ECC/docs/ROADMAP.md

ECC Roadmap

Status: draft for the maintainer to edit. Written 2026-09-02 against release 2.2.1. Dates are targets, not commitments. Numbers in brackets are the ones to confirm or change.

The two older planning docs stay as evidence and history: docs/ECC-2.0-GA-ROADMAP.md (2.0 milestones and control-plane deltas) and docs/ECC-PRO-SECURITY-ROADMAP.md (AgentShield and Pro conversion). This file is the short, current view.

Vision

ECC is the operating layer between a developer and whatever coding agent they run. One install, one set of skills, hooks, rules, and agents, working the same way in Claude Code, Codex, OpenCode, Cursor, Gemini, and the rest. The bar for everything that ships: simpler to read, faster to run, and traceable after the fact, for agents and humans alike.

Three things follow from that.

1. The repo is the product. Curated skills, hooks, and rules are the surface people install. Anything that is not installed, tested, or read by someone should not be in the tree.
2. Evidence over assertion. A harness change earns trust through a gate receipt, a capsule, and a reproducible verdict, not through a paragraph saying it works. The eval-harness frameworks exist to make that cheap.
3. Operator patterns travel. Approval loops, channel discipline, agreement generation, and e-sign placement were built for one desk. As generic skills they are useful to anyone running agents next to counterparties, customers, or money.

Where we are

• 2.2.1 shipped 2026-08-31 with guided manifest-driven setup, install-state ownership, repair and uninstall, and a signed release path.
• Catalog: [68] agents, [290] skills, [94] commands. The count is a liability as much as an asset. Overlapping and unreferenced skills exist.
• README is [2,200] lines with three install sections and a release-notes block that duplicates the changelog.
• Eval story: the eval-harness skill described a report format. The runnable layer (capsule journal, staged gate, replay fence, offline receipt) landed in September 2026 with a working example.
• AgentShield: steady free funnel, no bridge to a paid surface yet.

Plan

Track A: condense

Cut what nobody reads or installs. Merge what overlaps. One README that reads top to bottom in one pass. Exit criteria: no zero-reference tracked doc outside docs/releases/, no deprecated skill still shipped by default, README under [1,200] lines with one install path per harness.

Track B: evidence

Make the gate the default way a harness edit is judged. Wire the harness-optimizer agent and /harness-audit to emit gate receipts. Add capsule recording to the hooks that already log session activity. Then the next two plan slices: offline retrospective grouping over capsules (no new rollouts) and forced-compaction tests that prove pinned constraints survive.

Track C: operator skills

Ship the four desk-pattern skills (operator approval loop, counterparty channel discipline, master agreement generator with schedule append, e-sign field placement) and collect the first outside user of each. Fold what they report back into the skills before adding more.

Track D: distribution and revenue

Keep the release path boring: tag on main, CI green at the exact head, packed artifact tested on three platforms. Add the Pro conversion path AgentShield lacks: a hosted scan history and a PR-comment autofix loop. Details and scoring live in the security roadmap.

Next 90 days

Window: 2026-09-02 to 2026-12-01.

September

• Land the four PRs from the 2026-09-02 program: eval-harness frameworks, desk-pattern skills, condensation pass, this roadmap.
• README linear pass merged. Release notes move to CHANGELOG.md only.
• Delete list from the condensation survey executed, with catalog counts, manifests, and locale mirrors updated in the same PR.
• Decide the fate of continuous-learning v1 (deprecated since April): remove in [2.3.0] with a migration note, or keep as an archive outside the default install.

October

• harness-optimizer and /harness-audit produce gate receipts. A skill, hook, or agent change in this repo can cite a receipt in its PR.
• Capsule recording behind an opt-in hook flag, journaling tool calls and session boundaries with the default-deny payload allowlist.
• First taskset beyond the example: [20 to 60] tasks over one real skill family, with a held-out split and a reward-hack fixture.
• Skill catalog review: every skill has a test, a command, an agent, or a README mention, or it is marked for removal in [2.4.0].

November

• 2.3.0: condensation, eval frameworks, and operator skills in one release with the packed-artifact gate.
• Retrospective grouping over recorded capsules for one task family, report only, no promotion.
• Forced-compaction invariance test in CI for the pinned-state pattern.
• AgentShield Pro conversion CTA and hosted scan history behind a flag.

Decision points

• 2026-09-30: is the README under the line target with no test regressions? If not, cut scope on Track A rather than slipping the release.
• 2026-10-31: does a real taskset produce a stable verdict across three runs? If variance is high, hold Track B at receipts and do not start retrospective grouping.
• 2026-11-30: did any outside user adopt a desk-pattern skill? If none, stop adding operator skills and fold the four into a single guide.

Not on this roadmap

• Online reinforcement learning or weight updates from capsule data.
• Production transparency-log witnessing, GPU attestation, or key management inside the ECC package.
• Automatic merge or release driven by a gate verdict. The gate stops changes. A person promotes them.
• Any desk, payment, provider, or counterparty integration. Those belong to the systems that own them, not to a portable plugin.

How to edit this file

Change the bracketed numbers first. Move items between months freely. When a line ships, delete it here and record it in CHANGELOG.md. Keep the file under [200] lines.