GitHub Copilot CLI

GitHub Copilot CLI has a native hook system, so the adapter wires the same recall, queue, policy and error hooks as the other harnesses, in Copilot’s own config format. Support is broad (13 hooks, no postCompact) but rests on the adapter only: there is no working native-plugin route, and I could not run a live Copilot CLI to confirm behaviour. Adapter output below is verified in a sandboxed $HOME; the live-CLI claims come from the repo’s own docs and are labelled as such.

Prerequisites

Need	Why	Check
macOS or Linux	hooks are bash and Python
GitHub Copilot CLI with hooks	hook system GA Feb 2026 per the adapter docs	copilot --version
uv	runs every hook and installs the CLI	uv --version
reflect engine on PATH	recall and indexing	reflect --version
claude on PATH	the drain runs claude -p to write learnings	command -v claude
PyYAML	plugin/adapters/base.py imports yaml; stock python3 without it fails	python3 -c 'import yaml'

Install

# engine
uv tool install --force --upgrade --torch-backend cpu \
  'git+https://github.com/stevengonsalvez/ainb-reflect-memory.git[graph]'

# adapter
git clone https://github.com/stevengonsalvez/ainb-reflect-memory.git
cd ainb-reflect-memory
uv run --no-project --with pyyaml python plugin/adapters/copilot/copilot_adapter.py install --dry-run
uv run --no-project --with pyyaml python plugin/adapters/copilot/copilot_adapter.py install

--no-project stops uv creating a .venv in the checkout. If you prefer, pip install pyyaml and run python3 plugin/adapters/copilot/copilot_adapter.py install.

Flag	Effect
--dry-run	print the plan, change nothing
--no-hooks	skills and scripts only, no reflect.json
--no-bg-drain	skip the sessionStart drain entry (see limitations)
--force	overwrite a hand-written SKILL.md lacking the managed_by sentinel
--home DIR	treat DIR as $HOME (testing)

Re-run after updating the checkout; the deployed copy does not self-update.

Note

plugin/plugin.json carries only name, version, description, author, license and keywords: no hooks or skills field. plugin/copilot-hooks.json is not referenced by any manifest; the test suite uses it as a parity fixture. A Copilot plugin install therefore wires nothing by itself. Use the adapter.

What it deploys

~/.copilot/
├── hooks/
│   └── reflect.json              the one file the adapter owns, written whole
└── skills/
    ├── reflect/   SKILL.md  hooks/  scripts/  assets/  references/  reflect.toml
    ├── recall/    SKILL.md  hooks/  scripts/
    ├── status/    SKILL.md  scripts/
    ├── consolidate/  SKILL.md
    └── ingest/       SKILL.md

Each SKILL.md is the full plugin content plus managed_by: reflect-kb/adapters/copilot.

How hooks get wired

Copilot loads every *.json in ~/.copilot/hooks/ and combines them, so the adapter writes exactly one file, reflect.json, and never edits anyone else’s. The format differs from Claude and Codex on every axis:

	Claude / Codex	Copilot
Config	one shared settings.json / hooks.json	drop-in directory, one file per owner
Event names	SessionStart	sessionStart
Nesting	{matcher, hooks:[{...}]}	flat [{type, command}] plus top-level "version": 1
Timeout key	timeout	timeoutSec
Hook stdin keys	snake_case	camelCase

Every uv run command is prefixed REFLECT_HARNESS=copilot, which makes the recall hooks emit {"additionalContext": ...} and makes the shared stdin reader accept camelCase keys. The drain command is a ( ... &) subshell and carries no prefix. Example, as generated:

{
  "version": 1,
  "hooks": {
    "sessionStart": [
      { "type": "command",
        "command": "REFLECT_HARNESS=copilot uv run ~/.copilot/skills/recall/hooks/session_start_recall.py" },
      { "type": "command",
        "command": "(nohup ~/.copilot/skills/reflect/hooks/reflect-drain-bg.sh >/dev/null 2>&1 &) >/dev/null 2>&1",
        "timeoutSec": 5 }
    ]
  }
}

(Paths are written fully resolved, not with ~.) The 13 events:

Copilot event	Script (under ~/.copilot/skills/)	Role
sessionStart	recall/hooks/session_start_recall.py	top-3 recall as additionalContext
sessionStart	reflect/hooks/reflect-drain-bg.sh	detached drain
userPromptSubmitted	recall/hooks/user_prompt_submit_recall.py	prompt recall, deduped
preToolUse	reflect/hooks/pretooluse_context.py	policy lookup
permissionRequest	reflect/hooks/permission_request_reflect.py	permission lookup, arms watcher
notification	reflect/hooks/notification_reflect.py	arms watcher
postToolUse	reflect/hooks/posttooluse_minilearning.py	arms mini-learning watcher
postToolUseFailure	reflect/hooks/posttoolusefailure_minilearning.py	failure watcher
preCompact	reflect/hooks/precompact_reflect.py --auto --verbose	gate and queue
subagentStart	reflect/hooks/subagent_start_recall.py	subagent recall
subagentStop	reflect/hooks/subagent_stop_reflect.py	queue subagent transcript
agentStop	reflect/hooks/stop_reflect.py	slot update, session queue
sessionEnd	reflect/hooks/session_end_reflect.py	final queue
errorOccurred	reflect/hooks/error_occurred_reflect.py	error breadcrumb

postCompact is intentionally not wired: Copilot has no such event, and on the other harnesses it is bookkeeping only (it never drains or recalls), so nothing is lost.

Verify it works

jq '.version, (.hooks | keys)' ~/.copilot/hooks/reflect.json   # 1 and 13 event names
ls ~/.copilot/skills                                            # consolidate ingest recall reflect status

Smoke-test a recall hook with the Copilot env (exit 0, plain additionalContext object):

echo '{"sessionId":"t1","cwd":"'"$PWD"'"}' \
  | REFLECT_HARNESS=copilot uv run --script ~/.copilot/skills/recall/hooks/user_prompt_submit_recall.py
# {"additionalContext": ""}

Then run a Copilot session with a correction in it:

Evidence	Where
Transcript queued	new line in ~/.reflect/pending_reflections.jsonl, "harness": "copilot"
Drain ran	~/.reflect/drain.log
Learning landed	reflect stats, ~/.learnings/documents/
Hook failures	reflect errors count

Uninstall

uv run --no-project --with pyyaml python plugin/adapters/copilot/copilot_adapter.py uninstall

Removes ~/.copilot/hooks/reflect.json (and the hooks/ dir if empty; foreign *.json siblings are untouched) and the five adapter-managed SKILL.md files. The copied support directories remain:

rm -rf ~/.copilot/skills/reflect/{hooks,scripts,assets,references} ~/.copilot/skills/reflect/reflect.toml
rm -rf ~/.copilot/skills/recall/{hooks,scripts} ~/.copilot/skills/status/scripts
rmdir ~/.copilot/skills/{reflect,recall,status,consolidate,ingest} 2>/dev/null

Your knowledge base is untouched. Remove the engine with uv tool uninstall reflect-kb.

Limitations

• Per-prompt recall: repo says yes, I could not confirm. The adapter README and architecture doc state that userPromptSubmitted additionalContext reaches the model on Copilot CLI 1.0.66, and that the sessionStart and userPromptSubmitted envelopes were checked against the live binary. A docstring in user_prompt_submit_recall.py still says the output is ignored; it predates that validation. On older builds /recall is the manual fallback.
• Headless -p runs are unconfirmed. An earlier doc claimed sessionStart context is not injected in headless mode. Nothing in this repo confirms or refutes that.
• Capture needs claude. The drain runs claude -p. Without it, queued transcripts are not converted; recall still works and /reflect works by hand. --no-bg-drain removes the drain entry, so alone it leaves the queue unconsumed. The queue is shared across harnesses, so a Claude Code session on the same machine drains Copilot entries.
• Fewer skills. The adapter installs reflect, recall, status, consolidate, ingest. cost, errors-ack, corpus, export, slots are not installed; use the reflect CLI equivalents.
• Path anchors not rewritten. Unlike Codex, the Copilot adapter does not rewrite ${CLAUDE_PLUGIN_ROOT} anchors inside SKILL.md bodies yet, so a skill that cites them may point at paths that do not exist.
• Frozen copy. Re-run the adapter after upgrading.

See also: Hooks reference, Drain, Troubleshooting.