#!/usr/bin/env bash
# headless-coding-agent-ci.sh — run a coding agent non-interactively inside CI with hard caps.
#
# What it does: invokes `claude -p` in bare mode with a turn cap, a dollar cap, the dontAsk
# permission mode plus an explicit allow list, captures the JSON result, extracts the cost and
# the result text with Python (no jq needed), and fails the job when the agent exits non-zero
# or the estimated cost exceeds COST_CAP_USD.
#
# Inputs (environment):
#   TASK_PROMPT        the task (default: a read-only summary task)
#   MAX_TURNS     agentic turn cap (default 12)
#   MAX_BUDGET    --max-budget-usd passed to the agent (default 2.00)
#   COST_CAP_USD  fail the job above this total_cost_usd (default 1.50)
#   ALLOWED_TOOLS permission rules (default: Read plus pytest/git diff/git log)
#   DRY_RUN=1     replace the agent binary with a local Python stub that prints a JSON result
#                 shaped like the real one, so the wrapper can be exercised without credentials.
# Run:  bash headless-coding-agent-ci.sh          (needs ANTHROPIC_API_KEY unless DRY_RUN=1)
#       DRY_RUN=1 bash headless-coding-agent-ci.sh
# Requires: bash, python3 (3.13 tested); the `claude` CLI for real runs.
set -euo pipefail

TASK_PROMPT="${TASK_PROMPT:-Summarize what this repository does in five bullet points. Do not edit files.}"
MAX_TURNS="${MAX_TURNS:-12}"
MAX_BUDGET="${MAX_BUDGET:-2.00}"
COST_CAP_USD="${COST_CAP_USD:-1.50}"
ALLOWED_TOOLS="${ALLOWED_TOOLS:-Read,Bash(python -m pytest *),Bash(git diff *),Bash(git log *)}"
OUT="${OUT:-agent-result.json}"

if [[ "${DRY_RUN:-0}" == "1" ]]; then
  # A stand-in for the agent: prints one JSON object with the fields the wrapper reads.
  # It is NOT a model call; it exists so the control flow below can be tested offline.
  AGENT=(python -c 'import json,sys; print(json.dumps({"type":"result","subtype":"success","is_error":False,"num_turns":3,"duration_ms":1200,"total_cost_usd":0.0421,"session_id":"dry-run-session","result":"DRY RUN: the stub printed this instead of a model answer. Prompt length: %d chars" % len(sys.argv[-1])}))')
else
  AGENT=(claude --bare -p
         --output-format json
         --permission-mode dontAsk
         --permission-prompts none
         --allowedTools "$ALLOWED_TOOLS"
         --max-turns "$MAX_TURNS"
         --max-budget-usd "$MAX_BUDGET")
fi

echo "== running agent (turns<=$MAX_TURNS, budget<=\$$MAX_BUDGET, cap=\$$COST_CAP_USD, dry_run=${DRY_RUN:-0})"
set +e
"${AGENT[@]}" "$TASK_PROMPT" > "$OUT"
STATUS=$?
set -e
echo "== agent exit status: $STATUS"
if [[ $STATUS -ne 0 ]]; then
  echo "agent failed; raw output follows"; cat "$OUT"; exit $STATUS
fi

# Extract cost and result without jq; exit 3 when the cost cap is exceeded.
python - "$OUT" "$COST_CAP_USD" <<'PY'
import json, sys
path, cap = sys.argv[1], float(sys.argv[2])
with open(path, encoding="utf-8") as f:
    data = json.load(f)
cost = float(data.get("total_cost_usd") or 0.0)
print(f"== session {data.get('session_id')} subtype={data.get('subtype')} turns={data.get('num_turns')} cost=${cost:.4f}")
print("== result:")
print(data.get("result", "").strip())
if data.get("is_error"):
    print("== agent reported is_error=true"); sys.exit(2)
if cost > cap:
    print(f"== cost ${cost:.4f} exceeds cap ${cap:.2f}"); sys.exit(3)
PY
echo "== ok"