Skip to content

docs(spec): harness principles for the runtime agent, adopted from ECC (CLEAN-150) - #24

Open
maksymhryzodub-prog wants to merge 4 commits into
mainfrom
docs/CLEAN-150-ecc-agent-principles
Open

maksymhryzodub-prog wants to merge 4 commits into
mainfrom
docs/CLEAN-150-ecc-agent-principles

Conversation

@maksymhryzodub-prog

Copy link
Copy Markdown
Contributor

What

SpecKit specification specs/001-ecc-agent-principles/spec.md: which patterns from ECC the runtime adopts globally, for every agent built on it, written as prioritized user stories with testable requirements and measurable success criteria.

# Pattern Priority
US1 Guardrails that fire every time — deterministic rules around each tool call (block / transform / record), platform baseline + agent-level, fail-closed for guards, fail-open for observers P1
US2 Lean always-on context — rules only in the identity file, skills and tool details on demand, context budget visible to the operator P1
US3 Teachable memory with confidence — learned behaviours as trigger/action/confidence/evidence; reinforce, correct, decay, pin, prune, promote to skill P2
US4 Compaction at task boundaries, state saved first; hard limit stays as last resort P2
US5 Verified completion — claims checked against mutating tool outcomes before delivery P3
US6 Delegation with a completion contract — isolated workers, parallel, parent owns collection P3

Out of scope: installing ECC (it is a Claude Code plugin), its catalog, Memory Vault / Control Pane / AgentShield.

Sources read: ecc.tools, README, hooks/hooks.json, continuous-learning-v2, strategic-compact, verification-loop, rules/common/{agents,performance,security,hooks,coding-style}.md.

Why

The runtime is itself an agent harness; today its safety and behaviour live in prompt text (50–80% reliable by ECC's numbers), the admin identity file is already truncated at its size cap, and learned notes accumulate without weight or pruning. The spec sharpens existing roadmap items (#4 curator, #6 delegation, #10 safety, #15 hooks) and orders them.

Next

/speckit-plan on this spec → implementation tickets per user story.

Jira: https://dreamvention.atlassian.net/browse/CLEAN-150

🤖 Generated with Claude Code

maksymhryzodub-prog and others added 4 commits October 8, 2026 19:00
…C (CLEAN-150)

SpecKit specification fixing which ECC patterns the runtime adopts for
every agent: deterministic guardrails around tool calls, lean always-on
context with skills/tools on demand, confidence-weighted teachable memory,
compaction at task boundaries, verified completion, delegation with a
completion contract. Prioritized user stories, 30 functional requirements,
measurable success criteria; ECC itself is not installed.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Spec Kit 0.16.5 skills under .claude/skills/speckit-*, the shared
.specify/ templates/scripts/workflows (feature.json stays local), the
.worktreeinclude list so .env files follow new worktrees, and .gitignore
entries for the local second-agent runtime and .env.project.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…-150)

Phase 0/1 artifacts from /speckit-plan: research decisions R1-R15 anchored
to the loop, prompt, skill, session and task code; data model (rules,
tool-call records, instincts with confidence bands, context budget,
compaction checkpoint, Tool.effect, delegated worker); contracts for the
new tools, operator commands, config and file formats; quickstart
validation per user story. Two new agent slices (guard, instinct), the
rest extends the slices that own the behaviour.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…les (CLEAN-150)

90 tasks across setup, foundation, six user-story phases and polish;
tests first per phase, file paths per task, one PR and ticket per story.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant