel Branches 1/2/4 to use policy lookup
- Universalize normal algorithm: remove Anthropic-specific leakage from
Branch 3/4, use cost comparison instead of includes('opus') for tiered log
- Raise FLOOR_CEILINGS.critical to 10 (unlocks premium non-Anthropic models;
Anthropic path returns early via policy, unaffected)
- Fix misleading comment in resolveSubagentVariant about Perplexity suppression
- Mark SUBAGENT_MODEL_ALLOWLIST @deprecated (superseded by policy map)
- Raise maxRetries from 2 to 3 for full low->critical escalation ladder
- Surface rate-limit silent downgrade: log.warn with reason 'itpm-headroom-low'
and output note when ITPM headroom forces complexity to low
- Document clampComplexity agent-ceiling bypass (JSDoc + inline escalation comment)
- Improve Agent.get error for unknown subagent_type: list available agents
- Add packages/opencode/test/provider/variant-picker.test.ts (32 tests):
table-driven matrix: 7 variant configs x 4 complexity tiers
- Extend packages/opencode/test/tool/task-escalation.test.ts to 76 tests with
SUBAGENT_PROVIDER_POLICY coverage: passesPolicy unit tests, policy shape
assertions, resolveSubagentModel routing (opus/sonnet/haiku), proxy
detection for bedrock/vertex, openai/google/xai non-policy passthrough,
FLOOR_CEILINGS.critical=10 assertions
- Migrate 7 E2E test files to canonical personal-auth fixture for OAuth auth:
api-optimizations-e2e, cache-behavior-e2e, cache-drain-e2e,
cache-monitor-e2e, cache-monitor-lifecycle-e2e, cache-optimizations-e2e,
cache-optimizations-r1-r6-e2e
- Fix await lint warnings in api-optimizations-e2e.test.ts
Port and verification of upstream cache optimizations plus new E2E validation
through real Anthropic OAuth subscription path.
- Move git/PR guidance from bash.txt (uncached BP2 tool schema) to
anthropic.txt (cached BP1 system) -- saves ~161 bytes per request
repeated forever (ref: anomalyco#21345)
- Add E2E integration test cache-max-efficiency-e2e.test.ts exercising
the real OAuth subscription path with hafizludyanto credential.
Uses full OAuth plumbing (Authorization Bearer, ?beta=true query,
claude-cli/2.1.2 user-agent, oauth-2025-04-20 beta, Claude Code
system prefix). 11 tests validate cache_creation, cache_read,
hit rate >=85%, extended-cache-ttl-2025-04-11 header, no thinking
tokens, PR anomalyco#21535 queued message regression, model lock stability.
Measured hit rate: 93.7% on claude-sonnet-4-5-20250929 (sonnet-4-6
returns 404 on current OAuth subscription tier).
- Add skill opencode-prompt-cache-maximization/CLAUDE.md documenting
the 4-breakpoint strategy, 13 snapshot Maps, 8-dimension CacheGuard,
13 CacheMonitor assertions, cache-aware compaction, and debug toggles.
- Add skill opencode-anthropic-oauth-subscription/CLAUDE.md documenting
the OAuth subscription request envelope, token refresh flow
(client_id 9d1c250a-e61b-44d9-88ed-5944d1962f5e), auth.json layout,
model availability on subscription tier, and troubleshooting matrix.
Verified:
- bun typecheck clean
- variant-picker 32/32, task-escalation 76/76 (108 total, 0 fail)
- subagent-allowlist 99/106 (7 pre-existing gpt-5.2-codex fails, unchanged)
- Cache E2E 11/11, 93.7% hit rate
- 82 supporting unit tests pass (optimizations, cache-persist-snapshots,
prompt, message-v2)
Add a defensive guard in the LLM session layer that strips
`thinking: { type: "adaptive" }` from model parameters when the selected
model does not support adaptive thinking. Non-4.6 models (opus-4-5,
sonnet-4-5, haiku-4-5, and any other non-4.6 variant) would previously
receive the adaptive thinking option through user config or agent options
and respond with "adaptive thinking is not supported on this model" errors.
The guard detects these models by checking that the model ID does not include
"4-6" and removes the adaptive thinking config before the API call is made.
- Add packages/opencode/test/provider/adaptive-thinking-guard.test.ts
(17 tests) proving the guard strips adaptive thinking for non-4.6 models
and preserves it for claude-*-4-6 models
- Add packages/opencode/test/fixture/personal-auth.ts as canonical OAuth
fixture used across all E2E provider tests
Prevent Auth.set() from silently reassigning the active profile when a
non-default entry is written. The active profile now only changes when
no active profile exists yet OR when the caller explicitly passes
`options.active === true`. Previously any Auth.set() without
`active: false` would clobber the active profile, causing subagent OAuth
auth to silently swap credentials mid-session.
Recognize additional Claude.ai Max subscription usage-limit error signatures
so the TUI surfaces the correct "reached your Claude.ai usage limit" message
(and the opencode go upsell modal) instead of a generic API error.
- Add message regex patterns: /out.of.extra.usage/i, /extra.usage.exceeded/i,
/usage_limit_exceeded/i
- isUsageLimit() now matches body.error.type === "usage_limit_exceeded" and
body.error.code === "extra_usage_exceeded" in addition to the existing
insufficient_quota / usage_not_included codes
- Upstream error mapper branches on body.error.type === "usage_limit_exceeded"
and "extra_usage_exceeded" plus body.error.code === "extra_usage_exceeded",
returning type: "usage_limit" with a friendly message pointing at
claude.ai/settings/usage
- invalid_prompt handler reuses the captured errorMsg local
Port cache-optimization.test.ts to the canonical personal-auth fixture used
across all E2E provider tests. Removes manual auth.json reading and raw
fetch plumbing.
- Replace getToken() + raw fetch with checkPersonalAuth / setupAuth /
teardownAuth / createOAuthFetch from test/fixture/personal-auth
- Drop OPENCODE_E2E env gate; suite uses describe.skipIf(!HAS_AUTH) so it
runs automatically when ~/.local/share/opencode-personal/opencode/auth.json
has valid OAuth credentials
- Drop oauth-2025-04-20 beta from test headers (fixture handles OAuth
envelope, Bearer token refresh, claude-cli user-agent)
- Wire Auth.get('anthropic') access/expires into createOAuthFetch
Add hafizludyanto-sonnet-comparison.test.ts provider integration test
that resolves the hafizludyanto OAuth token from the opencode-personal
auth.json and compares Anthropic /v1/messages responses for
claude-sonnet-4-6 with and without the ?beta=true query flag.
Auto-skips via describe.skipIf when no OAuth token is present, so
the suite is a no-op on CI. Used to diagnose usage_limit_exceeded
behavior on the Extra Usage subscription path.
Bump cmd/opencode-search-engine: 18a207fe → 9a5b99506
- fix(indexer): hybrid search FTS + compaction blocking runtime
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
fairyhunter13
added a commit
to fairyhunter13/opencode
that referenced
this pull request
Apr 10, 2026
… cache hit-rate maximization
Introduce SUBAGENT_PROVIDER_POLICY as a declarative per-provider allow/deny
map for subagent model selection, replacing hardcoded Anthropic branching in
resolveSubagentModel. Future providers plug in via single map entries without
touching algorithm code. Universalizes the normal algorithm path so non-policy
providers (openai, google, xai, etc.) use fair dynamic tier-based selection
without Anthropic leakage.
Provider changes:
- Add SUBAGENT_PROVIDER_POLICY with anthropic entry (allow claude-sonnet-4-6)
- Add lookupSubagentPolicy() with proxy detection for claude-* via bedrock/vertex
- Add passesPolicy() helper (deny-before-allow precedence, undefined-safe)
- Refactor resolveSubagentModel Branches 1/2/4 to use policy lookup
- Universalize normal algorithm: remove Anthropic-specific leakage from
Branch 3/4, use cost comparison instead of includes('opus') for tiered log
- Raise FLOOR_CEILINGS.critical to 10 (unlocks premium non-Anthropic models;
Anthropic path returns early via policy, unaffected)
- Fix misleading comment in resolveSubagentVariant about Perplexity suppression
- Mark SUBAGENT_MODEL_ALLOWLIST @deprecated (superseded by policy map)
- Raise maxRetries from 2 to 3 for full low->critical escalation ladder
- Surface rate-limit silent downgrade: log.warn with reason 'itpm-headroom-low'
and output note when ITPM headroom forces complexity to low
- Document clampComplexity agent-ceiling bypass (JSDoc + inline escalation comment)
- Improve Agent.get error for unknown subagent_type: list available agents
- Add packages/opencode/test/provider/variant-picker.test.ts (32 tests):
table-driven matrix: 7 variant configs x 4 complexity tiers
- Extend packages/opencode/test/tool/task-escalation.test.ts to 76 tests with
SUBAGENT_PROVIDER_POLICY coverage: passesPolicy unit tests, policy shape
assertions, resolveSubagentModel routing (opus/sonnet/haiku), proxy
detection for bedrock/vertex, openai/google/xai non-policy passthrough,
FLOOR_CEILINGS.critical=10 assertions
- Migrate 7 E2E test files to canonical personal-auth fixture for OAuth auth:
api-optimizations-e2e, cache-behavior-e2e, cache-drain-e2e,
cache-monitor-e2e, cache-monitor-lifecycle-e2e, cache-optimizations-e2e,
cache-optimizations-r1-r6-e2e
- Fix await lint warnings in api-optimizations-e2e.test.ts
Port and verification of upstream cache optimizations plus new E2E validation
through real Anthropic OAuth subscription path.
- Move git/PR guidance from bash.txt (uncached BP2 tool schema) to
anthropic.txt (cached BP1 system) -- saves ~161 bytes per request
repeated forever (ref: anomalyco#21345)
- Add E2E integration test cache-max-efficiency-e2e.test.ts exercising
the real OAuth subscription path with hafizludyanto credential.
Uses full OAuth plumbing (Authorization Bearer, ?beta=true query,
claude-cli/2.1.2 user-agent, oauth-2025-04-20 beta, Claude Code
system prefix). 11 tests validate cache_creation, cache_read,
hit rate >=85%, extended-cache-ttl-2025-04-11 header, no thinking
tokens, PR anomalyco#21535 queued message regression, model lock stability.
Measured hit rate: 93.7% on claude-sonnet-4-5-20250929 (sonnet-4-6
returns 404 on current OAuth subscription tier).
- Add skill opencode-prompt-cache-maximization/CLAUDE.md documenting
the 4-breakpoint strategy, 13 snapshot Maps, 8-dimension CacheGuard,
13 CacheMonitor assertions, cache-aware compaction, and debug toggles.
- Add skill opencode-anthropic-oauth-subscription/CLAUDE.md documenting
the OAuth subscription request envelope, token refresh flow
(client_id 9d1c250a-e61b-44d9-88ed-5944d1962f5e), auth.json layout,
model availability on subscription tier, and troubleshooting matrix.
Verified:
- bun typecheck clean
- variant-picker 32/32, task-escalation 76/76 (108 total, 0 fail)
- subagent-allowlist 99/106 (7 pre-existing gpt-5.2-codex fails, unchanged)
- Cache E2E 11/11, 93.7% hit rate
- 82 supporting unit tests pass (optimizations, cache-persist-snapshots,
prompt, message-v2)
Add a defensive guard in the LLM session layer that strips
`thinking: { type: "adaptive" }` from model parameters when the selected
model does not support adaptive thinking. Non-4.6 models (opus-4-5,
sonnet-4-5, haiku-4-5, and any other non-4.6 variant) would previously
receive the adaptive thinking option through user config or agent options
and respond with "adaptive thinking is not supported on this model" errors.
The guard detects these models by checking that the model ID does not include
"4-6" and removes the adaptive thinking config before the API call is made.
- Add packages/opencode/test/provider/adaptive-thinking-guard.test.ts
(17 tests) proving the guard strips adaptive thinking for non-4.6 models
and preserves it for claude-*-4-6 models
- Add packages/opencode/test/fixture/personal-auth.ts as canonical OAuth
fixture used across all E2E provider tests
Prevent Auth.set() from silently reassigning the active profile when a
non-default entry is written. The active profile now only changes when
no active profile exists yet OR when the caller explicitly passes
`options.active === true`. Previously any Auth.set() without
`active: false` would clobber the active profile, causing subagent OAuth
auth to silently swap credentials mid-session.
Recognize additional Claude.ai Max subscription usage-limit error signatures
so the TUI surfaces the correct "reached your Claude.ai usage limit" message
(and the opencode go upsell modal) instead of a generic API error.
- Add message regex patterns: /out.of.extra.usage/i, /extra.usage.exceeded/i,
/usage_limit_exceeded/i
- isUsageLimit() now matches body.error.type === "usage_limit_exceeded" and
body.error.code === "extra_usage_exceeded" in addition to the existing
insufficient_quota / usage_not_included codes
- Upstream error mapper branches on body.error.type === "usage_limit_exceeded"
and "extra_usage_exceeded" plus body.error.code === "extra_usage_exceeded",
returning type: "usage_limit" with a friendly message pointing at
claude.ai/settings/usage
- invalid_prompt handler reuses the captured errorMsg local
Port cache-optimization.test.ts to the canonical personal-auth fixture used
across all E2E provider tests. Removes manual auth.json reading and raw
fetch plumbing.
- Replace getToken() + raw fetch with checkPersonalAuth / setupAuth /
teardownAuth / createOAuthFetch from test/fixture/personal-auth
- Drop OPENCODE_E2E env gate; suite uses describe.skipIf(!HAS_AUTH) so it
runs automatically when ~/.local/share/opencode-personal/opencode/auth.json
has valid OAuth credentials
- Drop oauth-2025-04-20 beta from test headers (fixture handles OAuth
envelope, Bearer token refresh, claude-cli user-agent)
- Wire Auth.get('anthropic') access/expires into createOAuthFetch
Add hafizludyanto-sonnet-comparison.test.ts provider integration test
that resolves the hafizludyanto OAuth token from the opencode-personal
auth.json and compares Anthropic /v1/messages responses for
claude-sonnet-4-6 with and without the ?beta=true query flag.
Auto-skips via describe.skipIf when no OAuth token is present, so
the suite is a no-op on CI. Used to diagnose usage_limit_exceeded
behavior on the Extra Usage subscription path.
Bump cmd/opencode-search-engine: 18a207fe → 9a5b99506
- fix(indexer): hybrid search FTS + compaction blocking runtime
Fix ReferenceError: revert is not defined in the session TUI route.
Two `const revert` declarations coexisted in the same component body:
an inner `const revert = session()?.revert?.messageID` inside an
onSelect callback (line 574) shadowed the top-level
`const revert = createMemo(...)` (line 1070). SolidJS's babel-plugin-solid
scope analysis captured the wrong binding in JSX memo closures, producing
a TDZ-style ReferenceError at runtime whenever the revert UI rendered.
- Rename top-level createMemo from `revert` to `revertState`
- Update 6 JSX usages (lines 1550, 1584, 1589, 1591, 1611x2) to revertState()
- Inner callback-local `const revert` preserved (no shadow conflict)
Regenerate packages/sdk/js/src/v2/gen/{sdk,types}.gen.ts to fix
pre-existing TS2305/TS2724 errors in @opencode-ai/plugin typecheck
(sdk.gen.ts imported AuthListResponses, EventTuiSessionNew,
IndexHealthResponses and others that no longer existed in types.gen.ts).
Add extractLocalDefs() pass in packages/sdk/js/script/build.ts: Zod 4's
z.lazy() emits $defs with __schemaX keys whose $refs point to
#/components/schemas/__schemaX (absolute) while the $defs themselves
are schema-local. Lift all local $defs into components.schemas before
createClient() so hey-api can resolve them. Without this the build
fails with: Missing $ref pointer "#/components/schemas/__schema0".
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
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
1 participant
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.
Issue for this PR
Fixes #21518
Type of change
What does this PR do?
prompt.ts:1483-1499wrapped queued user message text in<system-reminder>tags via in-memory mutation (p.text = [...]). Never persisted to DB. Each turn re-fetches from DB viahydrate()(JSON.parse) — gets original unwrapped text.stepresets to 0 perrunLoopcall, sostep > 1is false on first step of new turn. Same message appears unwrapped → provider prompt cache miss.Bug confirmed in v1.3.17, v1.4.0, and current
devHEAD — identical code at all three.Fix: Remove in-memory mutation from
prompt.ts. Addremindoption totoModelMessagesinmessage-v2.ts. Whenremind(=lastFinished?.id) is set, user text parts in messages with id > remind get wrapped during serialization. No mutation of stored parts. Same inputs = same output always = cache consistent.How did you verify your code works?
toModelMessageswithoutremind(old step 1) then withremind(old step 2+) — confirmed serialization differs (the bug). With fix applied (remindalways passed), output identical across calls.bun test test/session/)Screenshots / recordings
N/A — logic change, no UI.
Checklist