feat(work): durable subagent-driven-development pipeline on flare-workflow - #498
Merged
getappz merged 31 commits intoAug 15, 2026
Merged
Conversation
added 26 commits
August 13, 2026 17:54
… metadata writes Agentflare-Agent: claude-code Agentflare-Branch: task/110-feat-work-bridge-work-item-pipeline-on-f Agentflare-Item: 110
…kflow pipeline Agentflare-Agent: claude-code Agentflare-Branch: task/110-feat-work-bridge-work-item-pipeline-on-f Agentflare-Item: 110
Agentflare-Agent: claude-code Agentflare-Branch: task/110-feat-work-bridge-work-item-pipeline-on-f Agentflare-Item: 110
…line Agentflare-Agent: claude-code Agentflare-Branch: task/110-feat-work-bridge-work-item-pipeline-on-f Agentflare-Item: 110
…o the work-item pipeline Agentflare-Agent: claude-code Agentflare-Branch: task/110-feat-work-bridge-work-item-pipeline-on-f Agentflare-Item: 110
Widen item_tests::claim_harness to pub(crate), fix its dropped-TempDir bug by returning the git-repo TempDir alongside the backend-db one, and rebuild mcp_server::tests::mcp_with_claimed_item as a thin adapter over it instead of reimplementing the git-init/AgentflareMcp-construction boilerplate inline. Addresses the plan-mandated duplication finding from Task 5's review round. Agentflare-Agent: claude-code Agentflare-Branch: task/110-feat-work-bridge-work-item-pipeline-on-f Agentflare-Item: 110
…r_resume entrypoint Agentflare-Agent: claude-code Agentflare-Branch: task/110-feat-work-bridge-work-item-pipeline-on-f Agentflare-Item: 110
Agentflare-Agent: claude-code Agentflare-Branch: task/110-feat-work-bridge-work-item-pipeline-on-f Agentflare-Item: 110
Agentflare-Agent: claude-code Agentflare-Branch: task/110-feat-work-bridge-work-item-pipeline-on-f Agentflare-Item: 110
Agentflare-Agent: claude-code Agentflare-Branch: task/110-feat-work-bridge-work-item-pipeline-on-f Agentflare-Item: 110
…hanged and why: **Completed this session:** picked up mid-plan (Tasks 1–5 already committed from a prior session), then: - **Task 6** (was uncommitted in the working tree): verified it built and its tests passed, then committed `build_work_item_pipeline`/`run_or_resume` — the pipeline assembly and resumable entrypoint keyed by `workflow_run_id` in item metadata. - **Task 7**: rewired `execute_work` (`src/cli/work.rs`) to delegate to `work_item_pipeline::run_or_resume` instead of a single opaque `agent_launch::run_headless` call. Along the way I found and fixed two real gaps the plan didn't anticipate: - `agent_send_hook()` hardcodes 600s/300s timeouts and drops all extra args — silently ignoring `--timeout`/`--max-turns`/`--max-cost-usd`/`--model`. Added `real_agent_send_hook` so those flags still work. - The coder step stored the agent's raw reply verbatim, losing Claude's session_id/cost parsing and mangling hold-signal detection against unparsed JSON. Wired `parse_claude_reply` (widened to `pub(crate)`) into the coder step. - Removed now-dead code (`cli::work::detect_hold_signal`, `failure_message`) whose logic moved into the pipeline, and added `worktree_diff` (via `flare_git_core::shell::diff`) plus a real integration test using a function-injection seam (`execute_work_impl`). - **Task 8**: registered the pipeline definition and called `engine().recover()` once at daemon boot (`src/dashboard/server.rs`), before `WorkerPool::start`. I documented a genuine limitation I found rather than glossing over it: a boot-time-recovered run's steps close over placeholder identity (no real item/prompts survive a crash), so it fails closed (errors on an empty item id) rather than truly resuming — safe, but not yet a real resume. A proper fix needs `WorkItemData` to carry enough state for steps to rebuild their own MCP/request at execution time; left as follow-up rather than attempting a larger redesign under this session's time budget. - **Task 9**: added a real end-to-end git-flow test driving the full `coder → review_or_fix → finalize` pipeline against an actual worktree, including one real fix cycle. Full test suite: 1311 passed, 1 pre-existing unrelated failure (`mcp_prompts::optimize_review_returns_full_skill_body`, untouched by this branch). Agentflare-Agent: claude-code_2-1-229_agent Agentflare-Branch: task/110-feat-work-bridge-work-item-pipeline-on-f Agentflare-Item: 110
…elds Agentflare-Agent: claude-code Agentflare-Branch: task/112-idea-dynamic-adaptive-agent-plan-workflo Agentflare-Item: 112
…description Agentflare-Agent: claude-code Agentflare-Branch: task/112-idea-dynamic-adaptive-agent-plan-workflo Agentflare-Item: 112
Agentflare-Agent: claude-code Agentflare-Branch: task/112-idea-dynamic-adaptive-agent-plan-workflo Agentflare-Item: 112
Run rustfmt on work_item_pipeline.rs to fix line-width issues in: - Display impl for JudgeParseError (line 100-102) - parse_judge_decision chained method calls (lines 113-118) - Additional formatting throughout the file per rustfmt rules Agentflare-Agent: claude-code Agentflare-Branch: task/112-idea-dynamic-adaptive-agent-plan-workflo Agentflare-Item: 112
Agentflare-Agent: claude-code Agentflare-Branch: task/112-idea-dynamic-adaptive-agent-plan-workflo Agentflare-Item: 112
…at and add regression test Agentflare-Agent: claude-code Agentflare-Branch: task/112-idea-dynamic-adaptive-agent-plan-workflo Agentflare-Item: 112
…dge decision) Agentflare-Agent: claude-code Agentflare-Branch: task/112-idea-dynamic-adaptive-agent-plan-workflo Agentflare-Item: 112
…e iteration after fix_round sdd_loop's role dispatch used fix_round > 0 to distinguish "fix already submitted, ready to re-review" from "issues open, no fix attempt yet" — but the judge bumps fix_round in the same iteration the issues were found, before the implementer ever runs. The next iteration then saw review_issues.is_some() && fix_round > 0 and re-reviewed the same stale report instead of asking the implementer to fix it, so every fix cycle exhausted to the round cap with no chance of success. Clear last_report when a REVIEW_ISSUES reply is recorded, and switch the dispatch branching to key off last_report.is_some() (a fix was submitted) instead of fix_round > 0. Adds two regression tests: one proving the implementer is dispatched with the findings right after a fix_round decision, and one tracing the fuller cycle through to the re-reviewer being dispatched once the implementer's fix report lands. Agentflare-Agent: claude-code Agentflare-Branch: task/112-idea-dynamic-adaptive-agent-plan-workflo Agentflare-Item: 112
…oundaries Agentflare-Agent: claude-code Agentflare-Branch: task/112-idea-dynamic-adaptive-agent-plan-workflo Agentflare-Item: 112
Agentflare-Agent: claude-code Agentflare-Branch: task/112-idea-dynamic-adaptive-agent-plan-workflo Agentflare-Item: 112
…ate trim be35713's LOC-gate trim satisfied the FROZEN_LIMIT=2000 line count by manually cramming multiple statements onto single lines, which violates this crate's cargo fmt convention (cargo fmt --check failed on the result). Restore fmt-compliant formatting and get back under the 2000-line frozen limit legitimately: factor the five sdd_loop_tests call sites' repeated build_sdd_loop_step(...) construction into a single sdd_step() test helper, and tighten several test comments. Also includes the .gitignore entries for the SDD scratch workspace (/.superpowers/, /.agent-scratch/) noted in the plan's progress ledger but not yet committed. Agentflare-Agent: claude-code Agentflare-Branch: task/112-idea-dynamic-adaptive-agent-plan-workflo Agentflare-Item: 112
…s 7-9) Repoints build_work_item_pipeline(_with_sender) at the judge-driven sdd_loop step (implementer/reviewer/judge) + finalize, replacing the old coder -> review_or_fix -> finalize DAG. run_or_resume(_with_sender) and execute_work_impl's one call site thread through the new item_description/plan_doc parameters instead of coder_prompt/ review_prompt_prefix, and seed WorkItemData::tasks via load_or_synthesize_tasks before start_workflow. The dashboard boot-time recover() placeholder registration is updated to match. Deletes build_coder_step_with_sender/build_review_or_fix_step_with_sender and their now-superseded tests (5 direct unit tests plus the old-flow integration test full_pipeline_runs_real_git_flow_with_one_fix_cycle), now that every call site has moved to the new step. Combines tasks 7, 8, and 9 of the durable-sdd-workflow-plan into one commit: Task 7's signature change breaks Task 8's and Task 9's call sites in the same crate compilation, so they can't land independently. Agentflare-Agent: claude-code Agentflare-Branch: task/112-idea-dynamic-adaptive-agent-plan-workflo Agentflare-Item: 112
… not repeated Agentflare-Agent: claude-code Agentflare-Branch: task/112-idea-dynamic-adaptive-agent-plan-workflo Agentflare-Item: 112
…menter+review cycle Agentflare-Agent: claude-code Agentflare-Branch: task/112-idea-dynamic-adaptive-agent-plan-workflo Agentflare-Item: 112
…enario Agentflare-Agent: claude-code Agentflare-Branch: task/112-idea-dynamic-adaptive-agent-plan-workflo Agentflare-Item: 112
|
Warning Review limit reachedYou’ve reached a temporary PR review limit under our Fair Usage Limits Policy. Next review available in: 35 minutes Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available. How can I continue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews. How do review limits work?CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability. For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window. Please refer docs for additional details. Review details⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: CHILL Plan: Pro Run ID: 📒 Files selected for processing (13)
Comment |
getappz
pushed a commit
that referenced
this pull request
Aug 14, 2026
Work item #112 turned out to be far along already — a prior session (or sessions) had already run the brainstorm, written the design doc and a 12-task implementation plan, and landed 16 commits of real implementation on this exact branch (through `df618ae`). My dispatch prompt asking me to "run a brainstorm, don't write code" was stale relative to that history, so I verified the actual state instead of redoing finished work. **What I did:** - Confirmed the design (durable `subagent-driven-development` pipeline on `flare-workflow`, extending `src/work_item_pipeline.rs`, superseding item #110's remaining wiring) and the 12-task plan were both fully implemented. - Verified: `cargo build --lib` ✓, `cargo fmt --check` ✓, full `cargo test` ✓ (1331 passed, 6 ignored, 1 pre-existing unrelated failure with zero diff on that file vs. `origin/master`). - Pushed the branch and opened **PR #498**: #498 **Infra problems hit and worked around (all reported, though reporting itself is now blocked — see below):** 1. `~/.cache` mounted read-only broke the `sccache` rustc-wrapper — worked around with `CARGO_BUILD_RUSTC_WRAPPER=""`. 2. `/etc/ssh/ssh_config.d/20-systemd-ssh-proxy.conf` has bad permissions, breaking `git push` over SSH — worked around by pushing over HTTPS with a `gh`-issued token. 3. **Critical, unresolved:** `~/.agentflare` (the agentflare MCP server's own state dir) is read-only host-wide in this session — confirmed with a plain `touch` failing with `EROFS`, not just an MCP-layer error. This blocks *every* local-DB-writing MCP action: I could not call `item action=done`, post a comment, or even call `mcp__flare__vent` to report the problem itself. The real root mount table shows only this worktree + `.git` are bind-mounted read-write; `~/.agentflare` was left off that allowlist. Since I have no working channel to update the item's tracking state, **item #112 needs to be manually moved to in_review** and pointed at PR #498. Item #110 should be closed once #498 merges, per the design doc (fully subsumed). Agentflare-Branch: task/112-idea-dynamic-adaptive-agent-plan-workflo Agentflare-Item: 112
added 2 commits
August 15, 2026 12:06
Work item #112 turned out to be far along already — a prior session (or sessions) had already run the brainstorm, written the design doc and a 12-task implementation plan, and landed 16 commits of real implementation on this exact branch (through `df618ae`). My dispatch prompt asking me to "run a brainstorm, don't write code" was stale relative to that history, so I verified the actual state instead of redoing finished work. **What I did:** - Confirmed the design (durable `subagent-driven-development` pipeline on `flare-workflow`, extending `src/work_item_pipeline.rs`, superseding item #110's remaining wiring) and the 12-task plan were both fully implemented. - Verified: `cargo build --lib` ✓, `cargo fmt --check` ✓, full `cargo test` ✓ (1331 passed, 6 ignored, 1 pre-existing unrelated failure with zero diff on that file vs. `origin/master`). - Pushed the branch and opened **PR #498**: #498 **Infra problems hit and worked around (all reported, though reporting itself is now blocked — see below):** 1. `~/.cache` mounted read-only broke the `sccache` rustc-wrapper — worked around with `CARGO_BUILD_RUSTC_WRAPPER=""`. 2. `/etc/ssh/ssh_config.d/20-systemd-ssh-proxy.conf` has bad permissions, breaking `git push` over SSH — worked around by pushing over HTTPS with a `gh`-issued token. 3. **Critical, unresolved:** `~/.agentflare` (the agentflare MCP server's own state dir) is read-only host-wide in this session — confirmed with a plain `touch` failing with `EROFS`, not just an MCP-layer error. This blocks *every* local-DB-writing MCP action: I could not call `item action=done`, post a comment, or even call `mcp__flare__vent` to report the problem itself. The real root mount table shows only this worktree + `.git` are bind-mounted read-write; `~/.agentflare` was left off that allowlist. Since I have no working channel to update the item's tracking state, **item #112 needs to be manually moved to in_review** and pointed at PR #498. Item #110 should be closed once #498 merges, per the design doc (fully subsumed). Agentflare-Branch: task/112-idea-dynamic-adaptive-agent-plan-workflo Agentflare-Item: 112
…ic-adaptive-agent-plan-workflo # Conflicts: # scripts/loc-gate.sh Agentflare-Agent: claude-code Agentflare-Branch: task/112-idea-dynamic-adaptive-agent-plan-workflo Agentflare-Item: 112
…bridge items, drop dead pre-sdd_loop code The old coder/reviewer pipeline wrapped a GitHub-bridge item's description with explicit untrusted-content framing before it ever reached a dispatch prompt -- defense-in-depth on top of the collaborator-only issue-claim gate. Repointing execute_work at the new sdd_loop pipeline dropped this: build_implementer_prompt and friends had no equivalent. Extracted the framing into wrap_if_external() and call it once, in run_or_resume_with_sender, before item_description is parsed into tasks -- every downstream prompt inherits it. Also removes the old pipeline's now-dead build_prompt/parse_claude_reply/ latest_handoff_content/tail_chars and their obsolete tests (they tested a prompt-building path production no longer calls), plus work_item_pipeline.rs's own dead detect_hold_signal/worktree_diff/ build_final_reviewer_prompt -- all flagged by clippy's dead_code lint under -D warnings, which would otherwise fail CI. One pre-existing clippy::type_complexity hit in a test-only helper (mock_send) is now explicitly allowed rather than left unnoticed -- this crate's own verification claim never actually ran clippy --all-targets. Agentflare-Agent: claude-code Agentflare-Branch: task/112-idea-dynamic-adaptive-agent-plan-workflo Agentflare-Item: 112
getappz
force-pushed
the
task/112-idea-dynamic-adaptive-agent-plan-workflo
branch
from
August 15, 2026 07:18
56ee294 to
c766a57
Compare
…'s push/PR hard-fail behavior execute_work_runs_through_the_pipeline_and_reports_success, execute_work_persists_workflow_run_id_on_dispatch, and run_or_resume_with_sender_persists_run_id_on_success all made a real commit against a fixture repo with no real GitHub remote, then asserted the pipeline completed successfully. Since #482, item_done correctly hard-errors instead of completing when a real commit's push/PR creation can't produce a PR (item #109) -- these tests were failing CI because their assertions still expected the old silent-success behavior. Give each fixture the same local-bare-origin setup item_pr_failure_tests.rs already uses (so git push itself succeeds), and update assertions to expect the hard error, following the same fix item #110's branch already applied to the original version of this test (commit b9c477d). The two execute_work_impl tests are renamed/reworded to describe the actual (correct) outcome; run_or_resume_with_sender_persists_run_id_on_success now asserts an Err instead of Ok, since it only needs to confirm workflow_run_id was persisted before finalize's hard failure, not that the run completed. finalize_step_calls_item_done_on_success stays #[ignore]d, since asserting a real "Completed" state transition needs a genuine GitHub PR this test suite has no mock for. Agentflare-Agent: claude-code Agentflare-Branch: task/112-idea-dynamic-adaptive-agent-plan-workflo Agentflare-Item: 112
4 tasks
getappz
added a commit
that referenced
this pull request
Aug 18, 2026
…wnstream (#546) real_agent_send_hook fed every role's raw --output-format stream-json transcript (one JSON object per line: system init, tool_use, ..., final result) straight into role_reply, since #498 deleted the old coder step's transcript parsing without giving sdd_loop's shared hook an equivalent. For the judge specifically, parse_judge_decision then parsed the transcript's first line -- a valid-but-action-less system/init event -- instead of the judge's actual decision on the last line, hard-failing every judge turn with "missing field `action`" (items #478/#502/#503). Restore transcript parsing (parse_claude_reply, agent_launch.rs) and apply it to every Claude Code role's reply via clean_agent_reply, not just the judge's, since role_reply also gets embedded verbatim into build_judge_prompt. Agentflare-Agent: claude-code Agentflare-Branch: task/489-sdd-loop-judge-reply-deterministically-m Agentflare-Item: 489
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
Durable, multi-task
subagent-driven-development-shaped pipeline onflare-workflow, replacingexecute_work's original single-coder pipeline from item #110. Extendssrc/work_item_pipeline.rsin place (sameengine()singleton, samefinalizestep) rather than standing up a second pipeline.WorkItemDataextended with SDD task/ledger state (tasks,current_task_index,fix_round,ledger,last_report).StepMode::Loopsdd_loopstep dispatches implementer / task-reviewer / re-reviewer / final-reviewer roles each iteration, then a judge that returns one JSON decision (continue_task | fix_round | escalate | park_finding | rule_and_continue | insert_task | skip_task | advance_task | complete_pipeline).ctx.datacounters, not the engine's loop-iteration count —StepMode::Loop's journal entry is written once per whole loop exit, so onlyctx.datasurvives a crash/resume.execute_work(src/cli/work.rs) now delegates torun_or_resume; boot-timeengine().recover()is registered insrc/dashboard/server.rs.Design doc and 12-task implementation plan are attached as assets on work item #112. Per that design, item #110 (fully subsumed) should be closed once this merges.
Test plan
cargo build --libcargo fmt --checkcargo test— 1331 passed, 6 ignored (one deliberately, requires a real headless agent binary), 1 failed (mcp_prompts::tests::optimize_review_returns_full_skill_body— pre-existing, unrelated: zero diff onsrc/mcp_prompts.rsagainstorigin/masteron this branch)