diff --git a/src/agents/launcher.py b/src/agents/launcher.py
index 75b7cb8..dd2be7a 100644
--- a/src/agents/launcher.py
+++ b/src/agents/launcher.py
@@ -209,9 +209,15 @@ class AgentLauncher:
 
         # No git fetch/checkout here: ensure_worktree() already put the worktree on
         # the right branch. The agent simply runs inside its isolated work_path.
+        # Feature 4 (token usage): --output-format json makes claude emit a single
+        # result JSON (with usage + total_cost_usd) at the end of stdout. The log
+        # still captures it; _monitor_agent parses the trailing JSON after the run
+        # to record per-agent tokens/cost. _monitor_agent's failure handling keys
+        # off the process exit_code (not stdout shape), so this is safe.
         cmd = (
             f'cd {work_path} && '
             f'{self.CLAUDE_BIN} --print '
+            f'--output-format json '
             f'{model_flag}'
             f'"$(cat {task_file})" '
             f'--system-prompt "$(cat {system_prompt})" '
@@ -400,6 +406,17 @@ class AgentLauncher:
 
         notify_agent_finished(run_id, agent, exit_code, task_id=_task_id, duration_s=_duration_s)
 
+        # Feature 4: parse token usage / cost from the (json) run log and record
+        # it on the agent_runs row. Never fatal — a garbled/missing JSON records
+        # NULLs and logs a warning so a broken run can't crash the monitor.
+        try:
+            from ..usage import parse_usage_from_log, record_usage
+            _usage = parse_usage_from_log(output_path) if output_path else None
+            record_usage(run_id, _usage)
+        except Exception as e:
+            logger.warning(f"run_id={run_id}: usage accounting failed: {e}")
+            _usage = None
+
         # Commit and push any changes — in the per-branch worktree (ORCH-2 / S-4),
         # NOT in the shared /repos/<repo>. The worktree is already on `branch`
         # (ensure_worktree did the checkout), so no checkout is needed here.
@@ -490,6 +507,14 @@ class AgentLauncher:
                 from ..notifications import send_telegram
                 send_telegram(f"\u26a0\ufe0f {_wid}: Agent {agent} failed (exit_code={exit_code}). Check logs: /app/data/runs/{run_id}.log")
 
+        # Feature 4: post the per-agent usage comment under that agent's bot, and
+        # — for the deployer finishing the task — the per-task usage summary.
+        if exit_code == 0:
+            try:
+                self._post_usage_comments(run_id, agent, repo, branch, _usage)
+            except Exception as e:
+                logger.warning(f"run_id={run_id}: usage comment failed: {e}")
+
         # Auto-advance stage if agent finished successfully and QG passes
         if exit_code == 0:
             self._try_advance_stage(run_id, agent, repo, branch)
@@ -654,6 +679,32 @@ class AgentLauncher:
             logger.error(f"Auto-advance failed for run_id={run_id}: {e}")
 
 
+    def _post_usage_comments(self, run_id, agent, repo, branch, usage):
+        """Feature 4: post the per-agent usage comment (and Deployer summary).
+
+        - Always (on success, with a work_item_id): a per-agent finish comment
+          with token/cost, authored by the finishing agent's Plane bot.
+        - When the deployer finishes: also a per-task summary (SUM over
+          agent_runs GROUP BY agent), authored by the deployer.
+        """
+        from ..usage import usage_comment, task_summary_comment
+        conn = get_db()
+        row = conn.execute(
+            "SELECT id, work_item_id FROM tasks WHERE repo=? AND branch=?",
+            (repo, branch),
+        ).fetchone()
+        conn.close()
+        if not row:
+            return
+        task_id, work_item_id = row[0], row[1]
+        if not work_item_id:
+            return
+        plane_add_comment(work_item_id, usage_comment(agent, usage), author=agent)
+        if agent == "deployer":
+            plane_add_comment(
+                work_item_id, task_summary_comment(task_id), author="deployer"
+            )
+
     def _ensure_pr(self, repo: str, branch: str, run_id: int):
         import httpx
         owner = settings.gitea_owner
diff --git a/src/db.py b/src/db.py
index 0b77610..f8bfa9d 100644
--- a/src/db.py
+++ b/src/db.py
@@ -77,6 +77,13 @@ def init_db():
         "CREATE UNIQUE INDEX IF NOT EXISTS idx_events_delivery "
         "ON events(delivery_id) WHERE delivery_id IS NOT NULL"
     )
+    # Feature 4 (token usage): per-run token / cost accounting. Parsed from the
+    # claude --output-format json result by the launcher monitor. Idempotent
+    # ALTERs (no-op once the columns exist) so this is safe on the live prod DB.
+    _ensure_column(conn, "agent_runs", "input_tokens", "INTEGER")
+    _ensure_column(conn, "agent_runs", "output_tokens", "INTEGER")
+    _ensure_column(conn, "agent_runs", "cache_read_tokens", "INTEGER")
+    _ensure_column(conn, "agent_runs", "cost_usd", "REAL")
     conn.commit()
     conn.close()
 
diff --git a/src/usage.py b/src/usage.py
new file mode 100644
index 0000000..8968381
--- /dev/null
+++ b/src/usage.py
@@ -0,0 +1,268 @@
+"""Feature 4: token / cost accounting for agent runs.
+
+claude --output-format json emits a single result JSON object at the end of the
+run log with fields:
+  total_cost_usd
+  usage.input_tokens / output_tokens / cache_read_input_tokens /
+       cache_creation_input_tokens
+  modelUsage, num_turns, duration_ms
+
+This module parses that JSON out of a (text-or-json) run log, records the usage
+on the agent_runs row, formats a Plane comment for the finishing agent, and
+builds the per-task summary the Deployer posts on deploy/done.
+
+Everything here is defensive: a missing/garbled JSON never raises \u2014 we record
+NULL/0 and log a warning so a broken agent run can't crash the monitor.
+"""
+
+import json
+import logging
+
+from .db import get_db
+
+logger = logging.getLogger("orchestrator.usage")
+
+
+def parse_usage_from_text(text: str) -> dict | None:
+    """Extract the claude result-JSON usage from a run log's text.
+
+    The log may contain plain text before/after the JSON; with
+    --output-format json the JSON is the final object. We scan for the LAST
+    top-level '{' ... '}' that parses and carries usage/total_cost_usd.
+
+    Returns a normalised dict
+      {input_tokens, output_tokens, cache_read_tokens, cost_usd}
+    (ints / float, missing fields -> 0 / 0.0), or None if no usable JSON found.
+    """
+    if not text:
+        return None
+
+    candidate = _extract_last_json_object(text)
+    if candidate is None:
+        return None
+
+    usage = candidate.get("usage") or {}
+    if not isinstance(usage, dict):
+        usage = {}
+
+    cost = candidate.get("total_cost_usd")
+    if cost is None:
+        cost = candidate.get("cost_usd")
+
+    # If there is neither a usage block nor a cost, this isn't a result object.
+    if not usage and cost is None:
+        return None
+
+    def _int(v):
+        try:
+            return int(v)
+        except (TypeError, ValueError):
+            return 0
+
+    def _float(v):
+        try:
+            return float(v)
+        except (TypeError, ValueError):
+            return 0.0
+
+    return {
+        "input_tokens": _int(usage.get("input_tokens")),
+        "output_tokens": _int(usage.get("output_tokens")),
+        "cache_read_tokens": _int(
+            usage.get("cache_read_input_tokens", usage.get("cache_read_tokens"))
+        ),
+        "cost_usd": _float(cost),
+    }
+
+
+def _extract_last_json_object(text: str) -> dict | None:
+    """Return the last balanced top-level JSON object in `text` that parses.
+
+    Scans from the end for '}' and walks back to the matching '{' using a depth
+    counter (string-aware), trying json.loads on each candidate. Robust to log
+    lines or text emitted before the JSON.
+    """
+    # Fast path: the whole stripped text is the JSON.
+    stripped = text.strip()
+    try:
+        obj = json.loads(stripped)
+        if isinstance(obj, dict):
+            return obj
+    except (ValueError, TypeError):
+        pass
+
+    # Otherwise find the last balanced { ... } block.
+    end = len(text)
+    while True:
+        close = text.rfind("}", 0, end)
+        if close == -1:
+            return None
+        depth = 0
+        in_str = False
+        esc = False
+        start = None
+        for i in range(close, -1, -1):
+            ch = text[i]
+            if in_str:
+                if esc:
+                    esc = False
+                elif ch == "\\":
+                    esc = True
+                elif ch == '"':
+                    in_str = False
+                continue
+            if ch == '"':
+                in_str = True
+            elif ch == "}":
+                depth += 1
+            elif ch == "{":
+                depth -= 1
+                if depth == 0:
+                    start = i
+                    break
+        if start is not None:
+            blob = text[start:close + 1]
+            try:
+                obj = json.loads(blob)
+                if isinstance(obj, dict):
+                    return obj
+            except (ValueError, TypeError):
+                pass
+        end = close  # keep scanning earlier in the text
+
+
+def parse_usage_from_log(path: str) -> dict | None:
+    """Read a run log file and parse usage from it. Never raises."""
+    try:
+        with open(path, "r", encoding="utf-8", errors="replace") as f:
+            return parse_usage_from_text(f.read())
+    except OSError as e:
+        logger.warning(f"parse_usage_from_log: cannot read {path}: {e}")
+        return None
+
+
+def record_usage(run_id: int, usage: dict | None):
+    """Write parsed usage onto the agent_runs row. NULLs if usage is None."""
+    if usage is None:
+        logger.warning(f"run_id={run_id}: no usage JSON parsed, recording NULLs")
+        usage = {}
+    conn = get_db()
+    try:
+        conn.execute(
+            "UPDATE agent_runs SET input_tokens=?, output_tokens=?, "
+            "cache_read_tokens=?, cost_usd=? WHERE id=?",
+            (
+                usage.get("input_tokens"),
+                usage.get("output_tokens"),
+                usage.get("cache_read_tokens"),
+                usage.get("cost_usd"),
+                run_id,
+            ),
+        )
+        conn.commit()
+    finally:
+        conn.close()
+
+
+def fmt_tokens(n) -> str:
+    """Format a token count compactly: 1234 -> '1.2k', 2_500_000 -> '2.5M'."""
+    try:
+        n = int(n or 0)
+    except (TypeError, ValueError):
+        n = 0
+    if n >= 1_000_000:
+        return f"{n / 1_000_000:.1f}M"
+    if n >= 1_000:
+        return f"{n / 1_000:.1f}k"
+    return str(n)
+
+
+def fmt_cost(c) -> str:
+    """Format USD cost with 2 decimals: '$0.21'."""
+    try:
+        c = float(c or 0.0)
+    except (TypeError, ValueError):
+        c = 0.0
+    return f"${c:.2f}"
+
+
+# Pretty agent names for comments (mirrors STAGE_AUTHORS roles).
+AGENT_DISPLAY = {
+    "analyst": "Analyst",
+    "architect": "Architect",
+    "developer": "Developer",
+    "reviewer": "Reviewer",
+    "tester": "Tester",
+    "deployer": "Deployer",
+}
+
+
+def usage_comment(agent: str, usage: dict | None) -> str:
+    """Build the per-agent finish comment, e.g.
+    '\U0001f4bb Developer \u0433\u043e\u0442\u043e\u0432 \u00b7 45.2k in / 12.1k out \u00b7 $0.21'.
+    """
+    usage = usage or {}
+    name = AGENT_DISPLAY.get(agent, agent.capitalize())
+    icon = AGENT_ICON.get(agent, "\u2705")
+    return (
+        f"{icon} {name} \u0433\u043e\u0442\u043e\u0432 \u00b7 "
+        f"{fmt_tokens(usage.get('input_tokens'))} in / "
+        f"{fmt_tokens(usage.get('output_tokens'))} out \u00b7 "
+        f"{fmt_cost(usage.get('cost_usd'))}"
+    )
+
+
+AGENT_ICON = {
+    "analyst": "\U0001f50d",
+    "architect": "\U0001f4d0",
+    "developer": "\U0001f4bb",
+    "reviewer": "\U0001f50e",
+    "tester": "\U0001f9ea",
+    "deployer": "\U0001f680",
+}
+
+
+def task_usage_summary(task_id: int) -> dict:
+    """Aggregate agent_runs usage for a task.
+
+    Returns {total_in, total_out, total_cost, per_agent: [(agent, in, out, cost), ...]}.
+    """
+    conn = get_db()
+    try:
+        rows = conn.execute(
+            "SELECT agent, "
+            "COALESCE(SUM(input_tokens),0), "
+            "COALESCE(SUM(output_tokens),0), "
+            "COALESCE(SUM(cost_usd),0.0) "
+            "FROM agent_runs WHERE task_id=? GROUP BY agent ORDER BY agent",
+            (task_id,),
+        ).fetchall()
+    finally:
+        conn.close()
+    per_agent = [(r[0], int(r[1]), int(r[2]), float(r[3])) for r in rows]
+    total_in = sum(r[1] for r in per_agent)
+    total_out = sum(r[2] for r in per_agent)
+    total_cost = sum(r[3] for r in per_agent)
+    return {
+        "total_in": total_in,
+        "total_out": total_out,
+        "total_cost": total_cost,
+        "per_agent": per_agent,
+    }
+
+
+def task_summary_comment(task_id: int) -> str:
+    """Build the Deployer end-of-task summary comment (Feature 4, variant B)."""
+    s = task_usage_summary(task_id)
+    lines = [
+        f"\U0001f4ca \u0418\u0442\u043e\u0433\u043e \u043f\u043e \u0437\u0430\u0434\u0430\u0447\u0435: "
+        f"{fmt_tokens(s['total_in'])} \u0442\u043e\u043a\u0435\u043d\u043e\u0432 \u0432\u0445\u043e\u0434 / "
+        f"{fmt_tokens(s['total_out'])} \u0432\u044b\u0445\u043e\u0434 \u00b7 "
+        f"{fmt_cost(s['total_cost'])}"
+    ]
+    for agent, ti, to, cost in s["per_agent"]:
+        name = AGENT_DISPLAY.get(agent, agent.capitalize())
+        lines.append(
+            f"\u2022 {name}: {fmt_tokens(ti)} in / {fmt_tokens(to)} out \u00b7 {fmt_cost(cost)}"
+        )
+    return "\n".join(lines)
diff --git a/tests/test_usage.py b/tests/test_usage.py
new file mode 100644
index 0000000..59d059b
--- /dev/null
+++ b/tests/test_usage.py
@@ -0,0 +1,176 @@
+"""Feature 4: token / cost accounting tests.
+
+Covers:
+  * parse_usage_from_text on a REAL claude --output-format json result blob
+    (captured live from CLI 2.1.142), including a leading text line.
+  * parse on garbage / missing JSON -> None (never raises).
+  * record_usage writes the columns; NULLs when usage is None.
+  * fmt_tokens / fmt_cost formatting.
+  * usage_comment string format.
+  * task_usage_summary / task_summary_comment aggregate over agent_runs.
+
+DB is an isolated temp file; no network or subprocess.
+"""
+
+import os
+import tempfile
+
+os.environ.setdefault("ORCH_PLANE_API_TOKEN", "test-token")
+os.environ.setdefault("ORCH_GITEA_TOKEN", "test-token")
+
+_test_db = os.path.join(tempfile.gettempdir(), "test_orchestrator_usage.db")
+os.environ["ORCH_DB_PATH"] = _test_db
+
+import pytest  # noqa: E402
+
+from src import db as db_module  # noqa: E402
+from src.db import init_db, get_db  # noqa: E402
+from src import usage as U  # noqa: E402
+
+
+# Real claude --output-format json result object (captured from CLI 2.1.142).
+REAL_RESULT_JSON = (
+    '{"type":"result","subtype":"success","is_error":false,"duration_ms":1795,'
+    '"num_turns":1,"result":"Hi!","session_id":"abc",'
+    '"total_cost_usd":0.0560175,'
+    '"usage":{"input_tokens":45231,"cache_creation_input_tokens":7418,'
+    '"cache_read_input_tokens":18500,"output_tokens":12100,'
+    '"service_tier":"standard"},'
+    '"modelUsage":{"claude-opus-4-7":{"inputTokens":6,"outputTokens":7}},'
+    '"permission_denials":[]}'
+)
+
+
+@pytest.fixture(autouse=True)
+def setup_db(monkeypatch):
+    # get_db() reads settings.db_path live; pin it to our isolated DB.
+    monkeypatch.setattr(db_module.settings, "db_path", _test_db, raising=False)
+    if os.path.exists(_test_db):
+        os.unlink(_test_db)
+    init_db()
+    yield
+    if os.path.exists(_test_db):
+        os.unlink(_test_db)
+
+
+# --------------------------------------------------------------------------- #
+# parsing
+# --------------------------------------------------------------------------- #
+def test_parse_real_result_json():
+    u = U.parse_usage_from_text(REAL_RESULT_JSON)
+    assert u is not None
+    assert u["input_tokens"] == 45231
+    assert u["output_tokens"] == 12100
+    assert u["cache_read_tokens"] == 18500
+    assert abs(u["cost_usd"] - 0.0560175) < 1e-9
+
+
+def test_parse_with_leading_text():
+    """The agent may print text before the trailing JSON; we still find it."""
+    text = "some agent stdout line\nanother line\n" + REAL_RESULT_JSON
+    u = U.parse_usage_from_text(text)
+    assert u is not None
+    assert u["input_tokens"] == 45231
+    assert u["output_tokens"] == 12100
+
+
+def test_parse_garbage_returns_none():
+    assert U.parse_usage_from_text("not json at all { broken") is None
+    assert U.parse_usage_from_text("") is None
+    assert U.parse_usage_from_text(None) is None
+
+
+def test_parse_json_without_usage_returns_none():
+    assert U.parse_usage_from_text('{"hello":"world"}') is None
+
+
+def test_parse_from_log_missing_file_returns_none():
+    assert U.parse_usage_from_log("/no/such/file.log") is None
+
+
+# --------------------------------------------------------------------------- #
+# record_usage
+# --------------------------------------------------------------------------- #
+def _new_run(agent="developer", task_id=1):
+    conn = get_db()
+    cur = conn.execute("INSERT INTO agent_runs (task_id, agent) VALUES (?, ?)", (task_id, agent))
+    rid = cur.lastrowid
+    conn.commit()
+    conn.close()
+    return rid
+
+
+def test_record_usage_writes_columns():
+    rid = _new_run()
+    u = U.parse_usage_from_text(REAL_RESULT_JSON)
+    U.record_usage(rid, u)
+    conn = get_db()
+    row = conn.execute(
+        "SELECT input_tokens, output_tokens, cache_read_tokens, cost_usd "
+        "FROM agent_runs WHERE id=?", (rid,)
+    ).fetchone()
+    conn.close()
+    assert row["input_tokens"] == 45231
+    assert row["output_tokens"] == 12100
+    assert row["cache_read_tokens"] == 18500
+    assert abs(row["cost_usd"] - 0.0560175) < 1e-9
+
+
+def test_record_usage_none_writes_nulls():
+    rid = _new_run()
+    U.record_usage(rid, None)  # must not raise
+    conn = get_db()
+    row = conn.execute("SELECT input_tokens, cost_usd FROM agent_runs WHERE id=?", (rid,)).fetchone()
+    conn.close()
+    assert row["input_tokens"] is None
+    assert row["cost_usd"] is None
+
+
+# --------------------------------------------------------------------------- #
+# formatting
+# --------------------------------------------------------------------------- #
+def test_fmt_tokens():
+    assert U.fmt_tokens(6) == "6"
+    assert U.fmt_tokens(1234) == "1.2k"
+    assert U.fmt_tokens(45231) == "45.2k"
+    assert U.fmt_tokens(2_500_000) == "2.5M"
+    assert U.fmt_tokens(None) == "0"
+
+
+def test_fmt_cost():
+    assert U.fmt_cost(0.21) == "$0.21"
+    assert U.fmt_cost(0.0560175) == "$0.06"
+    assert U.fmt_cost(None) == "$0.00"
+
+
+def test_usage_comment_format():
+    u = {"input_tokens": 45231, "output_tokens": 12100, "cost_usd": 0.21}
+    c = U.usage_comment("developer", u)
+    assert "Developer" in c
+    assert "45.2k in" in c
+    assert "12.1k out" in c
+    assert "$0.21" in c
+
+
+# --------------------------------------------------------------------------- #
+# task summary
+# --------------------------------------------------------------------------- #
+def test_task_summary_aggregates_over_agents():
+    # two runs for the same task: developer + tester
+    for agent, ti, to, cost in [("developer", 1000, 200, 0.10), ("tester", 500, 100, 0.05)]:
+        rid = _new_run(agent=agent, task_id=42)
+        U.record_usage(rid, {"input_tokens": ti, "output_tokens": to,
+                             "cache_read_tokens": 0, "cost_usd": cost})
+
+    s = U.task_usage_summary(42)
+    assert s["total_in"] == 1500
+    assert s["total_out"] == 300
+    assert abs(s["total_cost"] - 0.15) < 1e-9
+    agents = {a for a, *_ in s["per_agent"]}
+    assert agents == {"developer", "tester"}
+
+    comment = U.task_summary_comment(42)
+    assert "1.5k" in comment       # total in
+    assert "$0.15" in comment       # total cost
+    assert "Developer" in comment
+    assert "Tester" in comment