A living list of work beyond the initial migration, organized into seven buckets. Each item carries a rough impact (1–5) and effort (1–5) score.
How to read the scores. Impact 5 = changes what users can do, not just how. Effort 5 = multi-week, touches several components, needs experimentation. Priority is not strictly impact÷effort — high-impact items are often worth doing even when expensive — but the ratio is a useful first filter.
- Shipped
- A. Reach — more hosts, more deployment targets
- B. Pipeline depth — more/richer stages
- C. Quality & safety — enforcement, sandboxing, scanning
- D. Observability & learning — telemetry, metrics, persistent learning
- E. Developer experience
- F. Integrations — where the team plugs in
- G. Innovation bets — speculative, future-oriented
- Priority queue
- Staying ahead of the curve — bets
Cross-references. Items tagged [cmp-E-N] were added or refined on 2026-06-03 after the comparative analysis against six adjacent AI-dev frameworks (comparative-analysis.md). Items tagged [hist-N] came from audit-archive/HISTORY.md § Between-cycle observations. Where multiple sources converge on the same idea, that's recorded inline.
Completed backlog items are preserved here so the active backlog tables stay scannable.
| Bucket | # | Item | I | E | Shipped |
|---|---|---|---|---|---|
| A | A1 | Gemini CLI adapter | 4 | 2 | v0.2.0 · CHANGELOG |
| A | A4 | Pluggable adapter discovery | 3 | 2 | landed · CHANGELOG |
| A | A6 | Native Windows validation and support | 2 | 2 | landed · Node 22 windows-latest portability smoke |
| B | B1 | Accessibility audit stage | 4 | 2 | v0.2.0 · CHANGELOG |
| B | B2 | Performance budget stage | 4 | 3 | v0.6.0 · CHANGELOG |
| B | B3 | Deploy cost gate | 4 | 2 | landed in PR #221 · CHANGELOG |
| B | B4 | Observability gate | 4 | 2 | v0.2.0 · CHANGELOG |
| B | B5 | Migration safety stage | 5 | 3 | v0.4.0 · CHANGELOG |
| B | B6 | Documentation gate | 3 | 2 | landed in PR #225 · CHANGELOG |
| B | B7 | Multi-language QA | 4 | 4 | Unreleased · Phase 19 · PR #264 |
| B | B8 | Cross-artifact consistency analyze [cmp-E-1] |
4 | 2 | v0.6.0 · CHANGELOG |
| B | B9 | Bounded workspace deltas [cmp-E-2] |
4 | 3 | v0.6.0 · CHANGELOG |
| B | B10 | Discover Standards preprocessing [cmp-E-5] |
3 | 3 | v0.6.0 · CHANGELOG |
| C | C1 | Filesystem-level allowedWrites enforcement |
4 | 4 | v0.6.0 · CHANGELOG |
| C | C2 | Secret scanning hook | 4 | 1 | v0.2.0 · CHANGELOG |
| C | C3 | License compatibility gate | 3 | 1 | v0.6.0 · CHANGELOG |
| C | C4 | Reproducible runs (recording half) | 4 | 4 | v0.4.0 · CHANGELOG |
| C | C5 | Capability-required permissions | 3 | 2 | v0.6.0 · CHANGELOG |
| C | C6 | Tamper-evident gate chain | 3 | 3 | v0.6.0 · CHANGELOG |
| C | C7 | eslint-plugin-security [hist-a] |
3 | 1 | v0.6.0 · CHANGELOG |
| C | C8 | CHANGELOG-per-PR fragments [hist-b] |
3 | 2 | v0.6.0 · CHANGELOG |
| D | D1 | OpenTelemetry tracing per stage | 5 | 3 | v0.2.0 · CHANGELOG |
| D | D2 | Gate-pass-rate dashboards | 4 | 2 | v0.2.0 · CHANGELOG |
| D | D3 | Lessons-learned across projects (org-shared) | 5 | 4 | v0.3.0 · CHANGELOG |
| D | D4 | Per-role per-model performance scores | 5 | 3 | v0.3.0 · CHANGELOG |
| D | D5 | Adaptive routing | 5 | 3 | v0.3.0 · CHANGELOG |
| D | D6 | Cost telemetry | 4 | 2 | v0.3.0 · CHANGELOG |
| D | D7 | Persistent project memory (embeddings index) | 5 | 4 | v0.2.0 · CHANGELOG |
| E | E1 | devteam status rich CLI output |
3 | 1 | v0.1.0 as devteam summary; Phase 11.1-11.3 updates · CHANGELOG |
| E | E2 | Web UI for pipeline runs | 4 | 4 | v0.2.0 · CHANGELOG |
| E | E4 | Live streaming output | 3 | 2 | landed in shared headless invoke helper · CHANGELOG |
| E | E5 | Pre-flight check (devteam doctor) |
3 | 1 | v0.1.0 plus Phase 14.2-14.3 updates · CHANGELOG |
| E | E6 | devteam replay <stage-id> |
3 | 3 | v0.4.0 · CHANGELOG |
| E | E7 | /goal integration for convergence-shaped stages |
3 | 2 | v0.6.0 · CHANGELOG |
| E | E8 | Codebase audit feature | 5 | 3 | v0.3.0 · CHANGELOG |
| E | E10 | Autonomous run watch mode | 3 | 1 | PR #268 · Phase 20 |
| E | E11 | Prototype mode | 4 | 2 | Unreleased · pre-SDLC packet for fast learning, optional host-run builds, feedback, and explicit promotion into a normal delivery track. See Phase 23 plan. |
| F | F1 | GitHub PR integration | 4 | 3 | v0.2.0 · CHANGELOG |
| F | F4 | CI runner integration | 4 | 3 | v0.4.0 (GitHub Actions only) · CHANGELOG |
| G | G1 | Multi-model peer review | v0.2.0 · CHANGELOG | ||
| G | G2 | Closed-loop acceptance criteria → exec spec → tests | v0.4.0 · CHANGELOG | ||
| G | G3 | Production feedback loop | landed · CHANGELOG | ||
| G | G4 | Red-team role between build and peer-review | v0.3.0 · CHANGELOG | ||
| G | G6 | Stage shopping (AI-inferred tracks) | v0.6.0 · CHANGELOG | ||
| G | G7 | Verification beyond tests | v0.4.0 · CHANGELOG | ||
| G | G8 | Long-context architecture continuity | v0.3.0 · CHANGELOG | ||
| G | G10 | Role tool budgets | landed · CHANGELOG | ||
| G | G11 | devteam run --repair — bug-fix intent mode (ADR-009) |
complete (Phase 10) · CHANGELOG |
| # | Idea | I | E | Notes |
|---|---|---|---|---|
| A2 | Cursor / Windsurf / Aider / Cline adapters | 3 | 3 | One per IDE-embedded agent. Each is an adapter, mostly install-payload work. |
| A3 | Cloud-runner adapter (e.g. AWS Lambda + Bedrock, Replit Agent) | 4 | 4 | Host adapter that runs one workstream on a remote worker, not the user's laptop. Enables long-running stages (multi-hour audits, big test suites). Phase 21 plan proposed for review. |
| A5 | API-direct adapter (no host CLI; talks to Anthropic / OpenAI / Google APIs directly) | 3 | 3 | For users who don't want to install claude-code or codex but still want orchestration. Lighter dependency footprint. |
| A7 | 4 | 2 | ✅ Unreleased · hosts/docker/ packages Stagecraft into a non-root Docker runner for unattended local orchestration against a mounted project, with runtime-only credentials and conservative lock handling. |
|
| A8 | Omnigent runtime adapter follow-through (#291) | 4 | 4 | Phase 24.1 ships a conservative omnigent host adapter. Follow-through tracks launch config (#292), prompt-file transport (#293), policy bridging (#294), session evidence (#295), and optional director-mode consolidation (#296). See Phase 24. |
No open items. B7 moved to Shipped in Phase 19.
| # | Idea | I | E | Notes |
|---|---|---|---|---|
| C1 | allowedWrites enforcement |
4 | 4 | ✅ v0.6.0 · CHANGELOG |
| C2 | 4 | 1 | ✅ v0.2.0 · CHANGELOG | |
| C3 | 3 | 1 | ✅ v0.6.0 · CHANGELOG | |
| C4 | 4 | 4 | ✅ v0.4.0 · CHANGELOG | |
| C5 | 3 | 2 | ✅ v0.6.0 · CHANGELOG | |
| C6 | 3 | 3 | ✅ v0.6.0 · CHANGELOG | |
| C7 | eslint-plugin-security[hist-a] |
3 | 1 | ✅ v0.6.0 · CHANGELOG |
| C8 | [hist-b] |
3 | 2 | ✅ v0.6.0 · CHANGELOG |
| C9 | [hist-c] |
3 | 2 | ✅ Unreleased · Phase 1/2 audit findings now require verified_by evidence, templates expose verification slots, and structural tests lock the contract. |
| # | Idea | I | E | Notes |
|---|---|---|---|---|
| D5 | D5 maturation — continuous adaptive routing | 5 | 3 | Today D5 proposes role-level swaps; the mature form re-routes the next run based on the prior run's outcomes automatically. Phase 17 makes per-workstream dispatch history durable and privacy-bounded; the gate stays shut pending ≥5 dispatches per (role, host) pair across ≥2 real user projects and cost telemetry. |
| # | Idea | I | E | Notes |
|---|---|---|---|---|
| E3 | VS Code extension | 3 | 3 | Sidebar with stage status, "run next stage" button, gate viewer. |
| E9 | Conversational stage mode [cmp-E-4] |
3 | 3 | devteam stage requirements --interactive opens a conversation with the PM subagent to refine the brief through Q&A before producing the artifact. Useful specifically for upstream stages (requirements, design, clarification) where the artifact benefits from refinement before being rendered. Architecture supports it (adapters could expose streaming-conversation alongside one-shot render). Implement if user feedback indicates the gate-driven loop is too rigid for upstream stages. Related to E7 but different mechanism: E7 is host-loops-until-condition; E9 is stage-manager-converses-with-agent. |
| E11 | Prototype mode | 4 | 2 | devteam prototype creates a lightweight pre-SDLC packet, can run the build prompt in a packet workspace, captures feedback, and records explicit discard/iterate/promote decisions. It is deliberately not a production gate track; promotion hands off into devteam run --feature-file ... --track <t>. |
| # | Idea | I | E | Notes |
|---|---|---|---|---|
| F2 | Jira/Linear ticket integration | 3 | 2 | devteam stage requirements --ticket FOO-123 pulls the ticket as the feature brief input. Gates link back to the ticket. |
| F3 | Slack/Discord notifications | 3 | 1 | Pipeline events (stage start, fail, escalate) post to a channel. Triggers for human checkpoints. |
| F5 | Pre-commit hook integration | 3 | 1 | Optional pre-commit hook that runs the relevant track for the change (nano if config-only, full if otherwise). |
These don't fit neatly in impact/effort because their value depends on how the field evolves. They are the items that would most meaningfully differentiate this tool.
Design specs include architecture diagrams (images). Stage 2 (design) and Stage 5 (review) accept image inputs. Principal can output a system diagram, not just prose. Visual reasoning is no longer a separate workflow.
Retrospective stage proposes changes to stages.js / roles/ / rules/ based on what worked. Proposals queue for human approval. The pipeline learns its own shape from operation.
The full evidence, effort/risk ratings, dependencies, and PR sequence now live in the current audit backlog and roadmap.
No ungated implementation item remains from the audit's immediate and targeted
improvement batches. Phase 16 completed privacy-safe readiness/export, Phase 17 made
dispatch evidence durable, Phase 18 added explicit accepted-resolution evidence for
H3, Phase 19 shipped polyglot verification in PR #264, and Phase 20 implements the
separable devteam run --watch operator UX without enabling active stall response. The
next capability horizon is real collection followed by review, not calendar-driven
activation. E9 conversational stage mode remains a discovery proposal until five real
users report upstream rigidity.
Completed from this audit cycle: dashboard HTML safety and lifecycle (PR #235),
native Windows CI evidence, support wording, and A6 promotion (PR #236), and bounded
durable transcript streaming (PR #237). Current-truth reconciliation removed the
remaining P1-3 ownership, vocabulary, comment, count, link, and provider drift.
Stable-fact consistency now locks schema vocabulary, Node/platform support, and the
absence of volatile test totals while leaving audit history untouched (audit P2-3).
The autonomous driver decomposition is complete: characterization, dispatch/transient,
and fix/ruling/merge transitions landed as three behavior-preserving slices while
run() retained lock, loop, effect, and final-persistence ownership (audit P2-2).
-
P3-1 — evidence readiness and export. Phase 16 implements the approved privacy model, read-only local readiness, consented aggregate export, identity lifecycle, strict bundle validation, and explicit multi-project analysis. See
plans/phase-16-evidence-readiness-and-export.md. Phase 17 adds allowlisted per-workstream dispatch events so D5 evidence accumulates during normal autonomous runs without reconstructing history from gate snapshots. Seeplans/phase-17-durable-evidence-instrumentation.md. Phase 18 adds explicit, hash-bound human acceptance for successful fix/retry resolutions so H3's derivability threshold can be measured without exporting recipe text. Seeplans/phase-18-accepted-resolution-evidence.md. This makes the gates below measurable; it does not open them. -
D5 maturation — continuous adaptive routing. Today D5 proposes role-level swaps; the mature form re-routes the next run based on the prior run's outcomes automatically. Evidence baseline (2026-06-14,
plans/adaptive-routing-evidence.md): zero real-run telemetry at review time. Phase 17 starts durable collection from real autonomous dispatches; it does not backfill old gates. Gate stays shut pending ≥5 durable dispatches per (role, host) pair across ≥2 real user projects and cost telemetry. ADR-007 Tier 1 (liveness heartbeat + observe-only stall probe) implemented in Phase 11.1; ADR-008 (advisory sweep +--fail-on-advisory) implemented in Phase 11.2; ADR-007 Tier 2 remains evidence-gated. -
H3 — Recipe factory (escalation→recipe learning) (Phase 3 of ADR-003 · design). Persist resolved escalations as semantically-indexed fix-recipes via the existing
core/memory/embedding store (D7);computeFixStepsconsults it on a FAIL signature before escalating, so recurring derivable failures resolve deterministically. Evidence review done (2026-06-14,plans/h3-ground-truth.md): zero real run logs/archives and no recurring unresolved class. Phase 18 makes explicit acceptance measurable under ADR-012. Gate stays shut pending ≥2 real projects each with ≥5 autonomous fix/retry runs, the same schema-bound failure accepted ≥3 times across both projects, and ≥80% derivability. Tracked by GitHub #142. -
ADR-005 standing grants. Keep deferred until at least 10 repair runs across 2+ projects and consequence-ceiling halt data establish which grants operators routinely approve. Tracked by GitHub #144.
-
ADR-007 Tier 2 active stall response. Keep deferred until real
stall-detectedevents calibrate frequency and threshold. Tracked by GitHub #145.
Five items that the comparative analysis or shifted context argues against investing in now:
- E3 — VS Code extension. Stagecraft sits above the IDE, not inside it. Building an editor extension works against that positioning, and IDE-native tooling is a crowded category.
- A2 — Cursor/Windsurf/Aider/Cline adapters. Supporting 30+ AI agents is a maintenance treadmill. Land A4 — Pluggable adapter discovery first and let the long tail be community-built.
- F2 / F3 / F5 — Jira / Slack / pre-commit integration. None changes what Stagecraft can do. Accept community PRs but don't invest core time.
- G9 — Self-modifying pipeline. Premature. Wait until multiple teams use the platform in different configurations before optimizing for any one signal.
Six positioning bets about where software development is heading.
Design the contract assuming 10× capability in 2 years. The schema (gate JSON), the contract (per-workstream gates merged to stage), and the routing layer should outlive the specific models we route to today.
Single-model agentic systems are giving way to multi-model coordinated systems. For non-trivial tasks, the stronger outcome usually involves 2–3 different model families. The routing layer is already built for this; the next step is making diversity structurally load-bearing (G1 multi-model peer review, D4/D5 adaptive routing).
The pipeline produces structured gate JSON so evals can be built on top. Every refinement should make the gate richer and more measurable. D1/D2/D4 are all expressions of this bet.
Today's pipeline is mostly stateless within a run; each new run starts fresh. Sustained coordination across projects, with continuous learning, requires the work in D3/D7/G8.
An agent with a deep, well-composed tool stack outperforms one that only writes text. Role briefs are the place to encode tool budgets per role (G10). Skills are early steps; tool negotiation is the mature form.
EU AI Act, US executive orders, SOC 2 controls: all require reproducible runs, audit trails, and documented decision provenance. C4 (reproducible runs), C6 (tamper-evident gate chain), and D1 (tracing) address this directly.
A coordinated team of specialized agents — each with a role, a tool budget, a gate contract, and shared memory — outperforms a single model. The model is a substrate. The team is the product.