otherness

Imported from pnz1990/otherness (AGENTS.md). Install upstream with npx skills add pnz1990/otherness. Copyright stays with the author.

otherness — Project Agent Context

What This Is

otherness is an autonomous software development system. A single session plays the roles of coordinator, engineer, adversarial QA, scrum master, and product manager — sequentially, in a loop, without human input. It operates on any GitHub-hosted project by reading AGENTS.md, docs/aide/, and .otherness/state.json, then autonomously generates work queues, implements features, reviews PRs, and ships releases.

otherness is now running on itself. When this agent improves otherness, those improvements deploy to all projects using otherness on their next run via the self-update mechanism (git -C ~/.otherness pull).

Status: Active. Core loop stable. Self-improvement mode enabled.

⚠️ GLOBAL DEPLOYMENT WARNING — READ BEFORE EVERY MERGE Changes to agents/standalone.md or agents/bounded-standalone.md deploy immediately to every project using otherness worldwide on their next session startup. These files are the highest-risk change surface. See §Change Risk Tiers below.

SDLC Process

The team process is embedded in ~/.otherness/agents/standalone.md. This file contains only project-specific context.

Project Config

PROJECT_NAME:   otherness
CLI_BINARY:     ""
PR_LABEL:       otherness
REPORT_ISSUE:   741
REPORT_URL:     https://github.com/pnz1990/otherness/issues/741
BOARD_URL:      ""
BUILD_COMMAND:  bash scripts/validate.sh
TEST_COMMAND:   bash scripts/test.sh
LINT_COMMAND:   bash scripts/lint.sh
VULN_COMMAND:   ""

Architecture

otherness is a collection of markdown instruction files read by OpenCode (the AI coding agent runtime). There is no compiled binary, no server, no database. The "product" is the quality of the autonomous agent loop those instructions produce.

Deployment model: ~/.otherness/agents/ is a git repo. Every agent self-updates via git -C ~/.otherness pull on startup. A merged PR to pnz1990/otherness:main deploys to all users on their next run.

Runtime dependencies: OpenCode (executes the agents), gh CLI (all GitHub interaction), git (VCS + worktree isolation + self-update), python3 stdlib (YAML parsing, JSON state).

State: .otherness/state.json on a dedicated _state branch. Never on main. Parallel sessions use branch-push as a distributed lock.

Skills: agents/skills/*.md — reusable checklists loaded by the engineer and QA phases. Grown by /otherness.learn.

Package Layout

~/.otherness/                  ← the repo root (git clone at this path)
  agents/
    standalone.md              ← thin orchestrator: state mgmt, startup, loop skeleton, hard rules
    bounded-standalone.md      ← scoped agent — 50 lines wrapping standalone.md
    onboard.md                 ← existing project onboarding
    otherness.learn.md         ← learning agent — internalize from open-source
    otherness.arch-audit.md    ← architectural audit agent
    gh-features.md             ← GitHub API reference
    phases/                    ← phase implementations (loaded by standalone.md)
      coord.md                 ← Phase 1: heartbeat, rate-limit, queue-gen lock, stale watchdog, claim
      eng.md                   ← Phase 2: spec/plan/tasks (speckit integration), implement, memory
      qa.md                    ← Phase 3: adversarial review, spec conformance, merge, archive
      sm.md                    ← Phase 4: triage, stale recovery, metrics, cross-project learning
      pm.md                    ← Phase 5: product validation, roadmap health, competitive check
    skills/                    ← reusable skill files
      declaring-designs.md     ← spec quality standard
      reconciling-implementations.md  ← QA checklist
      agent-coding-discipline.md      ← surgical changes, verifiable goals
      autonomous-workflow-patterns.md ← workflow design patterns
      architectural-audit.md   ← four-lens audit methodology
      contribution-hygiene.md  ← PR and commit discipline
      agent-responsibility.md  ← judgment and responsibility patterns
      role-based-agent-identity.md    ← role identity patterns
      ephemeral-pr-artifacts.md       ← PR operational patterns
      explicit-anti-patterns.md       ← anti-pattern documentation
      triage-discipline.md            ← SM triage process
      PROVENANCE.md            ← learning session audit trail
      README.md                ← skill index
  boundaries/
    README.md
    example.boundary
  docs/
    aide/
      vision.md
      roadmap.md
      definition-of-done.md
      progress.md
  .otherness/
    state.json                 ← on _state branch only
  .opencode/
    command/
      otherness.run.md
      otherness.run.bounded.md
      otherness.setup.md
      otherness.status.md
      otherness.upgrade.md
      otherness.learn.md
      otherness.arch-audit.md
      otherness.cross-agent-monitor.md
  otherness-config.yaml
  otherness-config-template.yaml
  AGENTS.md
  README.md
  onboarding-new-project.md
  onboarding-existing-project.md
  scripts/
    validate.sh                ← BUILD_COMMAND target
    test.sh                    ← TEST_COMMAND target
    lint.sh                    ← LINT_COMMAND target

Change Risk Tiers

This is the most important section. Every PR must be classified before merge.

CRITICAL tier rule: Any PR touching agents/standalone.md, agents/bounded-standalone.md, or agents/phases/*.md must post [NEEDS HUMAN: critical-tier-change] on the PR and NOT be merged autonomously. Leave it for human review.

What "Tests" Mean for otherness

otherness has no unit tests. Its correctness can only be validated by running it on a real project and observing behavior.

scripts/validate.sh (BUILD_COMMAND) performs these structural checks:

1. No hardcoded project paths — agent, phase, and skill files must not reference specific projects in executable instruction context. PROVENANCE.md and HTML comment provenance metadata are exempt.
2. Skill references resolve — every Load skill: read ... reference in standalone.md and agents/phases/*.md points to an existing file on disk or in the repo.
3. Required files present — all agent files, phase files, skill files, docs, and command files listed in the required set exist.
4. Self-update present — standalone.md contains the git -C ~/.otherness pull self-update block.

scripts/test.sh (TEST_COMMAND) runs all 4 validate checks plus:

5. Integration test — verify otherness is running correctly on the reference project (first non-otherness entry in monitor.projects in otherness-config.yaml) by checking the _state branch shows activity within the last 72 hours.

scripts/lint.sh (LINT_COMMAND) checks agent files for structural correctness using python3 (no external tools required): CRLF line endings, null bytes, required phase headers in standalone.md.

The Self-Improvement Loop

When otherness runs on itself, the improvement areas are:

1. Agent loop quality — does the standalone agent make better decisions? Measure: fewer [NEEDS HUMAN] escalations per batch on reference projects, shorter time-to-merge per item.
2. Skills depth — are the skills files growing from /otherness.learn sessions? Measure: PROVENANCE.md entries, skill file count.
3. Onboarding friction — does /otherness.onboard produce complete, accurate docs/aide/ files? Measure: how many manual edits are needed after onboarding.
4. Documentation accuracy — does README.md match actual behavior? Measure: PM validation journey passes.
5. Tooling — do scripts/validate.sh, scripts/test.sh, scripts/lint.sh catch real regressions? Measure: these scripts improve over time.

The PM phase must compare otherness against the community it came from (Hermes, Multica, Archon) every batch, and identify gaps worth closing.

Anti-Patterns (QA blocks PRs containing these)

Label Taxonomy

Product Validation Scenarios (PM Phase)

The PM agent validates otherness by observing it on reference projects. The reference project is the first non-otherness entry in otherness-config.yaml under monitor.projects. Commands below use $REFERENCE_PROJECT as a placeholder — replace with your actual project slug (e.g. owner/my-project).

Scenario 1: reference project is alive

REFERENCE_PROJECT=$(python3 -c "
import re
in_monitor=in_projects=False
for line in open('otherness-config.yaml'):
    if re.match(r'^monitor:',line): in_monitor=True
    if in_monitor and re.match(r'\s+projects:',line): in_projects=True
    if in_projects:
        m=re.match(r'\s+- (.+)',line)
        if m:
            r=m.group(1).strip()
            if not r.endswith('/otherness'): print(r); break
")
gh api "repos/$REFERENCE_PROJECT/branches/_state" \
  --jq '.commit.commit.committer.date'
# Must be within 72 hours. If older: otherness has stalled on the reference project.

Pass: _state shows commits within 72 hours.

Scenario 2: reference project has open PRs or recent merges

gh pr list --repo $REFERENCE_PROJECT --json number,title,state,createdAt \
  --jq '.[] | "\(.state) \(.createdAt) \(.title)"' | head -5

Pass: at least one PR opened or merged in the last 7 days.

Scenario 3: skills are growing

ls ~/.otherness/agents/skills/ | wc -l
# Must be ≥ 4 (the 4 we started with). Growing means /otherness.learn ran.
tail -20 ~/.otherness/agents/skills/PROVENANCE.md

Pass: PROVENANCE.md has at least one entry dated within the last 30 days.

Scenario 4: README matches actual behavior

Read README.md and cross-reference against agents/standalone.md. Check: every command listed in README exists in .opencode/command/. Check: every agent file listed in the stack diagram exists on disk. Open kind/docs issue for each discrepancy found.

Scenario 5: self-improvement is happening

OTHERNESS_REPO=$(python3 -c "
import re
for line in open('otherness-config.yaml'):
    m=re.match(r'^\s+repo:\s*(.+)',line)
    if m: print(m.group(1).strip()); break
")
gh pr list --repo $OTHERNESS_REPO --state merged \
  --json mergedAt,title --jq '.[] | "\(.mergedAt) \(.title)"' | head -10

Pass: at least one PR merged in the last 14 days. If zero: otherness has not improved itself recently. Investigate why and open an issue.

Files Agents Must Not Modify

• AGENTS.md (this file)
• docs/aide/vision.md
• docs/aide/roadmap.md
• otherness-config-template.yaml (change risk tier HIGH — QA only, not autonomous)

Future Risk: Global Deployment Model

Document for future consideration (Option B):

Currently, any merged PR to the otherness main branch deploys immediately to every project using otherness via git -C ~/.otherness pull. This is Option A — fast, simple, no versioning overhead.

The risk: a bug in agents/standalone.md breaks every otherness session everywhere simultaneously. Currently mitigated by: CRITICAL tier requiring human review, integration test on the reference project, and the fact that the agent loop is resilient (a bad instruction causes a [NEEDS HUMAN] rather than a crash).

When to upgrade to Option B (versioned releases):

• When otherness is used by >10 distinct repos
• When a CRITICAL tier regression has caused real damage to a non-test project
• When users need stability guarantees across sessions

Option B design: git tags as releases, projects pin to a tag in otherness-config.yaml, /otherness.upgrade bumps the pin with a changelog review step. The git pull self-update becomes git checkout <pinned-tag>.