feat(sisyphus): add GLM-5.x dedicated prompt builder and speed overlay#3736
Closed
islee23520 wants to merge 35 commits intocode-yeongyu:devfrom
Closed
feat(sisyphus): add GLM-5.x dedicated prompt builder and speed overlay#3736islee23520 wants to merge 35 commits intocode-yeongyu:devfrom
islee23520 wants to merge 35 commits intocode-yeongyu:devfrom
Conversation
- Add isGlmSisyphusHarnessModel for GLM-5/5.1/5-turbo detection - Route GLM harness models to specialized prompts in Sisyphus agents - Add Small Context Working Memory with state slices for GLM context optimization - Add GLM-specific context priorities and vision constraints for Sisyphus Junior - Add comprehensive tests for GLM prompt validation and routing
- Refactor sisyphus/glm.ts: 387→69 lines via overlay pattern (string replacement)
- Add isGlmThinkingModel() for GLM-5+ text models (excludes VLM)
- Add isGlmVisionModel() for GLM VLM variants (glm-4.6v, glm-5v-turbo)
- Oracle/Metis/Momus: GLM-5+ text → thinking: { type: enabled }, Claude → budgetTokens
- Sisyphus-Junior: GLM → thinking: { type: enabled } (was bare base)
- Sisyphus: GLM overlay injection + thinking config, fact-checked comments
- Update stale test: sisyphus-junior GLM now returns thinking
- Add 100-test factory benchmark (5 agents × 7 GLM variants + cross-agent guards)
- Add runtime benchmark script (scripts/benchmark-glm-thinking.ts)
Benchmark: 100 factory tests pass, 452 agent tests pass, typecheck clean
Verified: No GLM text model receives budgetTokens across any agent
Refs: code-yeongyu#3210, code-yeongyu#3256, code-yeongyu#3568
islee23520
changed the title
…rlay - New src/agents/sisyphus/glm-prompt.ts: 8-block GLM-specific Sisyphus prompt (DISPATCH→DELEGATE→COLLECT→SYNTHESIZE→DONE execution loop replacing EXPLORE→PLAN→ROUTE→EXECUTE→VERIFY→RETRY→DONE) - New src/agents/glm-prompt-quality.test.ts: 32 quality benchmarks across Instruction Compliance (10), Speed (10), Accuracy (9), Cross-Agent (3) - Extended src/agents/sisyphus-junior/glm.ts: SJ speed overlay with execution-first mindset, brief thinking, re-entry rule, exploration budget (2-iteration cap), tiered verification V1/V2/V3, token economy - Modified src/agents/sisyphus.ts: GLM routing from overlay string.replace to dedicated buildGlmSisyphusPrompt() builder (matches Kimi K2.x pattern) GLM-5.x does not support budgetTokens. Excessive thinking was controlled via prompt engineering: concise thinking mandate, re-entry rule (suppress re-verbalization for resolved turns), exploration budget hard stops, and tiered verification (V1/V2/V3) to avoid over-verification on trivial changes. Hephaestus delegation strategy included: sequential edits >= 3 automatically routed to Hephaestus (deep-thinking worker) to keep Sisyphus unblocked. All 140 GLM-related tests pass. Typecheck clean. AI slop removed.
islee23520
changed the title
![cubic-dev-ai[bot]](https://avatars.githubusercontent.com/in/1082092?s=60&v=4){kind=small}
There was a problem hiding this comment.
No issues found across 18 files
Confidence score: 5/5
- Automated review surfaced no issues in the provided summaries.
- No files require special attention.
Auto-approved: GLM-specific changes are strictly isolated using model-specific predicates. Non-GLM logic and configurations are preserved, verified by 140 tests including 32 new quality benchmarks.
- Add buildGlmSubagentVisionBlock() for concise subagent vision warnings - Apply vision constraint to Oracle, Metis, Momus (GLM branches) - Simplify GLM SJ speed overlay prompt (remove 4 redundant lines) - Remove redundant JSDoc from metis.ts, marketing language from sisyphus.ts - Centralize Sisyphus description as SISYPHUS_DESCRIPTION constant
There was a problem hiding this comment.
No issues found across 18 files
Confidence score: 5/5
- Automated review surfaced no issues in the provided summaries.
- No files require special attention.
Requires human review: Large PR (1800+ lines) modifies Sisyphus metadata descriptions globally, not just for GLM models. 45 failing tests reported, and complex model-routing logic changes require manual verification.
|
Looking forward to try these changes |
…unior prompt test
The SJ GLM prompt intentionally omits the .sisyphus/state/{plan-or-session}/
path and individual slice filenames (goal.md, decisions.md, etc.) that the
main Sisyphus GLM prompt includes. The test incorrectly expected the full
ledger path; align it with the lightweight memory contract.
# Conflicts: # src/agents/momus.ts
…r GLM Upstream removed GLM-specific thinking config from Momus, causing budgetTokens: 32000 to be applied to GLM models that do not support it. Restore the isGlmThinkingModel branch matching Metis and Oracle. Also add test coverage for: - Momus GLM thinking config without budgetTokens - Sisyphus call_omo_agent permission (allow) vs Hephaestus (deny)
There was a problem hiding this comment.
1 issue found across 5 files (changes from recent commits).
Prompt for AI agents (unresolved issues)
Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.
<file name="src/plugin-handlers/tool-config-handler.ts">
<violation number="1">
P1: Sisyphus’s `call_omo_agent` permission was changed from allow to deny, which can break its ability to delegate to other agents.</violation>
</file>
Tip: Review your code locally with the cubic CLI to iterate faster.
…="deep") delegation
Restore architecture and routing intent comments in types.ts, oracle.ts, and sisyphus/glm.ts that document design decisions.
There was a problem hiding this comment.
0 issues found across 3 files (changes from recent commits).
Requires human review: PR adds 2544 lines across 12+ files (new agents, prompt builders, config changes). Despite clean AI review and tests, the strict '100% sure no regressions' criteria cannot be met without manual review
…J, sisyphus Restore architecture intent comments: - metis.ts: agent role/responsibilities JSDoc - sisyphus-junior/agent.ts: routing order, BLOCKED_TOOLS intent - sisyphus.ts: Gemini overlay placement rationale, GLM thinking note
There was a problem hiding this comment.
0 issues found across 3 files (changes from recent commits).
Requires human review: Cannot be 100% sure of zero regressions: modifies core orchestration agents (sisyphus, metis, oracle, momus, sisyphus-junior) with new GLM routing logic, which could introduce edge-case misrouting or
… prompt The GLM_SJ_Speed_Optimizations section duplicated content already present in the base SJ prompt and Sisyphus system prompt. Only the GLM-specific context priorities and vision constraint are kept.
delegation-scorecard and event-metric-collector are test-only utilities with zero runtime consumers. Only referenced by scripts/benchmark-* Move to scripts/ if needed later.
|
You're iterating quickly on this pull request. To help protect your rate limits, cubic has paused automatic reviews on new pushes for now—when you're ready for another review, comment |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
GLM-5.x models overthink during Sisyphus orchestration. Since
budgetTokensis not supported, thinking restraint must come from prompt structure. This PR introduces a dedicated GLM prompt system that enforces a fast dispatch-first execution loop and prevents text-only models from attempting image analysis.Changes
glm-prompt.ts) with 8-block structure enforcing DISPATCH→DELEGATE→COLLECT→SYNTHESIZE→DONEVerification
bun run typecheck— cleanbun run build— cleanbun test— 6026 pass (45 pre-existing failures in tmux/background-agent, unrelated)Checklist
bun run typecheckpassesbun run buildsucceedspackage.json