This is the complete input contract for every harness workflow definition in this repository.
It is derived from the workflow definitions in workers/workflows/: an input is required
when it is listed in inputParameters and has no value in inputTemplate; every other listed
input is optional, with the shown default. Internal workflows are included for operators and
tests, but normally only the user-facing workflows are started directly.
Empty strings mean “use the normal runtime default” for that field. Model fields are blank by
default so the selected model profile can resolve them; do not replace them with a provider
default unless you are deliberately overriding the profile. *PromptTemplateSource records the
provenance of its companion override.
The following optional policy envelope is available on every workflow and is listed here once rather than repeated in every optional-parameter line below:
modelProfiledefaults to""; pass a profile name (for example,anthropic-standardoropenai-standard) or leave it blank for the configured default. TUI launches snapshot/modelsin that case; raw API/CLI callers retain bundled/repository defaulting.modelPolicydefaults to{}and carries the selected one-file user-policy snapshot.modelPolicySource,modelPolicySha256, andmodelsConfigdefault to""; they record the snapshot provenance and optionally select a contained repository policy path.modelOverridesdefaults to{}for explicit structured role overrides.
Every workflow resolves this envelope before agent work. The resulting role tier is carried into sub-workflows, dynamic forks, and scheduled children; workers record the selected profile, role, model, and source hashes in their task output. A nonblank legacy backend/model field remains an explicit override for that task role.
Required: repo, prNumber.
Optional: agent = ""; design = false; designDir = "docs/design"; designHumanApproval = true; designMaxIterations = 5; designMaxTurns = 250; designMaxBudgetUsd = 50; designModel = ""; designPromptTemplate = ""; designPromptTemplateSource = ""; fixPromptTemplate = ""; fixPromptTemplateSource = ""; maxApprovalRevisions = 2; maxBudgetUsd = 50; maxFixAttempts = 4; maxSubtasks = 4; maxTurns = 250; model = ""; openspecPlanAgent = ""; openspecPlanModel = ""; openspecMaxTurns = 500; openspecMaxBudgetUsd = 50; openspecHumanApproval = true; openspecMaxIterations = 5; planPromptTemplate = ""; planPromptTemplateSource = ""; allowAgentTestPlan = true; allowAgentAuthoredTests = true.
Addressing review feedback always routes through the full code_parallel decompose/fork/merge sub-workflow (the earlier engine choice between that and a single coding_agent session was removed — code_parallel was already the default, and the single-session path added a second, less capable code path with no independent value).
allowAgentTestPlan/allowAgentAuthoredTests are forwarded to both the initial verify pass and the address_pr_repair sub-workflow's own test_cycle call, so a PR-comment fix that touches a file with no directly-matching test (a shared helper, for example) is not permanently configuration_blocked — see test_cycle's entry below. Both default true here (unlike test_cycle/test_agent_fallback's own default of false) since this is the interactive PR-feedback loop where leaving a change permanently blocked on missing test-file mapping is worse than trying the agent fallback first; set either to false to restore the strict deterministic-only behavior.
Required: repoPath, candidateCommit, verification.
Optional: agent = ""; model = ""; maxTurns = 250; maxBudgetUsd = 50; fixPromptTemplate = ""; fixPromptTemplateSource = ""; allowAgentTestPlan = true; allowAgentAuthoredTests = true. This workflow returns remediation evidence only; it cannot push or comment. Defaults match address_pr, which is address_pr_repair's only caller and always forwards its own values explicitly; these defaults matter only for a standalone/direct invocation.
Required: repo, prNumber, repoPath, workspacePath, branch, candidateCommit.
Optional: feedback = ""; verificationState = "passed"; verification = {}; replyBody = ""; agentResult = ""; subtasks = []; totalTokens = 0; maxApprovalRevisions = 2; agent = ""; maxTurns = 250; maxBudgetUsd = 50; maxSubtasks = 4.
address_pr's human-approval gate for one already-verified candidate, extracted to its own sub-workflow to keep address_pr's own graph small. A verified candidate is never auto-published: the address_gate WAIT task always pauses for an explicit human approve/revise/stop decision, offering up to maxApprovalRevisions bounded, independently re-verified revision rounds first (each revise reruns code_parallel, marked optional: true so an infra failure there degrades to verification_blocked rather than failing the whole address_pr run). Resolves approvalState to exactly one of: approved, suppressed, blocked, verification_blocked, revision_exhausted (a still-pending state when maxApprovalRevisions is exhausted). address_pr's publication_gate only publishes when this resolves to approved.
Required: repoPath, candidateCommit, discoveryOutcome.
Optional: testMode = "targeted"; allowHeavySuites = false; includeBrowserTests = false; allowAgentTestPlan = false; allowAgentAuthoredTests = false; discoveryReason = ""; changedPaths = []; repairRoots = []; agent = ""; model = ""; maxTurns = 250; maxBudgetUsd = 50.
test_cycle's agent-assisted discovery/authoring fallback, extracted to its own sub-workflow to keep test_cycle's own graph small. A no-op that echoes candidateCommit/repairRoots straight back unless allowAgentTestPlan is true and discoveryOutcome is configuration_blocked. Tier 1: a read-only coding agent's proposed test plan, validated against this module's strictest gate (validate_agent_argv: the entrypoint must belong to a build system the repository actually shows evidence of, every path-like token must resolve to a real file, every claimed covered path must be a real changed path, and every changed file that looks like a test must be covered by some proposal). Tier 2, only reached when tier 1 also finds nothing and allowAgentAuthoredTests is true: a coding agent (pinned to the Claude backend, tools limited to Read/Grep/Glob/Write/Edit — no Bash) may author exactly one new test file, discovering the repository's own test convention itself rather than following any hardcoded per-language rule. The new file is never trusted on the agent's say-so: it must be the single new, test-shaped file (nothing else touched), must textually reference the changed code, and must pass a red/green check — failing when run against the candidate's pre-change baseline and passing against the real candidate change — before it is committed and treated as the resolved test plan. agentAuthoredTest/agentAuthoredTestPath in the output report whether this happened. It owns no branch: an accepted authored test is committed to whatever branch repoPath is already on.
Required: repoPath, candidateCommit.
Optional: testMode = "targeted"; maxFixAttempts = 4; testCommands = []; testPlanTemplate = ""; testPlanTemplateSource = ""; allowHeavySuites = false; includeBrowserTests = false; allowAgentTestPlan = false; allowAgentAuthoredTests = false; priorVerification = {}; allowedWriteRoots = []; agent = ""; model = ""; maxTurns = 250; maxBudgetUsd = 50; repairPromptTemplate = ""; repairPromptTemplateSource = "".
Runs the repository's tests for one exact candidate commit and fixes what fails. testMode = "targeted" derives an exact-test or affected-unit plan from the candidate's diff; testMode = "full" runs the repository-wide suite and is the mode to use before publishing a PR — full mode reruns the whole suite on every repair iteration, so reserve it for a final pre-publication gate, not a review-feedback loop. allowHeavySuites is honoured only in full mode and is what permits an integration or end-to-end target; it is never inferred. includeBrowserTests is also full-mode-only and adds the repository's dedicated Playwright/Cypress script when one exists — it never triggers a browser install: if binaries are not already provisioned on the host, this reports a diagnostic in rejectedCandidates rather than blocking the run. testCommands and testPlanTemplate are the two forms of the operator override and both outrank every other source: testCommands is a structured argv list; testPlanTemplate is the same shape as an inline JSON string, for a caller that only has text to pass, and is rejected outright if it starts with @ (the templating convention for resolving from a file in the checkout, which the candidate controls) rather than falling back to it. When both are supplied, testCommands wins, since it cannot suffer a JSON-parsing surprise. Below the operator override, the precedence is repository guides, then .conductor-code/verification.json, then build-system inference, and — only when every one of those comes up empty — test_agent_fallback (see above), gated on allowAgentTestPlan/allowAgentAuthoredTests. The agent is invoked at most once per candidate SHA and is never re-proposed after a repair; its accepted plan is carried forward the same way a priorVerification obligation is. Repair writes are restricted to allowedWriteRoots, falling back to the discovered changed paths.
The budget is maxFixAttempts fixes and one more test run than that, capped at 4 and 5. Because each loop iteration is fix-then-verify, the final run is always a verdict and no fix is ever left unvalidated. The loop also stops early when a fix produces no new commit, since re-running would repeat the identical result.
This workflow never fails. Every terminal condition completes with a testCycleState the caller branches on: tests_passed, no_tests_required, tests_failed_after_fix_budget, tests_failed_fix_unavailable, command_discovery_blocked, runtime_unavailable, candidate_commit_missing, or verifier_worker_unavailable. testsPassed is the boolean for a SWITCH gate.
One caveat: never failing is not the same as always completing. There are no execution deadlines anywhere in this harness, so if the isolated verification worker is not polling, test_discover and test_run never start and the workflow stays RUNNING rather than reporting verifier_worker_unavailable — that state only fires when a task actually fails. A caller that needs a bounded answer must watch queue depth or verification_health, not this workflow.
It owns no branch: fixes are committed to whatever branch repoPath is checked out on.
Required: repoPath, groupIds.
Optional: modelBuilder = ""; maxBudgetUsd = 50.0; preflight = false.
Required: repoPath, repo, headRepo, prNumber, branch, expectedHeadSha, candidateCommit.
Optional: replyBody = ""; reconcileModel = ""; reconcileMaxResolutionAttempts = 3; reconcileMaxBudgetUsd = 50.0; allowAgentTestPlan = true; allowAgentAuthoredTests = true. The caller must approve and locally verify the candidate first. This workflow performs the one candidate guard, remote-head guard, exact commit push, and success comment -- it completes once the branch is updated and does not wait on or poll CI; GitHub's own PR page reports check status independently.
When the remote-head guard finds the PR branch has moved since the candidate was checked out, it does not simply give up: it merges the remote's new commits in (reconcile_branch_drift, agent-resolved on conflict, bounded by reconcileMaxResolutionAttempts). A clean merge (the remote's changes never touched the candidate's own changed files) is trusted and pushed without re-verification; a resolved conflict is genuinely new content and is re-verified through test_cycle before it is ever trusted for publication (publicationState: "conflict_verification_failed" if that re-verification fails). Only resolution failing outright still reports publicationState: "branch_drift" with nothing pushed. reconcileState/reconcileTokens/reconcileCostUsd in the output report what happened, if anything.
Required: repo, kind, childWorkflow, number, revision, attempt.
Optional: childWorkflowVersion = 1; approvalMode = "human"; agent = "claude"; model = ""; judgeAgent = "claude"; judgeModel = ""; judgeMaxTurns = 50; judgeMaxBudgetUsd = 5; maxApprovalRevisions = 2; verificationProfile = ""; reviewPromptTemplate = ""; reviewPromptTemplateSource = ""; fixPromptTemplate = ""; fixPromptTemplateSource = ""; designPromptTemplate = ""; designPromptTemplateSource = ""; designJudgePromptTemplate = ""; designJudgePromptTemplateSource = ""; planPromptTemplate = ""; planPromptTemplateSource = ""; codePromptTemplate = ""; codePromptTemplateSource = ""; approvalJudgePromptTemplate = ""; approvalJudgePromptTemplateSource = "".
Required: repo, kind, number, revision. Optional: none.
Required: repoPath, task, wave.
Optional: agent = "claude"; model = ""; maxTurns = 500; maxBudgetUsd = 50; resumeSessionId = ""; feedback = ""; specContextPath = ""; contextPaths = []; codePromptTemplate = ""; codePromptTemplateSource = "".
Required: repoPath, instruction.
Optional: callerWorkflow = "code_parallel"; branchRunId = ""; changeBranch = ""; openspecPlanModel = ""; openspecMaxTurns = 500; openspecMaxBudgetUsd = 50; openspecHumanApproval = true; openspecMaxIterations = 5; codeModel = ""; codePromptTemplate = ""; codePromptTemplateSource = ""; maxTurns = 500; maxBudgetUsd = 50; maxFixAttempts = 4; precomputedPlan = {}; specContextPath = ""; usePrecomputedPlan = false; design = false; designDir = "docs/design"; designHumanApproval = true; designMaxIterations = 5; designMaxTurns = 500; designMaxBudgetUsd = 50; designModel = ""; designPromptTemplate = ""; designPromptTemplateSource = ""; inPlace = false; workspacePath = ""; contextPaths = []; allowAgentTestPlan = true; allowAgentAuthoredTests = true.
After merging the subtask branches this runs test_cycle in targeted mode, so the candidate it hands back has been tested and, where possible, repaired. maxFixAttempts bounds that repair loop. The result is diagnostic: testCycleState and testsPassed are returned for the caller to gate on, and a red suite never withholds the candidate commit. Each subtask also runs its own targeted test_cycle with a single fix attempt before its branch is merged. allowAgentTestPlan/allowAgentAuthoredTests (both default true here, unlike test_cycle/test_agent_fallback's own default of false) are forwarded to the post-merge test_cycle call so a targeted verification that cannot map a changed file to an exact test tries the agent-assisted fallback before ending command_discovery_blocked.
callerWorkflow is forwarded to design_docs (when design:true) so its human-review checkpoint
correctly identifies the real top-level launcher; a caller that embeds code_parallel
(issue_to_pr, address_pr) passes its own name here instead of the default.
Required: worktreePath, workflowId, loopId, prompt, modelResolution, bestCommit.
Optional: promptTemplate = ""; promptTemplateSource = ""; maxTurns = 250; maxBudgetUsd = 50.0; checks = []; findings = []; accepted = false; round = 1; maxRounds = 8; plateauCount = 0.
Required: repoPath, name, prompt, model.
Optional: promptTemplate = ""; promptTemplateSource = ""; templateKey = "code"; promptContext = {}; maxTurns = 250; maxBudgetUsd = 50.0; specContextPath = ""; contextPaths = []; allowedTools = []; allowedWriteRoots = []; allowAgentTestPlan = true; allowAgentAuthoredTests = true.
allowAgentTestPlan/allowAgentAuthoredTests (both default true, unlike test_cycle/test_agent_fallback's own default of false) are forwarded to this subtask's own single-fix-attempt test_cycle call, same rationale as code_parallel above.
Required: repoPath, instruction, callerWorkflow.
Optional: contextPaths = []; maxTasks = 25; planAgent = ""; planModel = ""; planPromptTemplate = ""; planPromptTemplateSource = ""; planMaxTurns = 500; planMaxBudgetUsd = 50; planMaxIterations = 5.
Author and get human approval for a dependency-DAG implementation plan (id/description/dependsOn/
files/acceptanceCriteria/checks per task), extracted from feature_campaign's own plan_loop so
it can be reused -- unlike design_loop, there was no existing second implementation to
reconcile with (code_parallel's own "plan_loop" is a fully-automatic validator-repair loop with
no human involved, a different concept entirely), so this is a standalone extraction that
faithfully preserves plan_loop's prior behavior: adjustable per-iteration turn/budget via the
same checkpoint-decision shape feature_campaign's other checkpoints use, and
continue/revise/adopt_edits/stop actions. Never hard-fails: an unapproved plan resolves
with approved:false (and stopped distinguishing an explicit stop from exhausting
planMaxIterations) for the caller to act on. callerWorkflow must be the caller's own
top-level workflow name, matching design_docs'/pr_draft_approval's same requirement.
allowedWriteRoots (the approved plan's own declared file scope) is surfaced in output since a
sub-workflow's internal task refs -- unlike a plain DO_WHILE loop's -- aren't reachable from the
caller once the loop ends.
Required: instruction, repoPath, callerWorkflow.
Optional: designAgent = ""; designDir = "docs/design"; designMaxBudgetUsd = 50; designMaxIterations = 5; designMaxTurns = 500; designModel = ""; designPromptTemplate = ""; designPromptTemplateSource = ""; humanApproval = true; failClosed = true; contextPaths = [].
Shared by every caller that needs a design phase: code_parallel (and so issue_to_pr and
address_pr, both of which embed code_parallel) and feature_campaign both delegate their
design phase here instead of each maintaining their own loop. callerWorkflow must be the
caller's own top-level workflow name -- required, no default, since design_docs is reachable up
to two levels deep (issue_to_pr -> code_parallel -> design_docs) and the TUI's approval
dispatch needs the real top-level name, not an intermediate sub-workflow's. humanApproval picks
a human WAIT checkpoint (which may also request a bounded turn/budget
increase, or explicitly stop early, via the same checkpoint-decision shape feature_campaign's
own checkpoints use) or an automated agent design-judge session. failClosed (default true,
matching every existing caller's prior behavior) TERMINATEs the whole workflow FAILED when a
design is never approved; feature_campaign passes false so an unapproved design degrades its
own campaign outcome instead of failing the entire run.
Required: repoPath, instruction.
Optional: contextPaths = []; model = ""; maxTurns = 500; maxBudgetUsd = 50.
Required: repoPath, instruction, changeBranch.
Optional: openspecPlanModel = ""; openspecMaxTurns = 500; openspecMaxBudgetUsd = 50.0; openspecHumanApproval = true; openspecMaxIterations = 5.
Required: repoPath, changeName, goal.
Optional: feedback = ""; model = ""; maxTurns = 500; maxBudgetUsd = 50.0.
Required: repoPath, changeName, artifact, goal.
Optional: feedback = ""; model = ""; maxTurns = 500; maxBudgetUsd = 50.0.
Required: none.
Optional: repo = ""; repoPath = ""; workspacePath = ""; instruction = ""; issueNumber = 0; keepWorktree = true; changeBranch = ""; inPlace = false; contextPaths = []; createPr = false; requirePrApproval = false; maxApprovalRevisions = 2; prBase = "main"; prTitle = ""; prBody = ""; prDraft = false; designDir = "docs/design"; designAgent = "claude"; designHumanApproval = true; designModel = ""; planAgent = "claude"; planModel = ""; codeAgent = "claude"; codeModel = ""; reviewAgent = "claude"; reviewModel = ""; maxTurns = 500; maxBudgetUsd = 50; maxFixAttempts = 4; maxTasks = 25; maxParallelism = 6; maxWaves = 20; designMaxRevisions = 5; planMaxRevisions = 5; finalMaxRevisions = 3; useImportedPlan = false; importedPlan = {}; importedDesignLocation = ""; specContextPath = ""; designPromptTemplate = ""; designPromptTemplateSource = ""; planPromptTemplate = ""; planPromptTemplateSource = ""; codePromptTemplate = ""; codePromptTemplateSource = ""; reviewPromptTemplate = ""; reviewPromptTemplateSource = ""; revisionPromptTemplate = ""; revisionPromptTemplateSource = ""; allowAgentTestPlan = true; allowAgentAuthoredTests = true.
allowAgentTestPlan/allowAgentAuthoredTests (both default true, unlike test_cycle/test_agent_fallback's own default of false) are forwarded to the final full-suite test_cycle sub-workflow, same rationale as code_parallel/address_pr above.
The design phase itself is now the shared design_docs sub-workflow (also used by
code_parallel/issue_to_pr) rather than a separate inline implementation. designHumanApproval
(default true, matching prior behavior) picks a human WAIT checkpoint or an automated
agent design-judge session; a design that's never approved degrades the campaign's own
outcome to incomplete rather than failing the whole run (design_docs's failClosed:false).
The final verification phase itself is now the standalone final_verification sub-workflow
(no other workflow needs an analogous human-reviewed final-verification checkpoint, so this is
extraction-only, not a merge). It always runs when the campaign's own outcome is still running
by the time the final phase is reached -- an earlier phase's explicit stop is never silently
overridden. finalMaxRevisions (default 3, matching prior behavior) bounds its review/checkpoint
rounds before an unresolved run resolves incomplete.
requirePrApproval (only meaningful when createPr is also true) gates the verified candidate's
PR draft behind pr_draft_approval -- the same human-approval-with-bounded-revisions sub-workflow
issue_to_pr's approvePr uses -- before it's ever pushed. false (the default) publishes
automatically, matching prior behavior. maxApprovalRevisions (default 2, only meaningful
alongside requirePrApproval) bounds how many human-requested revision rounds pr_draft_approval
will run before a still-pending approval is relabeled revision_exhausted.
Two independent choices, each "at least one of," both runtime-enforced by campaign_defaults/
campaign_workspace (workspace_prepare) rejecting the run at start if unmet -- neither is
schema-required because either alone is legitimate:
- What to build:
instruction(free text) orissueNumber(a GitHub issue number, requiresrepo). WhenissueNumberis set, the campaign fetches that issue's title/body -- plus any issue/PR/doc/CI links the body itself references, resolved the same untrusted-evidence waypr_reviewresolves links in PR feedback -- and uses that as the effective instruction instead. The originating issue is surfaced in output (issueNumber/issueTitle/issueUrl), and whencreatePris also true the PR body gets a trailingCloses #<n>. - Where to work:
repo(a git URL orowner/nameslug, resolved the same waypr_review/address_pr/issue_to_prresolve it) clones into a fresh temp workspace and works from there;repoPathuses an existing local checkout directly;workspacePathinherits an already-prepared worktree (e.g. from a parent campaign). Passing a git URL asrepoPathfails fast with a clear error instead of being treated as a local path -- userepofor that case.
These two choices are orthogonal: an issue can drive implementation against either a freshly
cloned repo or an existing local repoPath checkout, and a plain instruction works the same
way against either source.
prBody is a summary source, not a free-form PR template. Publication normalizes it to one
## Summary paragraph; when empty, the effective instruction (from instruction or the fetched
issue) supplies the summary.
Required: repoUrl, instruction.
Optional: changeBranch = "conductor-harness-change"; base = ""; agent = "claude"; model = ""; codePromptTemplate = ""; codePromptTemplateSource = ""; prTitle = ""; maxTurns = 300; maxBudgetUsd = 50.0; allowAgentTestPlan = true; allowAgentAuthoredTests = true.
allowAgentTestPlan/allowAgentAuthoredTests (both default true, unlike test_cycle/test_agent_fallback's own default of false) are forwarded to the full-suite verify test_cycle call, same rationale as code_parallel/address_pr above.
Required: repoPath, callerWorkflow, plan, remainingTaskIds.
Optional: allowedWriteRoots = []; codeAgent = ""; codeModel = ""; codePromptTemplate = ""; codePromptTemplateSource = ""; reviewAgent = ""; reviewModel = ""; reviewPromptTemplate = ""; reviewPromptTemplateSource = ""; revisionPromptTemplate = ""; revisionPromptTemplateSource = ""; maxTasks = 25; maxParallelism = 6; maxWaves = 20; implementationMaxTurns = 500; implementationMaxBudgetUsd = 50; specContextPath = ""; contextPaths = [].
Schedule and implement an approved dependency-DAG plan in dependency-ordered, file-disjoint
parallel waves, extracted from feature_campaign's own implementation_loop so it can be
reused -- like dag_plan_approval, this is a standalone extraction: code_parallel's own
parallel-implementation mechanism is a single non-staged fan-out/fan-in with no per-wave human
checkpoint, a different concept entirely, so there's nothing to merge with. Each wave forks one
campaign_subtask per ready task, joins, integrates, reviews, and pauses at a human checkpoint
(continue/revise/adopt_edits/stop, with the same adjustable turn/budget mechanism
design_docs/dag_plan_approval use). Never hard-fails: resolves with outcome "running"
(plan fully implemented, or maxWaves reached) or "incomplete" (a human explicitly stopped).
callerWorkflow must be the caller's own top-level workflow name, matching
design_docs'/dag_plan_approval's same requirement.
Required: repoPath, callerWorkflow, instruction.
Optional: branch = ""; allowedWriteRoots = []; reviewAgent = ""; reviewModel = ""; reviewPromptTemplate = ""; reviewPromptTemplateSource = ""; finalMaxIterations = 3; finalMaxTurns = 500; finalMaxBudgetUsd = 50; specContextPath = ""; contextPaths = [].
A read-only-by-default review agent (with authority to reconcile direct worktree edits and fix
actionable checkpoint feedback) verifies the campaign against its instruction, then a human
reviews via the final_checkpoint WAIT task (continue/revise/adopt_edits/stop, with the
same adjustable turn/budget mechanism design_docs/dag_plan_approval/implementation_waves
use), extracted from feature_campaign's own final_loop so it can be reused -- neither
code_parallel nor issue_to_pr has an analogous human-reviewed final-verification checkpoint,
so this is a standalone extraction, not a merge. Never hard-fails: resolves outcome "verified"
(explicit continue, committed) or "incomplete" (an explicit stop, or finalMaxIterations
reached without either). callerWorkflow must be the caller's own top-level workflow name,
matching design_docs'/dag_plan_approval's/implementation_waves's same requirement.
Required: repo.
Optional: issueLabel = "conductor:auto"; approvalMode = "human"; agent = "claude"; model = ""; judgeAgent = "claude"; judgeModel = ""; judgeMaxTurns = 50; judgeMaxBudgetUsd = 5; maxApprovalRevisions = 2; verificationProfile = ""; designPromptTemplate = ""; designPromptTemplateSource = ""; designJudgePromptTemplate = ""; designJudgePromptTemplateSource = ""; planPromptTemplate = ""; planPromptTemplateSource = ""; codePromptTemplate = ""; codePromptTemplateSource = ""; approvalJudgePromptTemplate = ""; approvalJudgePromptTemplateSource = ""; maxNew = 1; maxActive = 1.
Required: repo, issueNumber.
Optional: base = "main"; approvePr = false; design = false; designDir = "docs/design"; designHumanApproval = true; designMaxIterations = 5; designMaxTurns = 500; designMaxBudgetUsd = 50; designModel = ""; designPromptTemplate = ""; designPromptTemplateSource = ""; maxApprovalRevisions = 2; openspecHumanApproval = true; openspecMaxIterations = 5; codePromptTemplate = ""; codePromptTemplateSource = ""; maxTurns = 300; maxBudgetUsd = 50; allowAgentTestPlan = true; allowAgentAuthoredTests = true.
allowAgentTestPlan and allowAgentAuthoredTests forward to the final full-suite test_cycle sub-workflow exactly as documented under test_cycle above; both now default true here too (matching address_pr/code_parallel/feature_campaign/github_demo), since leaving a change permanently blocked on missing test-file mapping is worse than trying the agent fallback first; the resulting agentAuthoredTest forces prDraft to true as defense in depth even when delivery otherwise passed.
approvePr was previously declared but never read (the gate was hardcoded to skip it) -- now wired for real: true routes the verified candidate's not-yet-published PR draft through pr_draft_approval (see below) before it's ever pushed, false (the default) auto-approves exactly as before. The TUI defaults it to true for interactive chat-launched runs.
Required: callerWorkflow, repoPath, workspacePath, branch, title, body, candidateCommit.
Optional: summary = ""; base = "main"; issueNumber = 0; subtasks = []; verificationState = "passed"; verification = {}; merged = []; conflicts = []; totalTokens = 0; totalCostUsd = 0; originalContext = ""; maxApprovalRevisions = 2; agent = ""; maxTurns = 250; maxBudgetUsd = 50; maxSubtasks = 4; codePromptTemplate = ""; codePromptTemplateSource = ""; openspecMaxIterations = 5.
issue_to_pr's human-approval gate for a not-yet-published PR draft (title/body/base/head,
no PR number yet), extracted to its own sub-workflow so any workflow that verifies a candidate
first and only wants to publish once a human is satisfied can reuse it (feature_campaign also
uses it, via its own requirePrApproval flag -- see below). A human must explicitly approve the
draft via the pr_gate WAIT task, optionally requesting up to maxApprovalRevisions bounded,
independently re-verified revisions (via code_parallel) first, resolving to exactly one of:
approved, suppressed, blocked, verification_blocked, revision_exhausted.
callerWorkflow must be the caller's own top-level workflow name (e.g. "issue_to_pr",
"feature_campaign") -- the TUI's approval dispatch branches on it, and without it a gate living
in this sub-workflow would report its own name instead of the caller's.
Required: workflowId, reason, failureStatus, failureTaskId, failedWorkflow.
Optional: none beyond the shared policy envelope. Conductor supplies every required input automatically when it invokes this as the failureWorkflow of issue_to_pr — it is not started by hand. When a run dies before reaching its publication gate, this pushes whatever branch survived and opens a draft PR carrying the failure reason, so failed work is delivered for review rather than stranded in a scratch worktree. PR creation is best-effort (optional): if it fails, the push still stands and the run reports pushed_no_pr.
Required: repoPath.
Optional: baseRemote = "origin"; baseBranch = "main"; agent = "claude"; model = ""; localReviewPromptTemplate = ""; localReviewPromptTemplateSource = ""; maxTurns = 250; maxBudgetUsd = 50.
Required: specSource, changeId. The target path is required unless the local-source-workspace option is enabled for an absolute local Git-backed source.
Optional: repoPath = ""; useSpecSourceWorkspace = false; workspacePath = ""; keepWorktree = true; specSourceType = "auto"; specRef = ""; specPath = ""; specWritebackRepo = ""; specWritebackRef = ""; executionMode = "auto"; instruction = ""; changeBranch = ""; archiveBranch = ""; base = "main"; agent = "claude"; model = ""; maxTurns = 500; maxBudgetUsd = 50; maxTasks = 25; maxParallelism = 6; maxWaves = 20; assessPromptTemplate = ""; assessPromptTemplateSource = ""; codePromptTemplate = ""; codePromptTemplateSource = ""; reviewPromptTemplate = ""; reviewPromptTemplateSource = ""; verificationPromptTemplate = ""; verificationPromptTemplateSource = "".
Required: repo.
Optional: approvalMode = "human"; agent = "claude"; model = ""; judgeAgent = "claude"; judgeModel = ""; judgeMaxTurns = 50; judgeMaxBudgetUsd = 5; maxApprovalRevisions = 2; verificationProfile = ""; fixPromptTemplate = ""; fixPromptTemplateSource = ""; approvalJudgePromptTemplate = ""; approvalJudgePromptTemplateSource = ""; maxNew = 2; maxActive = 2.
Required: repo, prNumber.
Optional: agent = ""; model = ""; approve = false; approvalMode = ""; reviewGuidance = ""; maxInvestigationPasses = 5; reviewPromptTemplate = ""; reviewPromptTemplateSource = ""; reviewInvestigationPromptTemplate = ""; reviewInvestigationPromptTemplateSource = ""; maxTurns = 250; maxBudgetUsd = 50. A clean review posts only LGTM and APPROVEs. With approve:true (or automation approvalMode:"human"), the TUI can privately investigate and refresh the review, approve, post human REQUEST_CHANGES feedback, or defer without publishing. Investigation history is durable workflow output and is never posted.
Required: repo.
Optional: approvalMode = "human"; agent = "claude"; model = ""; judgeAgent = "claude"; judgeModel = ""; judgeMaxTurns = 50; judgeMaxBudgetUsd = 5; maxApprovalRevisions = 2; verificationProfile = ""; reviewPromptTemplate = ""; reviewPromptTemplateSource = ""; approvalJudgePromptTemplate = ""; approvalJudgePromptTemplateSource = ""; maxNew = 5; maxActive = 5.
Required: none.
Optional: none beyond the shared policy envelope. This read-only operator workflow executes Java, Go, Python, Ruby, Node, npm, and npx probes on both the broad execution worker and the isolated verification worker.