UNPKG

@blundergoat/goat-flow

Version:

AI coding agent harness and local dashboard for Claude Code, OpenAI Codex, Google Antigravity, and GitHub Copilot - setup audits, guardrails, structured skills, deny hooks, and persistent learning loops.

github.com/blundergoat/goat-flow

blundergoat/goat-flow

46 lines (28 loc) • 2.86 kB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
# Footguns **Traps in the code itself.** A footgun exists whether or not an agent triggers it - it's a property of how the codebase is structured. Example: "renaming one doc breaks 5 others because of dense cross-referencing." The trap is in the architecture, not in what the agent did. If the agent did something wrong → `.goat-flow/learning-loop/lessons/` instead. If a proven approach should be reused → `.goat-flow/learning-loop/patterns/` instead. ## Generated Index `INDEX.md` in this directory is generated by `goat-flow index` - never edit it by hand. After adding, editing, renaming, or resolving entries, re-run `goat-flow index`; `goat-flow stats --check` fails while the index is stale. ## Automatic Capture Policy Manual edits are normal: an explicit request to add or update a durable learning-loop entry means edit the Markdown in the correct directory. Programmatic automatic capture from terminal sessions, quality reports, PR reviews, or agent output is disabled unless `.goat-flow/config.yaml` explicitly sets `learning-loop.auto-capture.enabled: true` and a CLI-owned writer for that target exists. Prompts may suggest an entry, but automatic durable writes belong to CLI-owned code after opt-in. Auto-capture candidates must follow Extract / Consolidate / Skip: - Extract only when the item changes a future READ/SCOPE/ACT/VERIFY decision. - Consolidate into an existing entry when the same root cause already exists. - Skip raw tool output, successful smoke runs, UI/deploy churn, chain-of-thought, screenshots, raw JSON/HTML, duplicate dumps, and "I read the docs" summaries. Every entry MUST include file path evidence with grep-friendly semantic anchors (function name, unique string, or `(search: "pattern")`) per ADR-024. Do not use line numbers - they go stale on every edit. Agents should scan only entries above the `## Resolved Entries` section. Resolved entries are historical record, not active traps. Prefer category bucket files such as `hooks.md`, `setup.md`, or `auditor.md`. Every bucket file MUST start with a YAML frontmatter block that includes BOTH a `category` and a `last_reviewed` date (ISO `YYYY-MM-DD`). `goat-flow stats --check` fails when `last_reviewed` is missing. ```yaml --- category: hooks last_reviewed: 2026-04-20 --- ``` Inside a bucket, add entries as `## Footgun:` blocks. Each entry MUST begin with a `**Status:**` line (one of `active`, `resolved`) followed by a `**Created:**` date line in `YYYY-MM-DD` form: ```markdown ## Footgun: <short name> **Status:** active | **Created:** 2026-04-20 | **Evidence:** OBSERVED <body> ``` Entries without `**Status:**` cannot be split into active-vs-resolved by the audit scanner. Legacy single-entry files still work during migration, but category buckets with the frontmatter + `**Status:**` contract are the preferred and audited format.