repo-context

name	repo-context
description	Per-runId repo binding. Agents do NOT auto-detect the repo from cwd — hooks/bind-repo.ps1 resolves --repo / defaultRepo at sessionStart and publishes the bound context to a sidecar JSON that every agent reads.
allowed-tools	Read

Repo Context — hook-resolved (R5 + R6)

Repo selection is explicit, not derived from git remote get-url origin. The active repo comes from one of:

1. --repo <name> argument on the slash command (operator types it)
2. .defaultRepo in ~/.copilot/state/config.json (admin pre-sets per machine)
3. Otherwise — hooks/bind-repo.ps1 writes an unresolved sentinel; agents halt and ask.

This skill exists for agents to read the sidecar, not to compute it. The deterministic resolution + lookup is in hooks/bind-repo.ps1 (per N2 — deterministic work is off the LLM).

What the hook publishes

hooks/bind-repo.ps1 fires on userPromptSubmitted and writes ~/.copilot/state/runs/<runId>/.repo-context.json with this shape:

{
  "resolved": true,
  "repoName": "CRM.Client.UnifiedClient",
  "repoUrl": "https://...",
  "primaryBranch": "master",
  "areaPath": "Proj\\Client\\UnifiedClient",
  "iteration": "Proj\\Current",
  "bugTags": "playwright; auto-fix",
  "defaultAssignee": "alice",
  "adoOrg": "https://dev.azure.com/contoso",
  "adoProject": "ProjectX",
  "kustoCluster": "https://contoso.kusto.windows.net/",
  "kustoDatabase": "TelemetryDb",
  "pipelines": { "ci": 23865, "nightly": 23878, "weekly": 23863, "multiPr": 23890 },
  "testEnv": {
    "appUrl": "https://...",
    "authEmail": "...",
    "certPath": "...",
    "storageStatePath": "state-storage-auth-file.json",
    "testCommand": "npx playwright test"
  }
}

If resolved is false, the agent halts with the sidecar's reason field (e.g. "no --repo and no defaultRepo").

If the named repo isn't in config.json .repositories[], the hook writes .hook-fatal and the agent halts in its first Startup step.

Agent usage — first checklist step

Every agent does this once at the top of its workflow (the agents already do this — see their ## Startup block):

# 1. Halt if hook flagged a fatal
RUN_DIR="$HOME/.copilot/state/runs/$RUN_ID"
[ -f "$RUN_DIR/.hook-fatal" ] && { cat "$RUN_DIR/.hook-fatal"; exit 1; }

# 2. Read the bound context
CTX="$RUN_DIR/.repo-context.json"
if ! jq -e .resolved "$CTX" >/dev/null 2>&1; then
    echo "repo-context not resolved. $(jq -r .reason "$CTX")"
    exit 1
fi

# 3. Bind the variables the agent needs
REPO_NAME=$(jq -r .repoName "$CTX")
AREA_PATH=$(jq -r .areaPath "$CTX")
ADO_ORG=$(jq -r .adoOrg "$CTX")
KUSTO_CLUSTER=$(jq -r .kustoCluster "$CTX")
# ... and so on

PowerShell equivalent:

$ctx = Get-Content "$env:USERPROFILE\.copilot\state\runs\$env:RUN_ID\.repo-context.json" | ConvertFrom-Json
if (-not $ctx.resolved) { throw "repo-context unresolved: $($ctx.reason)" }
$REPO_NAME = $ctx.repoName
$AREA_PATH = $ctx.areaPath

Bug routing — separate sidecar

/pwagent-triage needs to map a failing test's file path → ADO area path. That's a different sidecar written by hooks/route-bug.ps1:

# Agent stages the request by writing the pending marker:
echo '{ "filePath": "src/packages/contacts/SaveContact.tsx" }' > "$RUN_DIR/.pending-routing.json"

# Next postToolUse fires route-bug.ps1, which writes:
jq . "$RUN_DIR/.bug-routing.json"
# { "areaPath": "Proj\\Contacts", "assignee": "alice", "tags": "playwright; contacts", ... }

Env-marker signals — separate sidecar

/pwagent-triage reads .env-signals.json (written by hooks/env-marker-scan.ps1 after each test-run postToolUse). See triage.agent.md step 2 for the aggregation logic.

What this skill explicitly DOES NOT do

• Does not auto-detect repo from cwd. Operators set --repo explicitly or rely on defaultRepo. See SPEC.md §15.3.
• Does not mutate config.json. The hook reads it; agents read the bound sidecar; nobody re-derives.
• Does not repeat the bind logic. There is one source of truth: hooks/bind-repo.ps1. Agents are consumers.

When the hook hasn't run yet

If the operator skipped prompt-submitted somehow (raw ACP session/prompt without going through the runtime's hook chain), no sidecar exists. Agents detect this:

[ -f "$RUN_DIR/.repo-context.json" ] || {
    echo "ERROR: hooks/bind-repo.ps1 did not run. Re-invoke via 'copilot ask' (which fires S2 events) or set --repo manually."
    exit 1
}

Halt in this case — the agent has no business deriving repo context.