some recent changes#251
Merged
Merged
Conversation
- buildRalphSynthesisPrompt() accepts optional seedGoal param (AC-01): when provided, injects the pre-existing ## Goal block with an authoritative-preserve instruction so the model keeps all [decision] tags and constraints verbatim and only expands missing slots. - Improved RALPH_SYNTHESIS_PROMPT_BASE (AC-03): explicitly instructs the model to reconstruct decisions ([decision] tags), constraints, ACs with DoD bullets, and assumptions — no information from the conversation should be omitted. - ralph-promote-routes.ts (AC-02): inspects the last assistant conversationTurn for a ## Goal block via regex; if found, extracts the block and passes it as seedGoal to buildRalphSynthesisPrompt(). - Tests: updated synthesis-prompt.test.ts (replaced stale snapshot assertions, added seedGoal coverage); added three new route tests for seed detection (match, no-match, last-turn-only). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Keep schedule run records active after successful enqueue until the queued task reaches a terminal state. Add regression coverage for queue success, queue failure, and enqueue failure lifecycle boundaries. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Keep overlapped schedule timer fires from enqueueing duplicate work by recording a missed run and deferring the next timer until the active run completes. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
After commit f56a824a8, ScheduleExecutor.executeRun() became async and awaits task completion before returning. In test environments with no AI executor, queued tasks fail immediately, so POST /schedules/:id/run now returns a run with status 'failed'. Update the three affected integration test assertions to also accept 'failed' alongside 'running'/'completed': - schedule-concurrent.test.ts - schedule-handler.test.ts - schedule-pause-markers.test.ts Also fix a stale comment in schedule-mutation-during-run.test.ts that incorrectly described executeRun() as synchronous. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Use a validation-only system prompt for Ralph final-check tasks while keeping autopilot execution enabled. This prevents the final-check phase from receiving normal implementation-loop instructions that tell it to edit code and commit. Add regression coverage for final-check system-message construction and executor dispatch. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Render completed ask_user calls as read-only history cards in Activity, keep them visible outside whisper collapse, and improve the fallback tool summary for batched questions. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Adds a compact "Open PR" input row at the top of the Pull Requests tab queue panel. Bare numeric input resolves against the active workspace repo; full GitHub PR URLs are parsed and matched against registered workspaces by repository remote URL, then opened in the existing internal PR detail/conversation view (#repos/<repoId>/pull-requests/<n>/overview). Validation calls the existing single-PR API before navigating, so closed, merged, or non-listed PRs open while truly missing PRs surface an inline error. Unmatched URL repositories produce a "repository not registered" inline error with no auto-registration and no external navigation. No new persistent storage, feature flag, or top-level per-repo data path is introduced. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
- Add gitCommitLookup feature flag (disabled by default) wired through CLIConfig -> ResolvedCLIConfig -> RuntimeDashboardConfig -> SPA config - Add isGitCommitLookupEnabled() helper in SPA utils/config.ts - RepoGitTab: add handleCommitLookup callback that validates SHA pattern, checks loaded list first, then falls back to getCommit API for misses - RepoGitTab: search input shows '↵ open commit' hint and loading state when query looks like a SHA and feature is enabled; Enter triggers lookup - RepoGitTab: deep-link useEffect attempts getCommit API lookup when commit is not in the loaded history page (feature-gated) - RepoGitTab: render 'Opened commit' section above CommitList for directly opened commits not in the loaded page; calls handleSelect on click - RepoGitTab: show inline commitLookupError near search bar on failure - Add namespace-registry merge for gitCommitLookup feature flag - Add 21 new tests covering all new state, callbacks, UI elements, and safety constraints (no state-changing git commands in lookup path) Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Refresh all 15 outdated specs under packages/coc/specs/ to reflect the current dashboard SPA, including the Admin shell redesign (embedded Memory/Skills/Logs/Usage&Costs/Models/Servers panels), the Workflows tab unification (Workflows + Templates + AI Chat Templates + Prompt & Script Templates), the deprecated Plans/Tasks tab, the redesigned PR queue, and the regrouped Repository Settings sidebar. Bumps each affected spec to a new minor or major version with revision-history entries. Co-authored-by: Cursor <cursoragent@cursor.com>
- Create packages/forge/resources/bundled-skills/ultra-ralph/SKILL.md with five sections (grill, synthesis, execution, iteration, final-check), each section delimited by '## Section: <name>' headings (greppable) - Register ultra-ralph in bundled-skills-registry.ts - Add ultra-ralph to DEFAULT_BUNDLED_SKILLS in coc/src/config.ts for auto-install at server startup (never clobbers user edits) - Thin the five Ralph prompt call sites to a skill reference + minimal machine-contract lines only: - RALPH_GRILL_SUFFIX (chat-base-executor.ts) - RALPH_SYNTHESIS_PROMPT_BASE (synthesis-prompt.ts) - RALPH_BASE_INSTRUCTIONS (ralph-executor.ts) - RALPH_SPEC_CONTRACT_PROMPT (iteration-prompt.ts) - RALPH_FINAL_CHECK_BASE_INSTRUCTIONS + READ_ONLY_INSTRUCTIONS - Delete ralph-prompt-overrides.ts; replace with admin-prompt-overrides.ts (no RALPH_PROMPT_IDS) for diff classification override still in use - Remove getPromptOverride calls from all five Ralph prompt sites - Remove /api/admin/prompts Ralph entries from getBuiltInPrompts() - Remove Ralph from PromptsPanel GROUP_ORDER in admin SPA - Add Vitest test for ultra-ralph SKILL.md presence and section checks - Update all affected test files Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Running and queued tasks were always shown with the Copilot (green) color in ChatListPane regardless of their actual provider. The root cause was that serializeTaskSummary stripped payload.provider when building the slim payload for queue-list API responses. ChatListPane resolves colors via getTaskChatProvider, which checks task.payload.provider among other paths; with it absent all tasks fell through to the Copilot green fallback. Adds payload.provider to slimPayload so Codex (indigo) and Claude (coral) sessions are colored correctly in the running/queued sidebar list. Also adds four regression tests covering all three provider values and the undefined (no provider) case. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Ensures Ralph follow-on tasks (next iteration, final-check, gap-fix loops) are admitted ahead of unrelated exclusive backlog in the same workspace, preventing session interruption. Core mechanism: add continuationOfSessionId to QueuedTask/CreateTaskInput and a new insertAsContinuation() method in TaskQueueManager that scans the queue for the first exclusive task not belonging to the same session and inserts the continuation before it. Changes: - packages/forge/src/queue/types.ts: add continuationOfSessionId? to QueuedTask - packages/forge/src/queue/task-queue-manager.ts: add insertAsContinuation(), call it from insertTask() when task.continuationOfSessionId is set - packages/coc/src/server/ralph/types.ts: extend RalphFinalCheckStatus with 'queued' - packages/coc/src/server/ralph/enqueue-iteration.ts: thread continuationOfSessionId - packages/coc/src/server/ralph/enqueue-final-check.ts: set continuationOfSessionId on final-check payload; buildFinalCheckStartRecord returns status: 'queued' (AC-06) - packages/coc/src/server/ralph/orchestrate-final-check.ts: set continuationOfSessionId on gap-fix loop task (AC-03) - packages/coc/src/server/queue/queue-executor-bridge.ts: set continuationOfSessionId on next-iteration enqueue (AC-01) Tests: - forge/test/queue/task-queue-manager.test.ts: 6 new continuation ordering tests - coc/test/server/ralph/enqueue-final-check.test.ts: expect 'queued' start status, assert continuationOfSessionId is set - coc/test/server/queue-executor-bridge.test.ts: Ralph session queue continuity describe block with AC-01 (RALPH_NEXT) and AC-02 (RALPH_COMPLETE final-check) tests AC-04 (workspace-scoped) is guaranteed by per-repo TaskQueueManager instances. AC-05 (non-exclusive concurrency) is unaffected. AC-07 (scheduled Ralph semantics) verified via existing schedule tests (350 pass). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…imization Implements a simplified SkillOpt loop (arXiv:2605.23904) under scripts/skillopt/ that trains a SKILL.md document by driving the GitHub Copilot CLI non-interactively. AC-01 Copilot CLI driver (cli-driver.ts) - runCopilotCli(prompt, workdir, model) with --allow-all-tools - Typed CliError for non-zero exits; configurable timeout - captureGitDiff for post-run diff capture AC-02 Task corpus + schema (corpus.ts + corpus/tasks.json) - Documented Task schema (id, prompt, seedRef, visibleTests, hiddenTests, judgeRubric, split) - loadCorpus validates schema, split presence, unique IDs - 5-task seed corpus (3 train, 2 selection) AC-03 Isolated rollout (rollout.ts) - git worktree add --detach + skill injection - Injects skill as .github/skills/active-skill.md - Runs hidden tests before worktree cleanup - Always cleans up in finally block AC-04 Scoring (scoring.ts) - blendScores(hidden, judge, weights) -- pure, testable - Default weights w1=0.7 / w2=0.3 (configurable) - LLM judge via headless CLI call on the diff - Hidden tests NEVER placed in agent-visible prompt AC-05 Optimizer edit (optimizer.ts) - buildOptimizerPrompt from current skill + scored rollouts - parseOptimizerEdit: JSON block -> add/delete/replace - applyEdit: anchor-based bounded edit applicator - Malformed output = no-op candidate (run continues) AC-06 Held-out validation gate (gate.ts) - evaluateGate: accept iff candidateScore > bestScore (strict) - Records decision + scores for history AC-07 Loop + artifacts (loop.ts) - Bounded loop (max-steps), Ctrl-C safe SIGINT handler - Atomic writes (tmp -> rename) for best_skill.md - Appends step records to history.jsonl - Writes summary.json on completion AC-08 CLI ergonomics (skillopt.ts + README.md) - parseArgs interface with --help, --skill, --corpus, --out, --target-model, --optimizer-model, --max-steps, --w1, --w2, --timeout-ms - Pre-flight check: fails fast if copilot not on PATH - README with schema docs, optimizer prompt contract, algorithm Tests: 62 vitest tests; all CLI calls mocked for CI Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Extract shared owner-bridge dispatch in MultiRepoQueueRouter and pass reasoningEffort through follow-up routing. Add regression coverage for follow-up, steering, and ask-user dispatch across bridges. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
ClaudeSDKService only passed cwd to the agent-sdk query, so the Claude agent was sandboxed to the working directory and could not read out-of-repo files such as CoC skill definitions under ~/.coc/skills. Wire an additionalDirectories option through SendMessageOptions into the SDK query options. ClaudeSDKService.resolveAdditionalDirectories always grants access to ~/.coc and the system temp directory, merges any caller-provided directories, resolves them to absolute paths, and de-duplicates (case-insensitively on Windows). Add unit tests covering the auto-injected directories, caller-provided entries, and de-duplication. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
- Add deriveEffort() pure function (effortUtils.ts) that computes the effort override from a model's stored preference, supported efforts, and reasoning capability flag. - Add useProviderReasoningEfforts(provider) hook that fetches the per-provider, per-model effort preference map from GET /api/agent-providers/<provider>/models/reasoning-efforts. - Update NewChatArea to auto-derive effortOverride whenever provider or effective model changes. Uses userPickedForModelRef to preserve explicit user picks within the same (provider, model) combo while re-deriving on any model/provider swap. Draft restore no longer restores effortOverride; always re-derives from current preferences. - Update ChatDetail to initialize effortOverride from processDetails.config.reasoningEffort on first task load (§5.1) and re-derive on mid-conversation model override change (§5.4). Resets on taskId change. Uses chatEffectiveModelId so effort options and disabled state reflect the active override, not only the original session model. - Tests: deriveEffort unit tests (13), useProviderReasoningEfforts hook tests (7), source-level wiring tests for NewChatArea and ChatDetail (23). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…dled skill The Ralph prompt entries were removed from getBuiltInPrompts() when the five Ralph prompt call sites were thinned to reference the bundled ultra-ralph skill. Update the stale admin-prompts test to expect the remaining 8 built-in prompts and drop the Ralph group assertions. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…effort hook - chat-mode-executors: assert grilling suffix references the bundled ultra-ralph skill + '## Goal' contract instead of the removed inline directive text. - config snapshot: accept new feature-flag key 'gitCommitLookup'. - ChatDetail source-introspection: widen the taskId-reset window now that the effect resets additional refs. - NewChatArea (x2) and composer ordering: add getReasoningEfforts to the agentProviders client mock now that ChatDetail/NewChatArea call useProviderReasoningEfforts on mount. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.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.
No description provided.