E07 — Agentic-workflow landscape pass (context integrity + spec-time verification)
Pillar (primary): P9 — Self-improvement velocity
Pillar (secondary): P1 — Doc completeness
Status: done (2026-08-14)
Phase: Phase 1 — Foundation
Started: 2026-08-14 (chartered and executed same day, at maintainer direction)
Closed: 2026-08-14 — all five items shipped in v1.29.0
Owner: maintainer + AI coding agent
Frozen intent — Outcome and exit criteria approved by maintainer on 2026-08-14.
Agents do not edit them; halt and renegotiate instead.
Outcome (jobs-to-be-done)
When the maintainer reviews a peer agentic workflow and finds mechanisms that close real gaps, they
want those converted into scoped, individually-shippable items — so each either ships in a versioned
release or is explicitly rejected with reasoning.
Source: a review of a Unity-based agentic workflow (specialised agents per phase, context “start
tokens”, “caveman” output compression, a doc orchestrator). Five ideas passed the fit test; the rest
were rejected in the review itself.
Exit criteria (binary)
Out of scope (rejected in review, recorded here)
- Severity tiers on review findings (Blocker / Warning / Nit). We route by layer, not severity
— layer says where to fix, severity only says how loud to be. A second axis reintroduces the noise
the routing was built to remove.
- Named agent personas. Rejected for the same reason as in E06: token cost per turn, coupling to
persona-capable tooling, and our challenge-prompt + cross-AI + decision-ownership mechanisms
already cover the goals. (See “Open questions” — role briefs without personas may be worth
revisiting.)
- Split doc architecture, manual agent triggering, human story pre-filtering. Already present.
- Platform-specific (Unity) and context-window-size practices. We stay tool-agnostic.
Linked docs
- Pillars: P9, P1
- Precedent: CHANGELOG v1.20.0–v1.23.0 and v1.25.0 (prior landscape passes)
- Analysis: recorded in this charter and in the item Resolutions; the source was a described
workflow rather than a public repo, so there is no commit to pin.
Item roster
See ARCHIVE.md — all five closed. BACKLOG.md is empty.
Open questions
- Role briefs per phase. The source workflow assigns each phase a named agent with a role
description (spec generator, spec reviewer, plan generator, plan reviewer, implementer, deep
reviewer, doc orchestrator). We rejected personas, but the underlying idea — each phase has a
different optimal stance, and only two of ours are captured as paste-able prompts
(
AGENT_KICKOFF.md, AUTONOMOUS_LOOP.md) — is a real gap. Raised by the maintainer during
execution; deliberately not folded into this epic to avoid scope creep mid-flight.
Resolved 2026-08-19 — chartered as E08 and shipped in
v1.30.0 as templates/ROLE_BRIEFS.md. Personas stayed
rejected; the briefs carry a stance per phase and link to the rules rather than restating them.
A re-read of this epic’s source during that pass also surfaced two gaps E07 did not extract
(blast-radius verification, self-answered clarifying questions under unattended runs), both
filed on E08. This charter’s Outcome and Exit criteria are unchanged.
Risks
- Convention fatigue. This is the second landscape pass in one day; v1.25.0 added five
conventions and this adds five more. Adopters have to absorb them. Mitigation: three of the five
are single sections, and two live in templates where they are read rather than remembered.
- The canary invites over-trust. A detector that produces false positives and false negatives can
make a session feel verified. Mitigated by stating its limits in the same breath as the rule.