# Agent Delivery Harness v1 adoption decision

- **Date:** 2026-09-03
- **Decision:** Adopt with revisions
- **Operating mode:** Opt-in and human-supervised
- **Base:** `origin/dev` at `7b5384ae10be7992f4d3a506a25f19452b8456bd`

## Reconciliation

The adoption decision follows the completed evidence-backed delivery pilot. No
new trial was opened and no historical result was rewritten.

|                                                       PR | Purpose                             | Human merge to `dev` |
| -------------------------------------------------------: | ----------------------------------- | -------------------- |
| [#333](https://github.com/rikster/stratofusion/pull/333) | Initial workflow pilot              | 2026-09-02           |
| [#334](https://github.com/rikster/stratofusion/pull/334) | Trial 2 documentation               | 2026-09-02           |
| [#335](https://github.com/rikster/stratofusion/pull/335) | Trial 2 closeout/security follow-up | 2026-09-02           |
| [#336](https://github.com/rikster/stratofusion/pull/336) | Trial 3 documentation               | 2026-09-02           |
| [#338](https://github.com/rikster/stratofusion/pull/338) | Trial 4 UI refactor                 | 2026-09-02           |
| [#340](https://github.com/rikster/stratofusion/pull/340) | Trial 5 UI refactor                 | 2026-09-02           |
| [#341](https://github.com/rikster/stratofusion/pull/341) | Five-trial closeout                 | 2026-09-02           |

Open PR #339 was inspected before adoption work. It changes the provider-neutral
shared-workspace foundation, database schema, feature configuration, and
architecture documentation. This adoption slice avoids those product and
database files. It changes repository workflow documentation, templates, and
local-only tooling only.

## Evidence used for the decision

- Five trials completed.
- Four sole-maintainer usefulness responses were recorded: three `yes`, one
  `partly`; Trial 4 had no response.
- Four trials were timed at 7, 2, 2, and 2 minutes: 13 minutes total.
- Zero confirmed privacy incidents occurred.
- No independent reviewer evidence was produced.
- No timed baseline exists, so the evidence supports no speed, review-time,
  throughput, or 10x claim.

## Revisions adopted in v1

1. Human-confirmed T0-T3 risk tiers; automation can never lower a tier.
2. A bounded lifecycle with terminal states and at most two repair cycles.
3. Fresh-context, read-only review separated from implementation and repair.
4. A maximum of two writing agents and explicit acknowledgement that worktrees
   do not isolate external resources.
5. A validated, gitignored lease record; no scheduler or orchestration service.
6. A fixed human merge/release path through `dev`, a separate release PR, and
   immutable-SHA deployment.
7. Evidence-helper defaults that cannot fabricate a before/after pair, honest
   verification metadata, stronger navigation and image controls, and safe Git
   fingerprints.

## Branch-protection blocker

Read-only calls to GitHub's branch-protection endpoint for private `dev` and
`main` returned HTTP 403 with: “Upgrade to GitHub Pro or make this repository
public to enable this feature.” Enforced protection on both branches is a
prerequisite for unattended writing agents. v1 does not change the plan,
visibility, repository settings, or protections; unattended writers remain
disabled.

## Explicit exclusions

This decision does not install Symphony, a team framework, Greptile, Cursor
Cloud Agents, or another orchestrator. It does not change CI, branch settings,
providers, OAuth, billing, rclone, databases, deployment, production, or
customer-facing behavior. It does not merge, deploy, publish, or schedule the
accompanying drafts.

The original pilot article and social drafts were superseded by the v1 adoption
drafts rather than left as a second, outdated “continue the pilot” narrative.
The original pilot reports and measurements remain unchanged.

## Acceptance outcome

The narrow harness is adopted for supervised opt-in use once this change is
human-reviewed and merged. CI, CASA, scanners, and human authority remain
authoritative. The completed pilot scorecard remains the measurement source;
future harness work is operational follow-up, not an extension of the pilot.

## Rollback

Revert the v1 workflow files and tooling together. Leave both original pilot
reports and the merged PR history intact. Ignored local leases and evidence can
be discarded after their review window.