feat(symphony): factory autonomous SDLC workflow pack (ENG-2921)#1127
feat(symphony): factory autonomous SDLC workflow pack (ENG-2921)#1127hari (harivansh-afk) wants to merge 1 commit into
Conversation
A Symphony pack implementing the agent workflow from ENG-2921: five agent roles (idea-generator, cleaner, researcher, worker, reviewer) plus an escalate escape hatch that take an idea to a green, auto-merged PR on ix. - Two cron feeders (idea-generator, cleaner) file deduped Linear issues. - task-pipeline runs per issue on the factory:scope-needed label: scope -> implement -> bounded review/fix cascade -> auto-merge. The Worker<->Reviewer loop is an unrolled, flat when-cascade (Symphony is acyclic); MAX_ROUNDS=3 is the structural convergence cap. - Reuses the ix-playbook-agent bot path for PRs + auto-merge. - Pure pack + config; no runtime changes. All three .sym validated through the real parser + interpreter + materializer.
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
There was a problem hiding this comment.
AI review found issues in this pull request.
Verdict: patch is incorrect
Confidence: 0.91
The factory pack is not runnable as added: skills with partials fail to load, agent nodes lack a placement/cwd, the workflow depends on unsupported structured agent outputs, and skipped conditional branches can deadlock. Additional delegate/tool paths also cannot work with the current runtime wiring.
- P1
packages/symphony/workflows/factory/skills/researcher.md:29Move partials under the skills directory - P1
packages/symphony/workflows/factory/workflows/task-pipeline.sym:13Set a real placement for agent nodes - P1
packages/symphony/workflows/factory/_partials/decision-contract.md:3Do not gate on final-message JSON until it is parsed - P1
packages/symphony/workflows/factory/workflows/task-pipeline.sym:21Flattened gates can deadlock on skipped branches - P2
packages/symphony/workflows/factory/skills/worker.md:10Delegate cannot retrigger while the current run is active - P2
packages/symphony/workflows/factory/skills/idea-generator.md:1Declared tools are not available to these agents
| {{partial:ix-validation}} | ||
|
|
||
| {{partial:decision-contract}} |
There was a problem hiding this comment.
These skills include {{partial:...}}, but Symphony resolves partials relative to the skill file's directory as skills/_partials/<name>.md. This PR added the partials at workflows/factory/_partials, so the loader will return {:missing_partial, ...} for each referenced partial and the catalog will not publish the skill; prompt: skill "researcher" then fails at runtime. Move the shared partials under skills/_partials or update the loader to support this pack-level location.
| scope <- agent { | ||
| engine: codex | ||
| model: "gpt-5.3-codex" | ||
| effort: high | ||
| permissions: read_only | ||
| prompt: skill "researcher" |
There was a problem hiding this comment.
This agent omits location, so the envelope defaults to :local. The runtime only acquires a checkout/cwd for :host or :ixvm; for :local/:room, workspace_cwd returns :error and RoomEngineClient fails the turn with :missing_cwd. The same omission appears on the other new agents, so the factory workflows fail before any prompt runs unless each agent declares a real placement or local cwd support is added.
| End your turn by emitting EXACTLY ONE fenced JSON object as the final block of | ||
| your message. Nothing after it. Downstream `when` gates read its fields, so the | ||
| keys and types must match what your role declares. Booleans must be real JSON | ||
| booleans, not strings. |
There was a problem hiding this comment.
The contract assumes the agent's fenced JSON becomes node output fields, but the current synchronous engine response persisted by Symphony contains only thread/event/cost metadata and drops message text. As a result ${scope.scoped}, ${impl.completed}, and reviewer fields are absent even when the model emits the requested JSON, so the Worker/review gates cannot make decisions from agent output. Add fail-closed parsing of the final JSON into node output, or use a mechanism that already produces structured output, before relying on these fields.
| # Worker: implement the scoped plan on a branch and open a PR (only if scoped). | ||
| impl <- when ${scope.scoped} { | ||
| agent { | ||
| engine: codex | ||
| model: "gpt-5.3-codex" | ||
| effort: high | ||
| permissions: workspace_write | ||
| prompt: skill "worker" { plan: ${scope.plan}, affected: ${scope.affected_crates} } | ||
| } | ||
| } | ||
|
|
||
| # Review round 1 (only if the Worker completed and opened a PR). | ||
| review1 <- when ${impl.completed} { | ||
| agent { |
There was a problem hiding this comment.
After a closed when, the bound name is not added to the interpreter environment. If scope.scoped is false, impl is never bound but the next top-level statement still evaluates when ${impl.completed}; that becomes a gate placeholder with no dependency, which placeholders never run or retire, so the run deadlocks instead of ending as denied/skipped. The same pattern repeats for later review rounds after an approval. Keep later rounds inside the gate that creates their inputs, or emit explicit false/default outputs for skipped branches.
| 1. Read the `plan`. If you disagree with it (wrong, unsafe, or it needs more | ||
| research), you are allowed to **deny** or **delegate** - do not build | ||
| something you think is wrong. Comment your reasoning on the issue. | ||
| - deny -> label `factory:denied`. | ||
| - delegate -> label `factory:scope-needed` again with a comment on what the | ||
| Researcher must resolve. |
There was a problem hiding this comment.
The delegate path re-applies the same factory:scope-needed trigger label from inside the current Linear-triggered run. The Linear webhook controller dedupes events for issues that already have an active pending/running run, so that label update is ignored; when this run exits there is no second webhook to start Researcher. Delegate needs to enqueue after the current run is no longer active, use a distinct handoff path, or invoke Researcher within the current workflow.
| --- | ||
| description: Propose one high-value improvement or feature idea for ix and file it as a Linear issue, deduped against existing work. | ||
| tools: [linear_graphql, web_search] | ||
| --- | ||
|
|
||
| You are the Idea Generator in an autonomous SDLC factory for the **ix** repo | ||
| (custom rust-vmm VMM, ix-server RPC, two isolated regions, IPv6-as-identity). | ||
|
|
||
| Goal: produce **one** concrete, valuable idea per run and file it as a Linear | ||
| issue. Quality over quantity. One good idea beats three vague ones. | ||
|
|
||
| Steps: | ||
| 1. Inspect the checked-out ix workspace. Look for: missing capabilities, rough | ||
| edges, places competitors/prior art do better. Use web search for prior art | ||
| and competitor features when it sharpens the idea. | ||
| 2. **Dedup before filing.** Query Linear for open AND closed issues with similar | ||
| titles/scope (use `linear_graphql`). If the idea already exists or was | ||
| already rejected, DO NOT file it - return `created: false` and stop. | ||
| 3. If it is genuinely new, create a Linear issue: | ||
| - Clear problem statement (what + why, from a user/system perspective). | ||
| - A short solution sketch. | ||
| - Label it `factory:scope-needed` (this is what triggers the pipeline). | ||
| - Team: Engineering. Priority: your best judgment. |
There was a problem hiding this comment.
This feeder requires web_search and linear_graphql to dedupe and create Linear issues, but in the current engine path skill frontmatter is only loaded with the prompt metadata; RoomEngineClient submits tools: [] unless run options explicitly supply tools, and the runtime never derives that from skill.tools. These agents therefore cannot perform their primary action even with LINEAR_API_KEY configured. Wire skill.tools into the turn request or avoid depending on these tool names.
Blast radius
pie showData title Rebuilt checks by category
"image" : 15
"rust" : 2
"site" : 2
"agent" : 1
"blast" : 1
"eval" : 1
"lint" : 1
"symphony" : 1
flowchart LR
c0["ix-mcp"]
c1["blast-radius-test"]
c2["agent-context"]
c3["lint"]
c4["site-test"]
c5["site-case-tests"]
c0 --> k0["agent-context"]
c0 --> k2["eval"]
c0 --> k3["image-development-base"]
c0 --> k4["image-kernel-dev"]
c0 --> k5["image-minecraft"]
changed checks (24)
|
1 participant
What
A Symphony pack implementing the autonomous agent workflow from ENG-2921: five agent roles plus an escalate escape hatch that take an idea from conception to a green, auto-merged PR on ix.
How it works
idea-generator,cleaner) scan ix + prior art and file deduped Linear issues labelledfactory:scope-needed.task-pipelinetriggers per issue on that label: Researcher scopes -> Worker implements on a branch + opens a PR -> bounded Reviewer/Worker correction loop -> auto-merge whenbuildis green.whenbody holds exactly one statement, so the Worker<->Reviewer loop is an unrolled, flatwhen-cascade ordered by data deps.MAX_ROUNDS = 3is a structural convergence cap (noreview4exists). Open-ended macro arrows are cross-run via Linear labels.ix-playbook-agentbot path (squash via merge queue;mainauto-deploys), inheriting its.github/**/CODEOWNERS guardrails.Validation
All three
.symwere run through the real parser + interpreter + materializer (pure Elixir, no hex deps):cleaner.symidea-generator.symtask-pipeline.symThe
1 agent + 7 gatesshape confirms the lazy forward cascade and that envelope validation passes.Scope / notes
SYMPHONY_PACK_DIR=packages/symphony/workflows/factory.packages/symphony/docs/factory-workflow.md.${node.field}, tool names (linear_graphql,web_search), and widening the ix-sideix-playbook-agentauto-merge scope from playbook-only to code.Closes ENG-2921.
Note
Add autonomous SDLC factory Symphony workflow pack for ix
packages/symphony/workflows/factory/that implements a fully autonomous software development lifecycle for theindexable-inc/ixrepository.task-pipeline.symtriggers on thefactory:scope-neededLinear label and runs a gated scope → implement → review pipeline with a bounded 3-round reviewer/worker-fix correction loop and an escalation path on failure.idea-generator.symruns daily andcleaner.symruns every 4 hours to file deduplicated idea and cleanup issues respectively.Macroscope summarized a478e81.