Ss Coding Workflow

Coding Workflow

Chains the existing ss-* skills into a single execute-and-deliver flow:

ss-create-branch → ss-coding (built-in review) → review-acceptance loop → ss-create-pr

Core principle: thin orchestration. This skill never writes code itself. It calls the other skills in order, passes outputs along, and reports progress.

Review is not run twice. ss-coding already enforces a post-coding review internally and returns a verdict. This workflow never invokes ss-code-review separately — the review-acceptance loop below is driven entirely by that verdict.

Unlike ss-feature-workflow, there is no proposal or planning step here, and no manual gate by default — the input is either an already-prepared plan or a change instruction handed straight to coding.

When to Run

This is end-to-end orchestration: it creates a branch, spawns coding agents, runs reviews, and delivers the result. Only run it when the user explicitly asks for this skill or clearly wants the full execute-to-delivery pipeline. For an ordinary edit, a question, or a single-step request, do the work directly (or point to the one relevant ss-* skill).

Inputs

• Execution input, one of:
  • a path to an existing execution plan (produced by ss-plan), or
  • a plain-language change instruction (e.g., "extract OrderService's helper methods into OrderUtils"). If ambiguous which one it is, ask; when deciding autonomously, treat text with a recognizable task/file structure as a plan, otherwise as a change instruction.
• Delivery mode — full (default) or lite. See "Delivery Mode" below.
• Explicit branch name (optional) — used verbatim by ss-create-branch, skipping name derivation. Set this when a caller (such as ss-multi-repo-workflow) needs one unified branch name across repositories.
• Skip-gates / decide-autonomously — kept for consistency with the other workflows; this workflow has no gate by default, so skip-gates has little effect here.
• Worktree preference — forwarded to ss-create-branch.

Derive one stable input identity before pre-flight and carry it verbatim into branch/commit/PR references and result evidence:

• plan input → compact JSON with keys in this exact order and no extra whitespace: {"kind":"plan","path":"<canonical repo-relative POSIX path>","problem_id":<JSON string containing the exact persisted Problem ID, or null>};
• direct instruction → compact JSON with keys in this exact order and no extra whitespace: {"kind":"instruction","sha256_12":"<first-12-lowercase-hex-of-SHA-256>"}, hashing UTF-8 text after trimming, lowercasing, and collapsing each whitespace run to one space.

Use normal JSON escaping inside strings and compare the parsed object plus canonical serialization verbatim. Never derive identity from a later summary, branch name alone, or whichever artifact is newest. Repository-runner invocation does not weaken these checks.

Delivery Mode: full vs. lite

The quality gates — TDD, multi-agent code review, verification — are identical in both modes. Only the start and finish change:

Mode resolution: an explicit mode input wins; otherwise ask once at the start and use that answer for the rest of the run. Never persist the choice to a project config file.

Multi-Repo Routing

Checked once, before pre-flight — this workflow's input is the plan, so multi-repo nature is fully decidable at the entrance. Full rules in ../ss-references/multi-repo-detection.md (this table is a condensed copy; the reference file wins on conflict).

On hand-off: pass the original input, any existing artifacts, and the active automation flags to ss-multi-repo-workflow; print one line explaining the hand-off and terminate this workflow.

Iron Rules

Violating any of these means stopping and explaining to the user:

1. The orchestrator does no real work — only call other ss-* skills; never edit source directly.
2. Pre-flight checks ask, they don't abort — apply the autonomous default only when deciding autonomously.
3. No gate for review — run the loop instead — see "Review-Acceptance Loop". There is also no gate before delivery.
4. Resumable — on start, probe for existing artifacts and skip completed steps.
5. Never relax safety guardrails — automation flags only affect questions and approvals; they never bypass the hard rules built into the underlying skills.

Pre-flight Checks

Run once on start. Any unmet check defaults to asking the user, not terminating.

No source-fetching tool is needed when the input is a local plan file or plain text — the check only triggers for a URL input, same as the other workflows.

Resume Detection

Process

1. Pre-flight and resume, then detect the input type (plan vs. change instruction).

2. ss-create-branch (full mode only):

   • If an explicit branch name was given, use it verbatim and skip derivation.
   • Otherwise name the branch from the plan title, or from the change instruction's summary with a prefix that matches its nature (refactor → refactor/, small feature → feat/, small fix → fix/); ask if ambiguous. Forward the worktree preference. Record the branch name and worktree path.

3. ss-coding — full execution for a plan input; inline fast mode (handles small, low-file-count changes without a full plan) for a change instruction. It already covers TDD and test verification, and enforces its own post-coding review before returning a verdict.

4. Review-acceptance loop — adjudicate findings and feed valid fixes or invalid-finding reasons back into ss-coding until its fresh built-in review returns literal APPROVED, or escalate.

5. Delivery (full mode) — call ss-create-pr to archive any delta spec, rebase, and open the PR; report the link, plus a reminder that ss-cleanup removes the worktree and branch after merge. (lite mode) — commit with a conventional-commits message (push optional), then print a summary of files changed, tests run, and the review verdict. Resolve the final delivered commit after all delivery/rebase work. If it differs from the last test-verified ref, rerun full verification against that commit; if conflict resolution or delivery changed content, rerun ss-coding so its built-in review approves the delivered diff.

6. Structured result block — always end the final report with this block. A repository runner under ss-multi-repo-workflow returns it to the parent; in ordinary interactive use it is just a few extra lines. Values for object/array fields are compact valid JSON on one line:

   ---SS-RESULT---
   status: SUCCESS | FAILED | BLOCKED
   input_identity: {"kind":"plan","path":"docs/plans/example.md","problem_id":null}
   branch: <branch-name>
   delivery: <pr-url | local-commit:<7-64 lowercase hex SHA> | none>
   delivered_ref: <7-64 lowercase hex SHA | none>
   review_verdict: <APPROVED | NEEDS_CHANGES | CRITICAL_ISSUES | NOT_RUN>
   test_verified_ref: <7-64 lowercase hex SHA | none>
   test_evidence: {"command":"<full verification command or none>","result":"passed|failed|not-run","summary":"<concise result>"}
   tasks_completed: <N/N>
   changed_files: ["path/one", "path/two"]
   commands_run: [{"command":"<exact command>","result":"passed|failed|not-run","summary":"<concise result>"}]
   residual_risks: ["<remaining non-blocking risk>"]
   clean_worktree: <true | false>
   blocker: <one-line reason when status is not SUCCESS, else omit>
   ---END-SS-RESULT---
   

   Emit exactly one block, as the last thing in the final message, with every field except the conditional blocker present. Use [], none, and not-run rather than omitting evidence on failure paths.

   status: SUCCESS is valid only when all of these hold: input_identity exactly matches this invocation; delivery is an open PR or explicit local commit; delivered_ref is the final commit delivered at that location; review_verdict is APPROVED; test_verified_ref == delivered_ref; test_evidence.result is passed; its exact full verification command appears as a passed entry in commands_run; every task is complete; changed_files is non-empty; all residual risks are disclosed; clean_worktree is true; and blocker is absent. Any contradiction or missing evidence changes the status to FAILED or BLOCKED. Every escalation/abort path emits the block with blocker filled in.

Review-Acceptance Loop

After ss-coding returns its verdict, this workflow sets no gate and does not pause:

1. Read the verdict.
2. No unresolved findings (approved) → go to delivery.
3. Unresolved findings remain → judge each one: valid goes on the to-fix list; invalid (false positive, out of scope, by design, or suggestion-only) is recorded with a reason and left unfixed.
   • Scope-reduction findings get no leniency: a finding reporting unimplemented requirements, stubbed logic presented as complete, or scope silently deferred ("phase 2", "MVP", "simplified for now") must never be judged invalid as "by design" or "out of scope" unless the user explicitly approved that scope adjustment earlier — cite that approval.
4. All judged invalid → pass the complete adjudication record back to ss-coding and require a fresh built-in review of the current diff. Only a literal APPROVED result proceeds to delivery; if approval cannot be obtained, emit FAILED/BLOCKED rather than rewriting the prior verdict.
5. Valid findings remain → call ss-coding again for just those findings → back to step 1.

Convergence protection:

• Cap workflow-level rounds at 3, independent of ss-coding's internal rounds; if exceeded, escalate with the latest verdict plus the fixed/unresolved lists.
• Track recurrence at the workflow level: the same finding judged valid across 2 consecutive workflow rounds after being "fixed" is a stall — escalate. Don't rely on ss-coding's internal counter; it doesn't accumulate across calls.
• Record every "invalid" judgment and its reason. Never force a finding to "invalid" just to pass.

Edge Cases & Error Handling

Examples

Run ss-coding-workflow on docs/plans/2026-06-26-refund.md
Run ss-coding-workflow with "Extract OrderService's helper methods into OrderUtils"
Run ss-coding-workflow with "Add 2-decimal-place amount validation to /api/refund", deciding autonomously