feat(SYMPH-71): teach merge agent about GitHub merge queue#50
Closed
ericlitman wants to merge 95 commits intoOasAIStudio:mainfrom
Closed
Conversation
Three extensions to symphony-ts for multi-model autonomous pipeline: 1. Runner abstraction (Task 5.1): Extract runner interface from Codex client, add ClaudeCodeRunner and GeminiRunner via Vercel AI SDK providers. Per-state runner selection via YAML config. 2. State machine (Task 5.2): Multi-stage workflows with typed transitions (agent/gate/terminal stages), per-stage config overrides, rework loops with configurable limits. Backward compatible — no stages = flat dispatch. 3. Ensemble gate (Task 5.3): Gate stages spawn parallel review agents, collect two-layer verdicts (JSON gate + plain text feedback), aggregate results, post to Linear. Supports human and automated gates. 77 new tests (217 total), typecheck clean.
Pipeline configuration for the founder's autonomous dev pipeline: - WORKFLOW.md with YAML frontmatter: Linear tracker, 5-stage state machine (investigate → implement → review → merge → done), per-stage runner/model overrides, ensemble gate with Codex + Gemini reviewers - 6 LiquidJS prompt templates: global clauses (headless mode, scope discipline, design references, verify lines, $BASE_URL), investigate, implement, review-adversarial, review-security, merge - Hook scripts: after-create (git clone + install), before-run (fetch + rebase) - validate.sh: config validation (YAML parsing, file checks, stage flow)
- Gemini runner: replace static import with lazy dynamic import() for ESM-only ai-sdk-provider-gemini-cli (require() returns empty module) - Claude Code runner: add model ID mapping (claude-sonnet-4-5 → sonnet) so YAML config can use standard Anthropic model names - Claude Code runner: add AbortController to generateText() calls for subprocess cleanup on close() - Add @ai-sdk/provider + transitive deps to package.json - Add integration smoke test (skipped by default, RUN_INTEGRATION=1) - 9 new unit tests for model mapping, abort, and provider behavior Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Claude Code runner was spawning without bypassPermissions, causing the agent to stall waiting for interactive permission approval in headless mode. Adding permissionMode: "bypassPermissions" fixes E2E dispatch. Also adds WORKFLOW-flat.md for flat (no state machine) E2E testing. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…pace cleanup Three blocking issues prevented multi-stage pipelines from completing: 1. Stage transitions never fired: advanceStage only runs in onWorkerExit after ALL turns complete. Added [STAGE_COMPLETE] sentinel detection in the turn loop for early exit when agents signal stage completion. 2. Continuation turns lost stage context: buildContinuationPrompt had no stageName parameter. Added stage-aware continuation prompts with per-stage constraints (investigate/implement/merge). 3. Stale workspaces from prior runs: afterCreate hook only fires on new directories. Added workspace cleanup on fresh dispatch (attempt=null) before createForIssue. Also includes: stage-aware beforeRun hook (skip rebase on feature branches), runner/model overrides from stage config, Linear comment posting for investigation notes, and WORKFLOW-staged.md with full stage definitions. 234 tests passing, build clean. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Use trimEnd().endsWith() instead of includes() to prevent false early exit when an agent mentions [STAGE_COMPLETE] conversationally rather than as the final output signal. Found by: Gemini (P1 → triaged as P2) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Two P1 findings from Codex R2: 1. Workspace cleanup now only fires on initial stage (investigate) of staged pipelines. Flat dispatch and non-initial stages preserve existing workspaces, preventing data loss after service restarts. 2. Gate stages now claim the issue before firing ensemble review, preventing duplicate gate dispatch on subsequent poll ticks. Gate handler errors release the claim for retry. Found by: Codex (2 P1s) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…L thresholds Reviewers were operating blind (only saw issue metadata, no code). Now gate-handler fetches `git diff origin/main...HEAD` and includes it in the reviewer prompt. Reviewer `prompt` field renders as inline Review Focus instructions. Both Gemini reviewers in WORKFLOW-staged.md now have explicit PASS/FAIL criteria to prevent overly strict rework loops. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…ture failures Gemini rate limits (429) were treated as code review FAILs, causing infinite rework loops even when reviewers that ran approved the code. Changes: - runSingleReviewer retries up to 3 times with exponential backoff - Infrastructure failures return verdict "error" instead of "fail" - aggregateVerdicts ignores "error" results (only counts real pass/fail) - All-error still returns FAIL (can't skip review entirely) - formatGateComment shows "ERROR" label for infrastructure failures Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
fix: stage machine execution — early exit, stage-aware prompts, workspace cleanup
B1: parseReviewerOutput now detects rate-limit text in 200 responses (e.g. "You have exhausted your capacity") and returns verdict "error" instead of "fail", preventing false rework loops. B4: handleEnsembleGate now posts a Linear comment when max rework attempts are exceeded, so the issue doesn't silently stall. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Adversarial review P2: empty catch {} in escalation path swallowed
errors silently. Now logs console.warn with issue identifier and error.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
fix: rate-limit text detection + escalation comment for ensemble gate
Add linear_state field to stage definitions and escalation_state to top-level config. The orchestrator now updates Linear issue states on stage dispatch (In Progress), gate entry (In Review), and escalation (Blocked). Includes WORKFLOW-staged.md config and active_states fix for In Review. 253 tests, typecheck clean. E2E validated: MOB-28 Todo→In Progress→In Review→Done. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…#4) * feat: workpad system — structured progress tracking on Linear tickets Phase 11: Add structured workpad comments to Linear issues with sync_workpad dynamic tool for token-efficient updates, fileUpload media flow, and stage-specific workpad behavior (investigate creates, implement updates, merge finalizes). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: R1 adversarial review — 1 P2 + test coverage gaps - Check commentUpdate.success in response (Codex finding) - Add 3 tests: missing comment field, empty id, update success=false (Sonnet finding) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* feat: failure signal parsing — route agent failures by class (verify/review/spec/infra)
* fix: R1 adversarial review — escalation side effects + state corruption guard + empty message fallback
* fix: R1 adversarial review — 2 P1s + 1 P2
P1: Persist spec-failure escalations to tracker (updateIssueState + postComment)
P1: Persist review-escalation side effects when max rework exceeded
P2: Add test verifying reworkCount threading to spawnWorker
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: R2 adversarial review — prevent redispatch of escalated issues
* feat: max retry safety net + review rework routing (Wave 3)
- scheduleRetry bounded by maxRetryAttempts (default 5), escalates to Blocked
- Continuation retries exempt from limit (delayType: "continuation")
- handleReviewFailure routes through downstream gate's onRework
- onRework field on StageDefinition for YAML-driven rework targets
- Escalation fires side effects (updateIssueState + postComment)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: R1 adversarial review — 2 P1s + 1 P2
P1: Agent runner breaks early on [STAGE_FAILED: ...] signals (Codex finding)
- Without this, multi-turn agents could overwrite the failure signal,
and the orchestrator would never see it.
P1: completed set no longer permanently blocks resume (Codex finding)
- Issues in escalation state (Blocked) remain blocked.
- Issues moved to any other active state (Resume, Todo) get cleared
from completed and re-dispatched.
P2: lastCodexMessage empty string now filtered (Gemini finding)
- Both lastTurnMessage and lastCodexMessage check for empty strings
before being passed as agentMessage.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: R2 adversarial review — findDownstreamGate includes agent stages with onRework
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
…nce) (#8) * feat: failure signal parsing — route agent failures by class (verify/review/spec/infra) * fix: R1 adversarial review — escalation side effects + state corruption guard + empty message fallback * fix: R1 adversarial review — 2 P1s + 1 P2 P1: Persist spec-failure escalations to tracker (updateIssueState + postComment) P1: Persist review-escalation side effects when max rework exceeded P2: Add test verifying reworkCount threading to spawnWorker Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: R2 adversarial review — prevent redispatch of escalated issues * feat: max retry safety net + review rework routing (Wave 3) - scheduleRetry bounded by maxRetryAttempts (default 5), escalates to Blocked - Continuation retries exempt from limit (delayType: "continuation") - handleReviewFailure routes through downstream gate's onRework - onRework field on StageDefinition for YAML-driven rework targets - Escalation fires side effects (updateIssueState + postComment) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: R1 adversarial review — 2 P1s + 1 P2 P1: Agent runner breaks early on [STAGE_FAILED: ...] signals (Codex finding) - Without this, multi-turn agents could overwrite the failure signal, and the orchestrator would never see it. P1: completed set no longer permanently blocks resume (Codex finding) - Issues in escalation state (Blocked) remain blocked. - Issues moved to any other active state (Resume, Todo) get cleared from completed and re-dispatched. P2: lastCodexMessage empty string now filtered (Gemini finding) - Both lastTurnMessage and lastCodexMessage check for empty strings before being passed as agentMessage. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: R2 adversarial review — findDownstreamGate includes agent stages with onRework Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: prevent completed issues from being re-dispatched after merge Council Review Run 3 found that merged issues get re-dispatched because: 1. The merge/done stages had no linear_state, so the issue stayed "In Review" on Linear after completing the pipeline 2. The resume logic cleared the completed flag for ANY non-escalation active state, including "In Review" Two fixes (defense in depth): - Tighten resume guard: only "Resume" and "Todo" states clear completed flag - Add linear_state: Done to the terminal stage so issues move to "Done" on Linear when the pipeline finishes - advanceStage now fires updateIssueState for terminal stages with linearState 7 new regression tests covering all resume-guard scenarios and terminal linearState behavior. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: council R1 — add updateIssueState to dispatchIssue terminal path Council review found that the gate-to-terminal path in dispatchIssue() was missing the updateIssueState call, making the linear_state: Done config dead code for gate-based workflows. Every successfully merged issue hits this path (gate approval → continuation → dispatchIssue → terminal short-circuit) and would never update the tracker to "Done". Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: Phase 17 — stall timeout, heartbeat, turn_failed race, graceful shutdown (#7) * fix: Phase 17 — stall timeout, heartbeat, turn_failed race, graceful shutdown - Add stall_timeout_ms: 900000 (15min) to WORKFLOW-staged.md config - Add workspace file-change heartbeat to ClaudeCodeRunner (polls dir mtime every 5s, emits activity_heartbeat events to reset stall timer) - Fix turn_failed race in agent runner (check lastTurn.status after signal checks) - Fix graceful shutdown race in runtime-host (move resolveExit after waitForIdle, add pendingExitCode tracking, add agent_runner_starting/error diagnostic logs) 328 tests passing. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: council R1 — heartbeat polls .git/index, guard catches all non-completed, fix test - Heartbeat: poll .git/index mtime instead of workspace root dir (detects git staging/commits, not just root-level file creation) - Guard: change `status === "failed"` to `status !== "completed"` to also catch `cancelled` turns - Test: fix misleading test that used `completed` status when claiming to test `failed` + STAGE_FAILED interaction Council review: 3 P2s found, 3 fixed. Cross-exam eliminated 3/10 findings. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> * fix: Phase 20 — stress test bugs (heartbeat, merge race, hook resilience) 1. Heartbeat blind spot: ClaudeCodeRunner now watches workspace dir mtime alongside .git/index so review agents that never touch git still emit heartbeats and avoid stall timeout kills. 2. Merge abort race: reconcileRunningIssues() now skips terminal_state stop requests for workers in the final active stage (whose onComplete target is terminal). Prevents killing merge agents mid-flight. 3. beforeRun hook resilience: git fetch retries 3x with git lock handling, rebase is best-effort with abort fallback. Hook no longer fails the stage on git contention. 4. Stall timeout bumped from 15min to 30min. 332 tests (4 new), typecheck clean. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Closes open PRs and deletes remote branches when symphony-ts removes a workspace, preventing orphaned branches. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…#9) Extend CodexUsage interface with optional cacheReadTokens, cacheWriteTokens, noCacheTokens, and reasoningTokens fields. Extract these from the AI SDK provider's inputTokenDetails/outputTokenDetails. Add the 4 new fields to LOG_FIELDS, emit them conditionally in structured logs, accumulate them in LiveSession and CodexTotals, and add tests verifying extraction, accumulation, and absence when the provider doesn't report them. Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
… token accounting (#10) Emit a stage_completed log entry when a worker finishes, capturing accumulated LiveSession token counts (input, output, total, cache, reasoning), turn count, duration, and stage name. Adds stage_name and turns_used to LOG_FIELDS and stage_completed to ORCHESTRATOR_EVENTS. Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
… 28) Restructured Linear into 7 teams (SYMPH, JONY, HSDATA, HSUI, HSMOB, STICK, HOUSE) with distinct issue prefixes per product. Each team has a Pipeline project with unique slugId. Added per-product WORKFLOW files, a WORKFLOW template for onboarding new products, and a launcher script. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…rement (#11) - Add promptChars and estimatedPromptTokens fields to AgentRunnerEvent - Measure rendered prompt size in runner.ts turn loop before startSession/continueTurn - Add turn_number, prompt_chars, estimated_prompt_tokens to LOG_FIELDS - Log turn_number, prompt_chars, estimated_prompt_tokens in logAgentEvent (runtime-host.ts) - Add tests verifying prompt size fields are correct and turn 1 > turn 2 for long templates - Add tests verifying new fields appear in structured log entries Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
…l shutdown with worker abort and bounded timeout (#12) Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
…summary (#13) - Add 'shutdown_complete' to ORCHESTRATOR_EVENTS in src/domain/model.ts - Add 'workers_aborted' and 'timed_out' to LOG_FIELDS in src/logging/fields.ts - Change abortAllWorkers() to return the count of workers aborted - Track shutdownStart, workersAborted, and timedOut flag in shutdown() - Emit shutdown_complete log event after Promise.allSettled with workers_aborted, timed_out, and duration_ms fields - Add two new tests: shutdown_complete logged with correct fields, and timed_out=true when timeout fires Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
…atch and reconciliation summary (#14) - Add poll_tick_completed to ORCHESTRATOR_EVENTS in src/domain/model.ts - Add dispatched_count, running_count, reconciled_stop_requests to LOG_FIELDS in src/logging/fields.ts - Extend PollTickResult with runningCount in src/orchestrator/core.ts - Add duration timing around pollOnce() in runPollCycle() - Emit poll_tick_completed info log with dispatched_count, running_count, reconciled_stop_requests, duration_ms - Add tests for poll_tick_completed event emission and dispatched_count accuracy Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
…atch and reconciliation summary (#15) - poll_tick_completed already in ORCHESTRATOR_EVENTS (src/domain/model.ts) - dispatched_count, running_count, reconciled_stop_requests in LOG_FIELDS (src/logging/fields.ts) - logPollCycleResult() emits poll_tick_completed with dispatched_count, running_count, reconciled_stop_requests, duration_ms (src/orchestrator/runtime-host.ts) - Duration timing around pollOnce() in runPollCycle() - PollTickResult includes dispatchedIssueIds, runningCount, stopRequests (src/orchestrator/core.ts) - Tests: poll_tick_completed event logged after successful poll; dispatched_count reflects dispatched issues - Update conformance-test-matrix.md to document poll_tick_completed observability coverage Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
…vent (#16) Add five new fields to the stage_completed structured log event that give operators a single-event view of total token cost per pipeline stage: - total_input_tokens: sum of per-turn input tokens across the stage - total_output_tokens: sum of per-turn output tokens across the stage - total_cache_read_tokens: accumulated cache-read tokens across all turns - total_cache_write_tokens: accumulated cache-write tokens across all turns - turn_count: number of turns executed in the stage New LiveSession fields codexTotalInputTokens and codexTotalOutputTokens accumulate turn-level deltas. Per-turn deltas are computed correctly by resetting lastReported* counters on session_started, so each turn's absolute counter starts from zero for delta computation. Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(SYMPH-8): add accumulator fields and summation logic for stage-level token totals Add totalStageInputTokens, totalStageOutputTokens, totalStageTotalTokens, totalStageCacheReadTokens, and totalStageCacheWriteTokens to LiveSession. Accumulate turn deltas into these fields in applyCodexEventToSession(). Add single-turn, multi-turn, and zero-turn tests. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix(SYMPH-8): add missing totalStage accumulator fields to runtime-host test fixtures TypeScript compilation was failing because LiveSession object literals in runtime-host tests were missing the new totalStage* fields added to the LiveSession interface. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
…#19) Update total_input_tokens, total_output_tokens to use totalStage* accumulator fields. Add total_total_tokens. Make total_cache_read_tokens and total_cache_write_tokens conditional (omitted when zero) using totalStage* accumulators. Existing input_tokens, output_tokens, total_tokens preserve last-turn semantics. Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
…field (#20) Add StageRecord and ExecutionHistory interfaces to model.ts. Add issueExecutionHistory: Record<string, ExecutionHistory> to OrchestratorState and initialize it as {} in createInitialOrchestratorState. Add a thin scripts/test.mjs wrapper to translate --grep to vitest's -t flag so mocha-compatible verify commands work. Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
…#67) Change investigate.model and implement.model from claude-sonnet-4-5 to claude-opus-4-6 in both WORKFLOW-symphony.md and WORKFLOW-template.md. The review stage already uses Opus; merge stays on Sonnet for its mechanical 3-command operation. Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Add a "Pipeline" column to the running sessions table that shows total elapsed time from first_dispatched_at to generated_at for multi-stage issues. Single-stage issues (first_dispatched_at equals started_at) show "—" in the Pipeline column. Updates both the server-rendered HTML and the client-side SSE render function. Adds formatPipelineTime helper to dashboard-render.ts and a corresponding client-side JS function in the embedded script. - Add <th>Pipeline</th> after "Runtime / turns" in table header - Add <col> to colgroup (7 columns total) - Add formatPipelineTime() server-side TypeScript helper - Add formatPipelineTime() client-side JavaScript function - Update colspan from 6 → 7 in empty state and detail rows - Add unit tests in tests/observability/dashboard-render.test.ts - Update colspan assertion in dashboard-server.test.ts Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(SYMPH-43): add post-creation verification layer to freeze-and-queue.sh Adds verify_issue_creation() helper that queries a Linear issue by ID after each issueCreate/issueUpdate mutation and confirms project.slugId matches the expected PROJECT_SLUG and (for sub-issues) parent.id matches PARENT_ID. Logs warnings on mismatch; never exits. Called at all 4 creation sites: trivial issue, parent update, parent create, and sub-issue create loop. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix(SYMPH-43): move verify_issue_creation() before first call site Fix P1 bash ordering bug: verify_issue_creation() was defined at line 894 but called at lines 161, 673, 739, and 839. Bash requires functions to be defined (executed) before their call sites are reached at runtime. Moves the function definition to line 51, before the trivial mode block where the first call site lives. Also removes INVESTIGATION-BRIEF.md (out-of-scope file from previous commit). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * feat(SYMPH-56): implement fast-track label-based stage routing Add FastTrackConfig interface and fastTrack field to StagesConfig, parse fast_track YAML config key, validate fast_track.initial_stage references a defined stage, and apply fast-track routing in dispatchIssue() when an issue carries the configured label and has no cached stage. Update all existing StagesConfig test stubs to include fastTrack: null and add 7 new unit tests covering all specified scenarios. Enable fast_track config in WORKFLOW-symphony.md and WORKFLOW-template.md. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * chore: fix biome formatting in core.test.ts Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
The blockedBy enforcement fix (removing the todo-only guard) was already applied in commit 72ec271. This commit adds an explicit test for the Resume state scenario mentioned in SYMPH-50 — verifying that issues in Resume state are also rejected when their blockers are non-terminal. SYMPH-50 Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix: convert all timestamps from UTC to US Eastern All user-facing timestamps (dashboard, structured logs, session metrics) now display in America/New_York timezone instead of UTC. Added a shared formatEasternTimestamp() utility that handles DST transitions. The only exception is linear-normalize.ts which must stay UTC for Linear API compatibility. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: lint issues (import order, Number.parseInt) --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
… Todo after relations (#73) - Change sub-issue creation guard from TODO_STATE_ID to BACKLOG_STATE_ID so all sub-issues land in Backlog state at creation time - After relation-creation loops complete, identify unblocked sub-issues (those whose identifier does not appear on the right side of any pair in CREATED_RELATIONS) and promote them to Todo via issueUpdate - Blocked sub-issues remain in Backlog; logs show which were promoted vs blocked Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
…ticks (#75) Replace $(cat "$SPEC_TMPFILE") with direct variable passing ($SPEC_CONTENT and $sub_body) in all 6 GraphQL mutation call sites. The $(cat ...) pattern inside double-quoted -v flags caused bash to interpret backticks in spec content as command substitution, hanging the shell. Also move function definitions (resolve_team_from_project, resolve_all_states, LINEAR_CLI, state globals) before the --trivial code path so they are defined before being called in all execution paths. Remove unnecessary SPEC_TMPFILE temp file (no longer needed since descriptions are passed directly via variables). Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* feat(SYMPH-60): add auto-close parent issue on terminal state After marking a sub-issue as Done (or Cancelled), query its parent issue via Linear GraphQL. If the parent exists and ALL sibling sub-issues are in terminal states, transition the parent to Done. Best-effort — errors are logged, never block the pipeline. - Add LINEAR_ISSUE_PARENT_AND_SIBLINGS_QUERY in linear-queries.ts - Add checkAndCloseParent method on LinearTrackerClient - Add optional autoCloseParentIssue callback on OrchestratorCoreOptions - Fire callback in advanceStage() after terminal updateIssueState - Wire callback in runtime-host.ts using LinearTrackerClient method - Add 4 tests covering: fires on terminal, skips non-terminal, error resilience, and missing callback Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(SYMPH-60): add fastTrack: null to test stage configs for StagesConfig compat Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
…#76) Add "State: Backlog (promoted to Todo after relations)" to sub-issue dry-run output and update the summary line to mention the Backlog→Todo flow. Trivial mode output remains unchanged. Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
) Restructure freeze-and-queue.sh sub-issue creation to create issues directly in Todo state with sequential blocking relations added immediately after each sub-issue, eliminating the Backlog staging and promotion loop. Changes: - Create sub-issues using TODO_STATE_ID instead of BACKLOG_STATE_ID - Iterate creation loop via SORTED_INDICES for priority ordering - Add sequential blockedBy relation immediately after each sub-issue (except the first, which is immediately dispatchable) - Move create_blocks_relation() and verify_issue_creation() function definitions before the creation loop - Keep file-overlap relations as a second pass after creation - Remove BACKLOG_STATE_ID fallback for sub-issues (lines 600-611) - Remove "Moving unblocked sub-issues to Todo" promotion loop - Keep parent issue lifecycle unchanged (Draft creation, Backlog transition) Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* feat(SYMPH-66): add rebase failure class and orchestrator routing Add "rebase" to FAILURE_CLASSES and update the STAGE_FAILED regex to recognize [STAGE_FAILED: rebase] signals. Add handleRebaseFailure() private method to core.ts that mirrors the first half of handleReviewFailure() — checks current stage has onRework, calls reworkGate(), posts a rebase comment, and schedules continuation retry. Add formatRebaseComment() to gate-handler.ts. Add routing case in handleFailureSignal(). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * style(SYMPH-66): fix biome formatting in model.ts and test files Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
…#78) * feat(SYMPH-67): add rebase flow support to WORKFLOW prompts and hooks Add merge conflict detection and rebase rework flow to the pipeline: - Add on_rework: implement and max_rework: 2 to merge stage YAML config - Update merge stage prompt to check PR mergeability and write REBASE-BRIEF.md on conflicts - Update implement stage rework prompt to handle rebase rework via REBASE-BRIEF.md - Add REBASE-BRIEF.md import block to before_run hook Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(SYMPH-67): restore fast_track config blocks removed out-of-scope Restores the fast_track configuration in both WORKFLOW-template.md (commented-out example) and WORKFLOW-symphony.md (active config for trivial-labeled issues). These were accidentally removed in the initial implementation but are not part of the SYMPH-67 spec. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Update the dry-run section to show sub-issues created in Todo state with interleaved sequential blocking relations shown immediately after each sub-issue, rather than in a separate section. File-overlap relations remain as a separate second-pass section. Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* feat(SYMPH-70): Replace turn history with recent activity feed Add RecentActivityEntry type and ring buffer (max 10) to LiveSession. Populate from approval_auto_approved events by extracting tool name and context (file basename for Read/Edit/Write/Glob, truncated command for Bash, pattern for Grep). Pass through to RuntimeSnapshotRunningRow and render in dashboard detail panel with relative timestamps. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: apply biome formatting to session-metrics and recent-activity test Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix: add recent_activity field to dashboard-render test fixture The dashboard-render.test.ts fixture was missing the recent_activity field added by SYMPH-70 to RuntimeSnapshotRunningRow. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
…iled counts (#81) * feat(SYMPH-69): Add issue title, fix UTC timestamps, add completed/failed counts - Add issue_title field to RuntimeSnapshotRunningRow, populated from entry.issue.title - Format last_event_at through formatEasternTimestamp() instead of passing raw UTC strings from Codex events - Add completed and failed counts to RuntimeSnapshot.counts, computed from issueExecutionHistory final stage outcomes - Add Completed and Failed metric cards to dashboard header (both server-side HTML and client-side JS render()) - Display issue title below identifier in running sessions table - Verify pipeline total time works for both single and multi-stage issues via first_dispatched_at fallback Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(SYMPH-69): resolve rebase conflicts and align with format-timestamp.ts - Remove duplicate formatEasternTimestamp from runtime-snapshot.ts (moved to format-timestamp.ts by earlier PR #72) - Add invalid-date guard (returns "n/a") to format-timestamp.ts - Fix non-null assertion → use Array.at(-1) with optional chaining - Update tests to expect ISO-8601 Eastern format instead of human-readable - Add missing issue_title and completed/failed fields to test fixtures Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* Add sops+age encrypted secrets (.sops.yaml + .env.enc) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * ops: add HOME, NODE_ENV, SoftResourceLimits to plist; bump ThrottleInterval to 60s - symphony-ctl generate_plist: add HOME env var (launchd doesn't inherit it), NODE_ENV=production, SoftResourceLimits/NumberOfFiles 4096 - ThrottleInterval 30→60 for crash restart breathing room - Example plist updated to match for reference consistency Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Changes default dashboard bind address from 127.0.0.1 to 0.0.0.0 so the dashboard is accessible from other machines on the network. Needed for production server deployment where the dashboard is accessed remotely.
* fix: bind dashboard server to 0.0.0.0 for network access Changes default dashboard bind address from 127.0.0.1 to 0.0.0.0 so the dashboard is accessible from other machines on the network. Needed for production server deployment where the dashboard is accessed remotely. * fix: resolve symlinks in symphony-ctl SCRIPT_DIR When symphony-ctl is invoked via a symlink, SCRIPT_DIR resolved to the symlink directory instead of the real script directory. Wrap BASH_SOURCE with realpath before dirname.
When symphony-ctl is symlinked (e.g. /usr/local/bin/symphony-ctl -> ops/symphony-ctl), BASH_SOURCE[0] returns the symlink path, not the target. This caused SYMPHONY_ROOT to resolve to /usr/local instead of the project directory, which meant .env was never found and the plist was installed without LINEAR_API_KEY — causing a silent crash-loop. Use realpath to resolve the symlink before computing SCRIPT_DIR. Co-authored-by: Claw Dilize <clawdilize@pro16.local> Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Bug 1: Change all ID! to String! in GraphQL mutation variable declarations to match Linear's schema expectations - Bug 2: Refactor create_blocks_relation() to use parameterized variables (-v flags) with single-quoted heredoc instead of inlining UUIDs via bash interpolation. Also parameterize the parent→Backlog transition mutation. - Bug 3: Move create_blocks_relation() and verify_issue_creation() definitions above the trivial mode block so they are defined before being called Co-authored-by: Claw Dilize <clawdilize@pro16.local> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
ericlitman
force-pushed
the
eric/symph-71-teach-merge-agent-about-github-merge-queue-eliminate-polling
branch
from
f4dc013 to
6719615
Compare
ericlitman
changed the title
…CC provider (#87) * feat(SYMPH-73): scaffold Slack bot module wiring Chat SDK → AI SDK → CC provider Add slack-bot module that receives channel messages via @chat-adapter/slack, invokes Claude Code via ai-sdk-provider-claude-code streamText, and posts threaded replies with reaction indicators (eyes → checkmark). - src/slack-bot/types.ts: ChannelProjectMap, SlackBotConfig, SessionMap types - src/slack-bot/handler.ts: message handler with reaction lifecycle and paragraph chunking - src/slack-bot/index.ts: Chat instance setup with SlackAdapter + MemoryStateAdapter - .env.example: documents required environment variables - tests/slack-bot/handler.test.ts: 11 tests for CC invocation, threading, chunking - tests/slack-bot/reactions.test.ts: 3 tests for reaction lifecycle Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * feat(SYMPH-76): fix 3 bugs in freeze-and-queue.sh - Bug 1: Change all ID! to String! in GraphQL mutation variable declarations to match Linear's schema expectations - Bug 2: Refactor create_blocks_relation() to use parameterized variables (-v flags) with single-quoted heredoc instead of inlining UUIDs via bash interpolation. Also parameterize the parent→Backlog transition mutation. - Bug 3: Move create_blocks_relation() and verify_issue_creation() definitions above the trivial mode block so they are defined before being called Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(SYMPH-73): resolve biome lint and format errors blocking CI - Format package.json files array to single line - Fix useTemplate lint error in handler.ts (concat → template literal) - Reorder imports in index.ts for organizeImports compliance - Replace generator failingStream (useYield error) with plain async iterable object - Reformat test files to pass biome line-length checks Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> --------- Co-authored-by: Claw Dilize <clawdilize@pro16.local> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Remove projectId from sub-issue issueCreate mutations and defer it to a batch issueUpdate pass after blocking relations are verified. Add verify_issue_creation_parent_only() for sub-issue parent checks, verify_blocking_relations() to confirm inverseRelations direction, and assign_project_to_sub_issues() for post-verification assignment. Update dry-run output to describe the deferred flow. Co-authored-by: Claw Dilize <clawdilize@pro16.local> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
…ths (#90) After migrating to launchd, logs moved to ~/Library/Logs/symphony/<product>/ but analyze, cleanup, and usage text still referenced /tmp/symphony-*. Co-authored-by: Claw Dilize <clawdilize@pro16.local> Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
ericlitman
force-pushed
the
eric/symph-71-teach-merge-agent-about-github-merge-queue-eliminate-polling
branch
from
f053d28 to
8b43f1e
Compare
ericlitman
changed the title
Replace merge stage prompt to eliminate polling waste: - Add Merge Queue Context section explaining gh pr merge responses - Remove redundant mergeability pre-check (Step 1) - Replace sleep/poll loops with gh pr checks --watch --required --fail-fast - Add explicit DO NOT list (no retries, no --admin, no code mods) - Add merge queue rejection path signaling [STAGE_FAILED: rebase] - Preserve Step 2b (rebase brief) and Workpad sections unchanged Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
ericlitman
force-pushed
the
eric/symph-71-teach-merge-agent-about-github-merge-queue-eliminate-polling
branch
from
8b43f1e to
89ad5a9
Compare
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.
Summary
gh pr mergereturns when merge queue is active (queue added vs auto-merge enabled responses)gh pr mergehandles this internallygh pr checks --watch --required --fail-fast— a single blocking call that waits for all checks--admin, no code modifications in merge stage[STAGE_FAILED: rebase]Addresses token waste from polling (SYMPH-24: 1.47M tokens, SYMPH-52: 2.07M tokens on sleep/check loops).
Changes applied to both:
pipeline-config/templates/WORKFLOW-template.mdpipeline-config/workflows/WORKFLOW-symphony.mdTool Output > TypeScript
Tool Output > Tests
SAST Output
Verification
Merge Queue Contextsection present[STAGE_COMPLETE]and[STAGE_FAILED: rebase]signal contract unchangedTest plan
pnpm vitest runhas no new failures (done, all pre-existing)🤖 Generated with Claude Code